CoStar Data and Methodology · 2021. 1. 5. · CoStar’s data provides the raw material for a comprehensive view of U.S. commercial real estate. Per CoStar’s 10-K, as of January

CoStar Data and Methodology Third Edition August 2020 v3.0

2

TABLE OF CONTENTS 1 Introduction and Overview .................................................................................................. 4

1.1 A Note About the Covid-19 Outbreak ........................................................................... 5 2 The CoStar Data ................................................................................................................. 7

2.1 Commercial Data ......................................................................................................... 7 2.2 Multifamily Data ..........................................................................................................10

3 Geographic Definitions .......................................................................................................12 4 Historical Time Series: Fundamentals and Rent ...............................................................18

4.1 Supply .........................................................................................................................18 4.2 Vacancy and Demand .................................................................................................19

4.2.1 Commercial Vacancy and Demand..................................................................................... 20 4.2.2 Multifamily Vacancy and Demand ....................................................................................... 21

4.3 Commercial Rents ......................................................................................................25 4.3.1 Problems with Commercial Rents ....................................................................................... 26 4.3.2 Standardizing Service Types .............................................................................................. 30 4.3.3 Commercial Same-Store Rents .......................................................................................... 33 4.3.4 Estimating Rents in Properties Without Rent Information ................................................... 36 4.3.5 Comparing Same-Store Rents to Available Space-Weighted Rents .................................. 38

4.4 Multifamily Rents ........................................................................................................41 4.4.1 Same-Store Multifamily Rents ............................................................................................. 42 4.4.2 Concessions and Effective Rents ....................................................................................... 48

5 Performance Metrics ..........................................................................................................49 5.1 Approaches to Tracking Property Values ....................................................................49 5.2 The CoStar Transaction Data .....................................................................................52 5.3 Estimating Property Values .........................................................................................56

5.3.1 Identifying Comp Properties ................................................................................................ 58 5.3.2 Estimating Values in Large Traded Properties ................................................................... 59 5.3.3 Estimating Values in Large Untraded Properties ................................................................ 67 5.3.4 Estimating Values for Small Properties ............................................................................... 70 5.3.5 Accuracy of Value Estimates .............................................................................................. 70

5.4 Estimating Property Cap Rates ...................................................................................73 5.4.1 Spot Cap Rate Estimates .................................................................................................... 73 5.4.2 Cap Rate Trends ................................................................................................................. 73

6 Forecasting ........................................................................................................................75 6.1 Approaches to Commercial Real Estate Forecasting ..................................................76 6.2 Economic and Capital Markets Data ...........................................................................78

6.2.1 Economic and Capital Markets Inputs................................................................................. 78 6.2.2 Scenarios ............................................................................................................................ 78

6.3 Forecast “Guidelines”: The Market Models ................................................................79 6.3.1 The Supply Model ............................................................................................................... 80 6.3.2 The Vacancy Model ............................................................................................................ 85 6.3.3 The Rent Model ................................................................................................................... 87 6.3.4 Performance Models: NOI, Cap Rates, and Prices ........................................................... 88 6.3.5 Interventions to the Market Model ....................................................................................... 90

6.4 The Building-Level Forecast Model .............................................................................91 6.4.1 Building-Level Supply Forecasts ......................................................................................... 92 6.4.2 Allocating Demand Across Buildings .................................................................................. 93 6.4.3 Building-Level Rent Forecasts ............................................................................................ 94 6.4.4 Building-Level Performance ................................................................................................ 95 6.4.5 Interventions to the Building-Level Forecasts ..................................................................... 97

6.5 Backtesting And Model Validation ...............................................................................98 6.5.1 Backtesting the Market Models ........................................................................................... 98

CoStar Data and Methodology White Paper | Summer 2017

3

6.5.2 Backtesting the Building-Level Model ................................................................................. 99 7 The Future ....................................................................................................................... 100 8 Appendix .......................................................................................................................... 101

8.1 Changes in This Version ........................................................................................... 101 8.2 Change in Prior Versions .......................................................................................... 101

4

1 INTRODUCTION AND OVERVIEW

Since its founding in 1987, the CoStar Group has assembled an unparalleled library of building- and deal-level commercial and multifamily real estate information. Each day, more than 1,400 researchers, located in offices across the United States, Canada, the United Kingdom, and Europe canvass tens of thousands of space listings, lease deals, and transactions, while field research teams systematically document and photograph every building and construction project from specialized cars and aircraft equipped with military-grade surveillance technology. Each day, more than 4 million apartment rent observations flow into CoStar datasets via the firm’s ILS and marketing services, including Apartments.com, ApartmentFinder.com, ForRent.com, and LoopNet. And STR collects high-frequency data for more than 60,000 hotels around the world. This comprehensive and up-to-the-minute dataset has become indispensable to those in the commercial real estate industry, who use it to market space in their properties, to find space for their firms or clients, to identify comps as they consider buying or selling an asset, to benchmark their performance against peers, or to understand the competitive threat they face from new projects underway. CoStar’s Quantitative Analytics Group creates models and algorithms to extract from this voluminous dataset the most accurate and timely account of these trends in vacancies, rents, prices, and cap rates, whether for an individual property, a custom set of properties, a submarket or market, or the entire nation. The Quantitative Analytics Group also produces forecasts of the key real estate variables. Where other firms provide their clients with market-level forecasts, CoStar offers a forecast for every building, freeing our clients from the restrictions of geography and allowing them to view a forecast for any custom set of properties. Moreover, CoStar’s forecasting models update in real-time based on actual movements in the property, ensuring that our clients will never have to rely on forecasts based on months-old data. These property-level historical series and forecasts drive the CoStar Market Analytics service, by which CoStar clients can access full underwriting reports, rent comp reports, and market reports for all 390 CBSAs, as well as submarket reports for more than 3,000 submarkets. These reports feature narratives written by a team of more than 50 market analysts and economists who collectively represent the largest and most experienced team of commercial real estate experts in the industry. Taken together, CoStar’s comprehensive datasets, real-time property-level historical and forecasts series, and written analysis have transformed commercial real estate analytics and put local players on equal footing with the largest international institutional investors when it comes to market knowledge. This paper describes the underlying CoStar dataset, the models used to create property-level vacancy and rent series, the models used to estimate prices and cap rates, and the CoStar approach to forecasting and model validation. We have written

5

this paper for a non-technical audience, though the quantitatively inclined will find mathematical equations throughout (others can safely ignore these).

1.1 A NOTE ABOUT THE COVID-19 OUTBREAK The Covid-19 outbreak of 2020 resulted in a shutdown of economic activity not seen since the end of World War II. In April of 2020, the Bureau of Labor Statistics reported that the US economy lost 21 million jobs, and in a single month, the unemployment rate rose from 4% to 14%. The recovery also broke a record, as 4.8 million jobs were added in June 2020. The scale of the Covid-era economic data not only broke records, it broke chart axes, economic doctrine, and forecasting models. At CoStar, we had never entered figures like these into our forecasting model, not even in our most dire scenarios, or even as a test. In response to the Covid outbreak, we made several adjustments to our forecasting model. Vacancy Model We have adjusted our vacancy change model to use the trailing four-quarter average employment growth relative to the trailing 20-quarter median employment growth as the demand driver. Previously, the model used the current quarter employment growth relative to the trailing 20-quarter average. This change makes the vacancy forecast less dependent on any single quarter’s employment data, and smooths out and moderate vacancy movements. The use of the trailing median effectively removes the 2020Q2 employment drop from the historical trend. In addition, we changed our economic demand driver for retail and office to use the sector-specific employment: retail employment for the retail property type, and office-using employment for the office property type. Previously, we had used total employment as the demand driver for all property types. We believe that the use of sector-specific demand drivers captures the widening divergence in outcomes between the property types. Note that the regression models are still calibrated to the total employment history. In the forecast, the change sector-specific employment replaces the change in total employment as the inputs to the model. Rent Model We have adjusted our rent model to use a trailing four-quarter employment growth, in addition to the vacancy level, the vacancy change, and the long-term trend in rents. Previously, the model used the current quarter’s employment growth. The change makes rent movements less dependent on any single quarter’s employment data, while also smoothing out and moderating rent movements.

6

Supply Model We have recalibrated our supply model to use the period since 2008, rather than the full history. Our intent in making this change is to exclude the high-supply period in 2006 and 2007, especially for retail, and in doing so lower the amount of supply in the forecast, particularly for retail. Net Operating Income (NOI) Model Our price forecast derives from the cap rate and NOI forecast, per the equation cap rate = NOI / price. Thus, we require an assumption about NOI over the forecast period to arrive at a price forecast. The NOI model depends on trends in rent and vacancy. In the past, we had included assumptions about typical lease lengths in each market and property type. For example, office lease lengths average around five years, compared to just one year for multifamily. As a result, in our forecast, NOI for multifamily responded more quickly to changes in rent than did office NOI. We have adjusted the NOI model to use a trailing four-quarter change in rent for all property types. This change means that all property types will experience the same price change, given the same rent, vacancy, and cap rate trends. Under the prior model, office prices would have changed more gradually due to the longer lease terms. The use of the same trailing rent period also aligns the forecast for NOI (and by extension prices) with our historical model for price, set forth in section 5.3. Cap Rate Model We have changed our cap rate model to use the spread between the Baa corprorate yield and the 10-year Treasury yield (both as provided from Oxford Economics), rather than the absolute Baa corporate yield. This change removes the long-term upward trend in interest rates from our cap rate forecast. By using the Baa spread rather than the level, we are effectively assuming that the risk-free rate will remain flat at current levels. Section 6.2 outlines CoStar’s use of Oxford Economics data and forecasts in detail.

7

2 THE COSTAR DATA

Commercial real estate investors and financiers face particular data needs and challenges. Unlike stocks or bonds, property prices have no ticker, and buildings do not report quarterly financials or provide earnings guidance. To the contrary, owners and tenants keep their own counsel: U.S. REITs are not required to, and do not deign to, publically report the valuations for their underlying assets. Investors have had to rely on information they have on their own assets, broker advice and reports, and NCREIF performance trends to arrive at a view of market trends and outlook. This “hands-on-the-elephant” approach has characterized commercial real estate research and analysis for decades. But with the typical institutional real estate deal trading above $100 million and the total market capitalization of the asset class now approaching $15 trillion, investors who seek a more objective and systematic approach can achieve superior returns. The CoStar Group’s mission is to provide transparency and efficiency to this notoriously opaque market. The firm has invested billions of dollars to create and maintain a research apparatus that can track availabilities, vacancies, asking rents, and transactions for more than six million properties in real-time. This apparatus includes more than 1,800 researchers who, every day, contact market participants to inquire about the status of listings and properties; ILS services like Apartments.com, through which more than a million new rent data points each day flow into CoStar datasets; marketing platforms like LoopNet; STR, the trusted partner with whom the world’s hotels share their operating data; and sophisticated automated data collection technology to “scrape” rents from community websites.

2.1 COMMERCIAL DATA CoStar’s data provides the raw material for a comprehensive view of U.S. commercial real estate. Per CoStar’s 10-K, as of January 30, 2015 the CoStar database included approximately:

- 1.3 million sale and lease listings; - 4.5 million total properties; - 8.2 billion SF of sale and lease listings; - 5.8 million tenants; - 2.3 million sales transactions valued in the aggregate at approximately $5.7

trillion; and - 16.5 million digital attachments, including building photographs, aerial

photographs, plat maps, and floor plans.

8

This highly complex database contains hundreds of data fields, including:

- Location - Mortgage and deed information - Site and zoning information - For-sale information - Building characteristics - Income and expense histories - Space availability - Tenant names - Tax assessments

- Lease expirations - Ownership - Contact information - Sales and lease comparables - Historical trends - Space requirements - Demographic information - Number of retail stores - Retail sales per SF

This data constitutes an effective census of the commercial real estate universe, as well as detailed pictures of individual assets. Exhibit 1 presents the landing page on the CoStar product for 1331 L Street in Washington, the headquarters location of the CoStar Group:

Exhibit 1: CoStar Data for 1331 L Street, Washington, DC

CoStar began collecting data in different markets at different times, so some markets have longer time series than others. For the office and industrial property types, most major markets have data back to at least 2000. Smaller markets and the retail property type typically begin in 2006. Exhibit 2 shows the analytic start date for fundamentals data for major markets by property type.

9

Exhibit 2: Start Dates for Major Markets by Property Type

Market Office Start Industrial

Start Retail Start Dallas/Ft Worth 1982Q1 1995Q1 2000Q1

Atlanta 1986Q4 1990Q1 2000Q1

Washington DC 1993Q1 1993Q1 2006Q1

Baltimore 1994Q1 1994Q1 2006Q1

New York City 1994Q3 1994Q3 2006Q1

Northern New Jersey 1995Q2 1995Q2 2006Q1

Chicago 1996Q2 1996Q2 2006Q1

Los Angeles 1996Q2 2000Q4 2006Q1

Long Island (New York) 1996Q3 1996Q3 2006Q1

Westchester/So Connecticut 1996Q3 1996Q3 2006Q1

Orange (California) 1996Q4 1996Q4 2006Q1

East Bay/Oakland 1997Q1 1997Q1 2006Q1

San Francisco 1997Q1 1997Q1 2006Q1

South Bay/San Jose 1997Q1 1997Q1 2006Q1

Philadelphia 1997Q3 1997Q3 2006Q1

Boston 1998Q1 1998Q1 2006Q1

Sacramento 1998Q4 1998Q4 2006Q1

Houston 1999Q1 1999Q1 2006Q1

Phoenix 1999Q1 1999Q1 2006Q1

San Diego 1999Q2 1999Q2 2006Q1

Orlando 1999Q3 1999Q3 2006Q1

South Florida 1999Q3 1999Q3 2006Q1

Tampa/St Petersburg 1999Q3 1999Q3 2006Q1

Denver 1999Q4 1999Q4 2006Q1

Jacksonville (Florida) 1999Q4 1999Q4 2006Q1

Charlotte 2000Q1 2000Q1 2006Q1

Cincinnati/Dayton 2000Q1 2000Q1 2006Q1

Cleveland 2000Q1 2000Q1 2006Q1

Columbus 2000Q1 2000Q1 2006Q1

Detroit 2000Q1 2000Q1 2006Q1

Inland Empire (California) 2000Q1 2000Q2 2006Q1

Memphis 2000Q1 2000Q1 2000Q1

Seattle/Puget Sound 2000Q1 2000Q1 2006Q1

St. Louis 2000Q1 2000Q1 2006Q1

Indianapolis 2000Q3 2000Q3 2006Q1

Kansas City 2000Q3 2000Q3 2006Q1

Pittsburgh 2000Q3 2000Q3 2006Q1

Raleigh/Durham 2000Q3 2000Q3 2006Q1

Austin 2000Q4 2000Q4 2006Q1

Nashville 2000Q4 2000Q4 2000Q4

Portland 2003Q1 2003Q1 2006Q1

Hampton Roads 2005Q2 2005Q2 2005Q3

Richmond VA 2005Q2 2005Q2 2005Q2

Las Vegas 2005Q3 2005Q3 2005Q3

San Antonio 2005Q3 2005Q3 2005Q3

Tucson 2005Q3 2005Q3 2005Q3

10

Salt Lake City 2005Q4 2005Q4 2005Q4

Birmingham 2006Q1 2006Q1 2006Q1

Greensboro/Winston-Salem 2006Q1 2006Q1 2006Q1

Greenville/Spartanburg 2006Q1 2006Q1 2006Q1

Hartford 2006Q1 2006Q1 2006Q1

Milwaukee/Madison 2006Q1 2006Q1 2006Q1

Minneapolis/St Paul 2006Q1 2006Q1 2006Q1

Oklahoma City 2006Q1 2006Q1 2006Q1

Providence 2006Q1 2006Q1 2006Q1

Southwest Florida 2006Q1 2006Q1 2006Q1

Toledo 2006Q1 2006Q1 2006Q1

Tulsa 2006Q1 2006Q1 2006Q1

Data for other markets begin in 2007. Same-store rent series (discussed in detail in Section 4.3 and 4.4) extend back to at least 2000 for all office, multifamily, and industrial markets, and to 2006 for all retail properties.

2.2 MULTIFAMILY DATA CoStar began researching the multifamily space in 2013 and has amassed a data set of more than 450,000 apartment properties totaling more than 11 million units, consistent with the firm’s mission of recording every property. For comparison, the U.S. Census Bureau’s American Housing Survey records total multifamily stock of about 9 million rental units in rental properties with at least 20 units as of 2017, while other multifamily analytics firms track fewer than 100,000 properties. The apartment leasing market differs in important ways from that of other property types. Most significantly, the market consists of millions of individual potential renters, as opposed to the thousands of firms seeking office, retail, or warehouse space. Whereas an office landlord may lease an entire building to a single tenant, apartment landlords must market and lease each unit individually. Rents for similar units can vary widely within the same building depending on the lease term, the unit’s location in the property, its views, and other often intangible qualities. Office landlords have an incentive to list and market all the space they have available; apartment landlords might feel they can improve their negotiating position by marketing only a few units, even if many more are in fact available. Moreover, the asking rent for an office space typically does not change much day to day—but apartment rents can vary day to day, or even hour to hour, based on the lease-rent optimization (LRO) programs employed by many apartment communities. Finally, apartment renters, most of whom will only search for a new apartment a few times, will typically not pay hefty fees for access to a listing service like CoStar. Because of these differences, CoStar’s traditional research approach—directly asking property managers about the availability, vacancy, and asking rents in their buildings— does not translate as well to tracking apartments. Researchers at best gain visibility into broad categories of units or models at a community, rather than the unit-level data most relevant to property managers. And human data collection can, at best, achieve a

11

monthly update frequency across a large number of properties. But today’s multifamily industry sets rents daily, and so monthly, or worse, quarterly data becomes outdated within days of collection. Moreover, the rent figures provided by a human respondent may be estimates or approximations across the entire community, rather than precise unit-level figures that renters actually use to decide where to live. The vast majority of CoStar’s rent data comes via Apartments.com and the CoStar Group’s other ILS platforms. Each day, the firm adds more than 4 million rent data points that automatically flow into our datasets, for more than 50,000 properties. We augment this high-frequency data with more traditional data collection methods. CoStar’s multifamily research team calls some 20,000 properties a month to collect vacancy rates and rents. CoStar has also developed automated data collection technology to “scrape” daily rent information from property websites. Finally, to extend the data history, CoStar purchased RealFacts’ apartment rent data set, which tracks some 12,000 properties back into the 1990s. Exhibit 3 shows the scale of CoStar’s rent data :

Exhibit 3: CoStar Apartment Rent data, Neartown/River Oak Submarket in Houston

The scale and quality of CoStar’s rental data allows us to create accurate high-frequency analytic series that can show a market’s reaction to events in near-real time. For example, Exhibit 4 shows our daily rent series for Houston during Hurricane Harvey in September 2017, and more recently, during the Covid outbreak.

12

Exhibit 4: Daily Rent Series for Houston

3 GEOGRAPHIC DEFINITIONS

CoStar’s data represents an effective census of commercial real estate, including more than 6 million properties. By offering data and forecasts at the individual property level, CoStar clients may aggregate the data in any way they see fit. However, for the purposes of defining market coverage and for the convenience for our clients, we have established market and submarket definitions. CoStar’s market definitions generally match the Core Based Statistical Areas (or “CBSA”), including the Metropolitan Divisions as defined by the Office of Management and Budget, ensuring that the economic data that underlies the forecasts correspond to the real estate data1. Within each market, CoStar researchers and analysts, in consultation with local market participants, have drawn submarkets corresponding to each of the four main property types. On average, each major market contains 25 submarkets. Exhibit 5 shows the submarket map for the Washington-Northern Virginia-Maryland office market.

1 For certain markets, including Los Angeles, New York, and Boston, we use the smaller MSA definitions.

13

Exhibit 5: Washington, D.C. Office Submarket Map

14

Exhibit 6 zooms into the downtown area, for which CoStar maintains nine distinct office submarkets:

Exhibit 6: Downtown Washington, D.C. Office Submarket Map

CoStar also reports data for submarket clusters, which represent aggregations of submarkets (as in New York Midtown or Downtown Washington). Exhibit 7 presents the 88 office submarkets and corresponding clusters for the Washington, D.C., market (defined as CBSA 47900).

15

Exhibit 7: Washington, D.C. Office Submarket List

Research Market Submarket Name Cluster Name

Washington, D.C. Crystal City Alexandria/I-395 Area Washington, D.C. Eisenhower Ave Corridor Alexandria/I-395 Area Washington, D.C. I-395 Corridor Alexandria/I-395 Area Washington, D.C. Old Town Alexandria Alexandria/I-395 Area Washington, D.C. Pentagon City Alexandria/I-395 Area Washington, D.C. Bethesda/Chevy Chase Bethesda/Chevy Chase Washington, D.C. Capitol Hill Capitol Hill Area Washington, D.C. Southwest Capitol Hill Area Washington, D.C. NoMa Capitol Hill Area Washington, D.C. Capitol Riverfront Capitol Hill Area Washington, D.C. CBD Downtown DC Washington, D.C. East End Downtown DC Washington, D.C. West End Downtown DC Washington, D.C. Great Falls Dulles Corridor Washington, D.C. Herndon Dulles Corridor Washington, D.C. Reston Dulles Corridor Washington, D.C. Route 28 Corridor North Dulles Corridor Washington, D.C. Route 28 Corridor South Dulles Corridor Washington, D.C. Bowie E Prince George's County Washington, D.C. Greater Upper Marlboro E Prince George's County Washington, D.C. N Arlington/E FallsChurch East Falls Church Washington, D.C. Fauquier County/Vint Hill Fauquier County/Vint Hill Washington, D.C. Frederick Frederick Washington, D.C. Georgetown Georgetown/Uptown Washington, D.C. Uptown Georgetown/Uptown Washington, D.C. Annandale Greater Fairfax County Washington, D.C. Fairfax Center Greater Fairfax County Washington, D.C. Fairfax City Greater Fairfax County Washington, D.C. Falls Church Greater Fairfax County Washington, D.C. McLean Greater Fairfax County Washington, D.C. Merrifield Greater Fairfax County Washington, D.C. Oakton Greater Fairfax County Washington, D.C. Tysons Corner Greater Fairfax County Washington, D.C. Vienna Greater Fairfax County Washington, D.C. Gaithersburg I-270 Corridor Washington, D.C. Germantown I-270 Corridor Washington, D.C. I-270 Corridor North I-270 Corridor Washington, D.C. North Bethesda/Potomac I-270 Corridor Washington, D.C. North Rockville I-270 Corridor Washington, D.C. Rockville I-270 Corridor Washington, D.C. Leesburg/West Loudoun Leesburg/Route 7 Corridor

16

Washington, D.C. Route 7 Corridor Leesburg/Route 7 Corridor Washington, D.C. Manassas Manassas/Route 29/I-66 Washington, D.C. Route 29/I-66 Corridor Manassas/Route 29/I-66 Washington, D.C. Beltsville/Calverton N Prince George's County Washington, D.C. College Park N Prince George's County Washington, D.C. Greenbelt N Prince George's County Washington, D.C. Landover/Largo/Captl Hts N Prince George's County Washington, D.C. Lanham N Prince George's County Washington, D.C. Laurel N Prince George's County Washington, D.C. Northeast Northeast/Southeast Washington, D.C. Southeast Northeast/Southeast Washington, D.C. Outlying Montgmery Cnty E Outlying Montgmery Cnty E Washington, D.C. Outlying Montgmery Cnty W Outlying Montgmery Cnty W Washington, D.C. Outlying PG County South Outlying PG County South Washington, D.C. Ballston R-B Corridor Washington, D.C. Clarendon/Courthouse R-B Corridor Washington, D.C. Rosslyn R-B Corridor Washington, D.C. Virginia Square R-B Corridor Washington, D.C. Branch Avenue Corridor S Prince George's County Washington, D.C. NatHbr/OxnHill/FtWash S Prince George's County Washington, D.C. Pennsylvania Ave Corridor S Prince George's County Washington, D.C. Huntington/Mt Vernon SE Fairfax County Washington, D.C. Springfield/Burke SE Fairfax County Washington, D.C. Kensington/Wheaton SE Montgomery County Washington, D.C. North Silver Spring/Rt 29 SE Montgomery County Washington, D.C. Silver Spring SE Montgomery County Washington, D.C. Woodbridge/I-95 Corridor Woodbridge/I-95 Corridor Washington, D.C. Washington County Washington County Washington, D.C. Spotsylvania County Greater Fredericksburg Washington, D.C. Stafford County Greater Fredericksburg Washington, D.C. Fredericksburg City Greater Fredericksburg Washington, D.C. King George County Greater Fredericksburg Washington, D.C. Charles County Charles County Washington, D.C. Calvert County Calvert County Washington, D.C. Clarke County Clarke County Washington, D.C. Frederick County Frederick County Washington, D.C. St Mary's County St Mary's County Washington, D.C. Warren County Warren County Washington, D.C. Winchester City Winchester City Washington, D.C. Morgan County Morgan County Washington, D.C. Berkeley County Berkeley County Washington, D.C. Jefferson County Jefferson County Washington, D.C. Hampshire County Hampshire County Washington, D.C. Fort Belvoir Army Reserve Fort Belvoir Army Reserve

17

Washington, D.C. Quantico Marine Base Quantico Marine Base Washington, D.C. Caroline County Caroline County

Finally, we use the concept of a “slice” to group together similar properties. We define slices as follows:

Exhibit 8: Slice Definitions

Property Type Name Includes

Office Class A 4 & 5 Star properties

Office Class B 3 Star properties

Office Class C 1 & 2 Star properties

Industrial Logistics Warehouse & Distribution subtypes

Industrial Specialized Industrial All other Industrial subtypes

Industrial Flex All Flex properties

Multifamily Class A 4 & 5 Star properties

Multifamily Class B 3 Star properties

Multifamily Class C 1 & 2 Star properties

Retail Malls Malls

Retail Neighborhood Center Neighborhood Centers

Retail Power Center Power Centers

Retail Strip Center Strip Centers

Retail General Retail Stand-alone retail locations

Retail Other Other retail locations, including airport retail and retail in other public locations.

18

4 HISTORICAL TIME SERIES: FUNDAMENTALS AND RENT

Commercial real estate analysis begins with the fundamentals: Supply, demand, and vacancy. CoStar’s comprehensive property-level data on property size, date built, and tenancy makes these metrics easy to track—we need only add up the supply and demand, and calculate the vacancy rate. Rents pose a greater analytic challenge. CoStar collects asking rent data on spaces available for lease. Thus, only properties with availability report a rent, and that set of properties changes constantly. The changing mix of properties that report rents period to period confounds analyses of rent trends, since the appearance of a new large block at a high or low rent relative to its peers can give the appearance of large swings in rents. To control for the effect of changing composition, and to ensure every property has a rent in every period, we have constructed a “same-store” rent series, covered in Section 4.3. This section details the specific methodology by which CoStar creates submarket- and market-level series for commercial real estate fundamentals and rents, starting with supply.

4.1 SUPPLY In commercial real estate, “supply” can denote the total stock, usually expressed in rentable building area (RBA) or units (for multifamily assets), of commercial real estate space in a market. In addition, “supply” can refer to the expected completions of construction projects now underway. For clarity, this paper will refer to the current stock of commercial real estate as “stock” and to deliveries as “supply.” CoStar tracks construction start and end dates for every property in the dataset, and can use the construction dates to create full supply histories, going back decades. Exhibit 9 presents annual change in the supply of office space for the Los Angeles market.

19

Exhibit 9: Los Angeles Office Supply Since 1982

To arrive at the amount of supply (i.e., new deliveries) over time, we need only add up the properties delivered in that period. Note that the accuracy of this measure decreases over time, since properties demolished prior to the date at which CoStar begin tracking the market will not appear in the supply data, and as a result the CoStar data may understate the total completions in the distant past.

4.2 VACANCY AND DEMAND CoStar defines vacancy as space that is not physically occupied by a tenant. CoStar uses the term “availability” to denote space available for lease; a space can be available but not vacant if the landlord has begun marketing the space in anticipation of the current tenant moving out. “Demand” denotes the total occupied space in a market, and the vacancy rate as 1 – (Demand / Stock). CoStar has tracked the vacancy and availability of commercial real estate for nearly 30 years. At any given time, CoStar is tracking more than 600,000 individual spaces for lease; at regular intervals, a researcher contacts each listing agent or landlord to check on the status of the listing, including the current asking rent and whether a tenant is physically occupying the space. These agents and landlords usually have an incentive to provide information, since CoStar provides a platform to market space. Many of them have received calls from CoStar research for years, and many have developed a personal relationship with their CoStar research. From these research efforts, CoStar can establish the vacancy and availability rate for every commercial real estate asset at any point in the recent past, as well as when each space has leased. For multifamily properties, lease terms are shorter, and a property has as many tenants as units (or more, as some tenants choose to live together as roommates), resulting in more turnover and more volatility in vacancy rates. Seasonality also affects multifamily

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

WASH

Source: CoStar As of October 2017

(1M)

0M

1M

2M

3M

4M

5M

6M

7M

82 84 86 88 90 92 94 96 98 00 02 04 06 08 10 12 14 16

Quarterly Net Change in Total Stock (millions of SF)

20

vacancy. Because of this, CoStar tracks multifamily vacancy on a monthly cycle, rather than the quarterly cycle used for commercial properties (and competitors in the multifamily space). Furthermore, we cannot safely assume that, in the absence of information, a multifamily property is fully leased, nor can CoStar research teams examine tenant lists or lobby directories. These factors make tracking multifamily occupancy more difficult and the results less precise. However, these challenges affect all firms in the business of collecting and analyzing multifamily data. More data is the only solution, and the sheer scale of CoStar’s data collection apparatus results in a much larger and more accurate dataset than our competitors.

4.2.1 Commercial Vacancy and Demand

CoStar aggregates the building-level tenancy data to arrive at submarket- and market-level vacant space. Combined with the supply figures, we can derive the vacancy rate. Exhibit 10 presents office supply, demand, and vacancy for Los Angeles office.

Exhibit 10: Office Fundamentals for Los Angeles Office

CoStar’s building-level vacancy data begins around 2000 for most major metros; a handful have longer histories, including Atlanta, Dallas-Fort Worth, Washington, New York, Chicago, Los Angeles, and Baltimore, while coverage of most smaller markets (and retail) began in 2006. Prior to these dates, CoStar backcasts major markets using broad market trends collected by Property & Portfolio Research, a consulting and forecasting firm acquired by CoStar in 2009. Exhibit 2 (above) shows the date when the fundamentals time series begins for each major market.

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA

LOSA


0%

2%

4%

6%

8%

10%

12%

14%

(6M)

(4M)

(2M)

0M

2M

4M

6M

8M

96 98 00 02 04 06 08 10 12 14 16

Quarterly Change in Total Stock (Supply) Quarterly Change in Occupied Space (Demand) Vacancy

Quarterly Net Change in Total Stock (millions of SF) Vacancy Rate

21

4.2.2 Multifamily Vacancy and Demand

On average, CoStar researchers collect about 15,000 vacancy data points each month. In total, CoStar has at least one vacancy observation for more than 170,000 properties, and at least five data points for 130,000 properties. For perspective, a competitor to CoStar’s multifamily analytics collects data on about 35,000 properties in total. In total, CoStar’s vacancy data includes more than 15 million observations as of Summer 2020, at the time of writing. This substantial research effort, however, falls short of collecting vacancy data every month for every configuration in every property—more than 100 million data points in total. To fully populate the time series and provide a vacancy in every month for every configuration in every multifamily property, we must turn to statistical methods to estimate some missing data.

4.2.2.1 Filling in Missing Multifamily Data

The statistical process to fill in missing multifamily vacancy data includes two steps: First, fill in missing data points in properties for which we have vacancy data (about 170,000 properties out of the 450,000 total multifamily properties in the dataset, and in 83,000 out of 122,000 properties with at least 50 units). Second, we will estimate vacancy rates in the properties for which we have never collected vacancy data. Step one—filling in missing data points—takes two forms. Where a gap exists between observations, we linearly interpolate (i.e., draw a straight line) between the two observations. To populate the data before the first observation and after the most recent observation, we draw from the trend in stabilized vacancy rates from the same bedroom type in the same submarket. Exhibit 11 illustrates this process for a real property.

22

Exhibit 11: Creating the Multifamily Vacancy Time Series

Step two requires estimating the vacancy rate at each point in time for properties for which CoStar has never collected a vacancy observation. As noted, about 280,000 multifamily properties in the CoStar dataset do not have any vacancy data. This apparent shortcoming in data collection in fact underscores CoStar’s commitment to researching every property, regardless of its size: These properties tend to be quite small, and often do not have a leasing manager or landlord whom a CoStar researcher could call. To estimate the vacancy in properties for which we have never collected vacancy information, we apply the average vacancy from other stabilized properties in the same submarket with similar characteristics. Filling in these missing properties aligns the supply dataset with the vacancy dataset, so that we can accurately estimate aggregate demand.

4.2.2.2 Estimating Multifamily Lease-Up

The methods outlined above apply to stabilized multifamily properties. Recently delivered properties in lease-up exhibit different vacancy dynamics as they lease up. CoStar’s research and analyst teams prioritize collecting data on recently delivered properties to ensure the most accurate trends in lease-up. However, for the purposes of estimating lease-up in projects that delivered in the past and to fill in any missing data points in recent deliveries, we have developed an algorithm to estimate lease-up trends. Exhibit 12 presents an analysis of vacancy observations for properties completed since 2014, where the x-axis indicates months since the delivery date. The data shows that properties are about 55% leased in the first month, and leasing more or less follows an exponential curve eventually approaching the market average.

PropertyID

9489515

9489515

9489515

9489515

9489515

9489515

9489515

9489515

4/30/2011 0:00

5/31/2011 0:00

6/30/2011 0:00

7/31/2011 0:00

8/31/2011 0:00

9/30/2011 0:00

10/31/2011 0:00

11/30/2011 0:00

12/31/2011 0:00

1/31/2012 0:00

2/29/2012 0:00

3/31/2012 0:00

4/30/2012 0:00

5/31/2012 0:00

6/30/2012 0:00

7/31/2012 0:00

Source: CoStar. PropertyID 9489515 As of December 2017

0

5

10

15

20

25

2010 2011 2012 2013 2014 2015 2016 2017

Estimated Vacancy Vacancy Observations

Number of Vacant Units

Vacancy follows market trend prior to first observation

Straight-line interpolation between observations

23

Exhibit 12: Multifamily Lease-Up Trends

We apply a function approximating this shape where we need to estimate vacancy in a recently delivered property. We estimate vacancy in the first month after delivery to be midway between the stabilized vacancy rate in the market and 100%—a number typically around 53%. We estimate that vacancy will decline following the curve in Exhibit 12 over the next 18 months converging to the market rate. We also apply this algorithm for properties in lease-up that have some real data, filling in any gaps and trending the vacancy rates before the first observation and after the most recent using this same equation.

4.2.2.3 Backcasting Multifamily Data

CoStar’s collection of multifamily data began in 2012. In addition to filling in missing data, we also extend the vacancy series for multifamily properties using trends based on data from the United States Census Bureau, which tracks rental vacancy rates for the 75 largest MSAs. CoStar also acquired the RealFacts dataset, which includes rent and vacancy data for approximately 12,000 properties with data extending back to the 1990s. We use these data to estimate longer time series. Finally, we draw on the metro-level apartment rent trends collected by Property & Portfolio Research since the early 1990s. We apply these trends to each individual property to extend all multifamily series back to 2000.

4.2.2.4 National Multifamily Supply, Demand, and Vacancy

By implementing the algorithms described above, we have created full vacancy series for all 450,000 multifamily properties in the CoStar data set. These series make full use of the millions of vacancy data points across nearly 200,000 properties to estimate vacancy trends and levels, filling in gaps between observations and providing

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31


0%

10%

20%

30%

40%

50%

60%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Vacancy Rate By Month Since Delivery

24

approximate vacancy levels for property for which we have not collected any information. Creating a full time series for every property allows us to present high-quality estimates of aggregate supply and demand, either at a national or market level. Exhibit 13 presents our estimate of national multifamily supply, demand, and vacancy based on all 450,000 properties and 11 million units in the dataset.

Exhibit 13: National Multifamily Supply, Demand, and Vacancy

The increasing period-to-period volatility since 2012 indicates the improving quality of CoStar’s data. Prior to 2012, the series draws on broad, metro-level trends from the U.S. Census and legacy Property & Portfolio Research data. After 2012, the series is increasingly based on high-frequency, high-volume data collected by CoStar.

25

4.3 COMMERCIAL RENTS As part of CoStar’s daily data gathering, researchers collect the asking rent for available spaces in commercial properties. In total, the dataset contains more than 3.5 million asking rent observations across the industrial, office, and retail property types. Exhibit 14 presents graphically the scale of CoStar’s commercial rent data for 4 & 5 Star office properties in Boston’s Financial District submarket. Each line indicates the asking rent for a particular space in a property.

Exhibit 14: Boston Financial District 4&5-Star Office Rents

Exhibit 15 adds the available space-weighted average line to the graph. This view effectively shows how much per square foot a tenant would pay were they to rent all the space in the market in a particular quarter.

Souce: CoStar As of November 2017

$0

$10

$20

$30

$40

$50

$60

$70

$80

$90

$100

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Asking Rent

26

Exhibit 15: Boston Financial District 4&5-Star Available Space-Weighted Rent

4.3.1 Problems with Commercial Rents

This view of the space market clearly has some flaws. While relatively stable from 2000 to 2007, thereafter the aggregate line exhibits wild fluctuations, particularly in the most recent period. These fluctuations have little to do with actual market forces; rather, they result from changes in the set of spaces on the market. Exhibit 16 presents an attribution of movements in the aggregate series.

Exhibit 16: Attribution of Aggregate Rent Movements, Boston Financial District

The columns provide the reasons for the rent movements. The dark blue shows the aggregate rent movements directly attributable to observed changes in the underlying rent data. The decline in rents from 2001 to 2004, for example, resulted from observed cases where properties actually lowered asking rents on spaces available for lease.


$0

$10

$20

$30

$40

$50

$60

$70

$80

$90

$100

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Asking Rent

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

Source: CoStar As of January 2018

($15)

($10)

($5)

$0

$5

$10

$15

$20

$0

$10

$20

$30

$40

$50

$60

$70

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Due to: Change in Rent Change in Size of Space New Rent Observation

Dropped Rent Observation Combination of Factors Space-Weighted Rent

Asking Rent Effect on Aggregate Asking Rent Series

27

The lighter blue color, on the other hand, indicates aggregate rent movements that resulted from rents disappearing from the dataset, either due to the space being leased, withdrawn, or the listing manager no longer choosing to report a rent. In the first half of 2008, for example, rents fell about 10% simply because the dataset included fewer high rents. This same dynamic explains the large recent swings: Rents fall when expensive space goes off the market, and rents rise when cheaper space goes off the market. The dark gray color shows the effect of new rents entering the dataset, either due to new space on the market or to properties reporting a rent on an existing space which they had previously withheld. When higher rents enter the time series, as at the end of 2015, the aggregate rent series rises. On the other hand, if lower rents appear, as in 2017, rents will fall. The orange columns show the effect of changes in the amount of space offered. For example, if a property had 100,000 SF for rent at $100, and successfully leased 50,000 SF, then the weight on the $100 rent would fall by half, causing a decline in rents. The remaining effect, classified as “Combination of Factors,” shows rent movements that do not have a single cause. In these cases, both the size of the space and the asking rent have changed, making it impossible to attribute the movement to a single effect. Exhibit 15, above, shows why an available space-weighted average of the asking rents can present perverse trends that have little to do with actual market forces. In fact, in the extreme, these dynamics can cause the aggregate available-space weighted series to show the exact opposite of the expected result. For example, consider the East Cambridge/Kendall Square submarket, just across the Charles River from Boston’s Financial District. A leading center for technology and pharmaceutical research, the Kendall Square submarket has seen vacancies fall from 12% in 2012 to just 2% in 2018. Available space-weighted rents, however, have fallen over the same time, as shown in Exhibit 17 (the spike in the second half of 2017 notwithstanding). As seen in the graph, the loss of high rent observations (shown in the light blue columns), either due to lease-up or properties withholding rents, and changes in the amount of space (in orange) account for much of the decline—dynamics that have nothing to do with actual rent movements.

28

Exhibit 17: Attribution of Aggregate Rent Movements, East Cambridge/Kendall Square Submarket

Houston’s office market shows the opposite trend. Vacancies in the market rose from 11% in 2014 to 16% in 2017 as energy prices crashed. However, the available space-weighted rents shown in Exhibit 18 present a relatively healthy rental market, with rents flat since 2017—largely as a result of lower rents dropping out of the dataset, as indicated by the light blue columns.

Exhibit 18: Attribution of Aggregate Rent Movements, Houston Office Market

These perverse dynamics can distort even the most prestigious office markets. Exhibit 19 shows the available space-weighted rent for the Plaza District in New York City. The sharp decline in 2009Q4 resulted from a number of the most expensive properties choosing to no longer report a rent to CoStar researchers.

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519


($15)

($10)

($5)

$0

$5

$10

$15

$20

$0

$10

$20

$30

$40

$50

$60

$70

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017




14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14

14


($1)

($1)

$0

$1

$1

$2

$0

$5

$10

$15

$20

$25

$30

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017




29

Exhibit 19: Attribution of Aggregate Rent Movements, New York Plaza District

Finally, scarce data in small markets (or submarkets in large markets) can produce series that can resemble a random walk, as in the example of the Scranton industrial market, shown below in Exhibit 20.

Exhibit 20: Attribution of Aggregate Rent Movements, Scranton Industrial

These problems do not indicate any underlying deficiencies with the CoStar data. No other firm has more asking rent information. Rather, the series shown above indicate that we need a more sophisticated approach to using these data—an approach that can control for the changing composition of the set of properties reporting rents.

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

1528

200

50

200

3


($30)

($25)

($20)

($15)

($10)

($5)

$0

$5

$10

$0

$20

$40

$60

$80

$100

$120

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017




42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

42540

200

50

200

3


($2.00)

($1.50)

($1.00)

($0.50)

$0.00

$0.50

$1.00

$1.50

$2.00

$0

$1

$2

$3

$4

$5

$6

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017




30

4.3.2 Standardizing Service Types

In the United States, we face another problem unique to office properties: Landlords report a myriad of different rent terms—triple net, full service, plus electric, and so on. To properly capture rent trends and levels, we must try to normalize the various service types. Exhibit 21 presents the different categories, along with the frequency of each service type in the CoStar space-for-lease database as of December 2017:

Exhibit 21: Office Rent Service Types as of December 2017

Service Type Percentage

Full Service Gross 33.2%

Modified Gross 20.1%

Triple Net 19.8%

Negotiable 13.8%

Plus Electric 5.7%

Plus All Utilities 2.4%

Net 1.6%

TBD 1.3%

Tenant Electric 1.0%

+ Elec & Clean 0.3%

Plus Cleaning 0.3%

Utilities & Char 0.2%

Double Net 0.2%

Service types vary across the country, but also within markets, submarkets, and even properties, making accurate analyses of rent growths and levels difficult. To correct this problem, we must try to standardize rents by adding estimates of the netted-out expenses to each asking rent. For example, we would add an estimate of annual electric costs per square foot to all “plus electric” rents. But what are annual electric expenses in, say, New York? CoStar’s own dataset includes expense data for about 50,000 properties (as well as tax data on nearly every property), and CoStar researchers continue to add expense data wherever and whenever possible. But CoStar’s own data does not amount to a statistically viable sample of expense, both due to a relatively small sample size (especially at a market or submarket level) and to a geographic and quality bias: The expense data skews toward large, high-quality assets in major markets. Fortunately, a number of other industry sources provide expense data for office properties, including the Institute of Real Estate Management the National Council of Real Estate Fiduciaries (NCREIF), and public CMBS data. CoStar has used these three data sources, as well as our own data, to estimate typical expenses by building quality and location. The Institute of Real Estate Management’s (IREM) Income/Expense provides anonymized building-level financial data across a wide range of income and expense

31

categories for more than 9,000 properties. CoStar has acquired this full dataset for use in estimating expenses. NCREIF also provides income and expense data. Per NCREIF’s policies, it requires a minimum number of properties to present aggregated data; using the custom query tool, CoStar has downloaded the expense data first at a zip code level, then at a CBSA and CBD/suburban level; then at a state and CBD/suburban level, and finally at a regional and CBD/suburban level. CoStar also draws on publically available financial data for the properties which serve as collateral in CMBS pools, a dataset of more than 1,800 properties. The CMBS data adds another dimension to CoStar’s pool of expense information, as it includes mostly smaller assets located outside of major markets. We combine the information from these four sources to estimate expenses across the relevant categories at a zip code level, a submarket level, a submarket-cluster level, a CBSA and CBD/suburban level, a state and CBD/suburban level, and a regional and CBD/suburban level, further slicing each aggregate by quality: 1 & 2 Star, 3 Star, and 4 & 5 Star. We assign to each property the expense estimate from the lowest geographic level that meets statistical confidence tests. In general, properties in major markets use zip code or submarket-level expense estimates, while properties located outside of major metros tend to use market, state, or region-level estimates. These spot estimates of expenses are based on data as of 2013. To create time series of expenses, we apply same-store trends based on NCREIF data at a regional and CBD/suburban level for each expense category. Exhibit 22 shows CoStar’s estimated expenses by category for all office properties in the New York market.

32

Exhibit 22: New York Plaza District Office Expense Estimates

Note that the “Other” category includes management and administration, services, and miscellaneous expenditures. We can then add these estimated expenses to office rents depending on the service type to “gross up” the various net rent categories, allowing us to compare rents across properties and markets on a full-service basis.2 We now have all the raw material necessary to construct same-store rent series.

2 Note that the other commercial property types are much more consistent in the types of rents reported. Most retail rents in the CoStar dataset are triple net. Industrial rent service types do vary, but typically not within a market. CoStar does not at present offer estimates of industrial or retail expenses.

33

4.3.3 Commercial Same-Store Rents

Estimating same-store rent series for commercial properties involves creating full time series for every property, so that we can aggregate properties without regard to whether they have a rent or not. Creating full time series requires us to make three types of estimations: First, we must estimate rents between known observations. Second, we must estimate rents before the first known observation and after the most recent. Third, we must estimate rents in properties for which CoStar does not include any rent information. To estimate rents between, before, and after known rents, we will assume that rents in the subject property would have followed the same trend as rents in nearby, similar properties. We construct these trends from observed rent changes for individual properties. For office properties, we also consider the floor, so as to not mistake a rent in a more desirable floor for an increase. For example, we would not want to mistake the appearance of a top-floor rent in the Empire State Building at $100 per square foot as a large increase over a $20 lower-floor rent in the preceding period.

Exhibit 23 shows the raw asking rent data at the space level for 4 & 5 Star properties in Boston’s Financial District, along with the trendline derived from these data. Note that for the purposes of creating the trend line, we only include observations that show a change (either positive or negative) in order to increase the volatility of the trend lines. The trend lines also include changes between non-adjacent rent observations in the same property, assuming a constant growth rate between the non-adjacent observations.

Exhibit 23: Boston Financial District 4&5-Star Office Rents with Trendline

We can now apply these trend lines to individual properties. To fill in the gaps between rents, we pivot the submarket trend line by adding a constant equal to the quarterly difference in total growth between the trend line and the rents at the start and end of the


$0

$10

$20

$30

$40

$50

$60

$70

$80

$90

$100

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Asking Rent

34

gap. We extend rents before the first observation and after the last observation by applying the rates of change from the trend line. Exhibit 24 illustrates how we apply the trend lines to create a full time series for a particular property.

Exhibit 24: Construction of Same-Store Rent Series for Specific Property

At this point, every property for which CoStar has at least one rent observation will have a full same-store time series dating back to at least 2000 for office and industrial, and to at least 2006 for retail (the date at which CoStar began covering the retail sector). Markets in which CoStar has collected data for a longer period will have earlier start dates. The resulting aggregate series, however, understate the actual volatility in rents. This is because the process described above assumes that, for the purpose of creating the trend lines, rent growth between non-adjacent rent observations is constant over time. In fact, rent movements between these points would more likely have followed the market trend. We can correct this problem by recreating the trend lines from the now-complete building level data. Whereas the initial trend line construction assumed a constant rate of growth between non-contiguous rent observations, this second iteration uses the shape of the trend line. By repeating this process, the results eventually converge to a true same-store view of the market, based on the observed changes in real rents. Exhibit 25 illustrates how the trend line for London’s Mayfair submarket changes with each additional iteration of the process, with the solid black line indicating the final series.

3/31/2000

6/30/2000

9/30/2000

########

3/31/2001

6/30/2001

9/30/2001

########

3/31/2002

6/30/2002

9/30/2002

########

3/31/2003

6/30/2003

9/30/2003

########

3/31/2004

6/30/2004

9/30/2004

########

3/31/2005

6/30/2005

9/30/2005

########

3/31/2006

6/30/2006

9/30/2006

########

3/31/2007

6/30/2007

9/30/2007

Souce: CoStar: PropertyID 7457537 As of November 2017

£0

£5

£10

£15

£20

£25

£30

£35

£40

£45

£50

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Trendline Observed Asking Rents

Asking Rent

35

Exhibit 25: Evolution of London Mayfair Trendline

Each iteration applies a more and more rounded trend line, enhancing the shape of the aggregate line and eventually converging to a line that most nearly approximates the movements in real rents. We can represent this process mathematically as follows: First, we calculate growth rates as

𝐺𝑟𝑜𝑤𝑡ℎ0,𝑡 =𝑅𝑒𝑛𝑡𝐵,𝑡

𝑅𝑒𝑛𝑡𝐵,𝑡−1 where 𝑅𝑒𝑛𝑡𝐵,𝑡 ≠ 𝑅𝑒𝑛𝑡𝐵,𝑡−1

Where 𝐺𝑟𝑜𝑤𝑡ℎ0,𝑡 denotes the initial calculation of growth, and 𝑅𝑒𝑛𝑡𝐵,𝑡 denotes the rent

in building 𝐵 at time 𝑡. For the purposes of calculating growth, we exclude rents that do not change from period 𝑡 − 1 to period 𝑡 in order to increase the volatility of the index. Using growth rates 𝐺𝑟𝑜𝑤𝑡ℎ0..𝑛 we can create an initial index 𝐼𝑛𝑑𝑒𝑥0,𝑡 as:

𝐼𝑛𝑑𝑒𝑥0,𝑡 = 100 𝛱0

𝑡(1 + 𝐺𝑟𝑜𝑤𝑡ℎ0,𝑡)

We can now use this index to fill in missing point between rent observations for a given building 𝐵, where building 𝐵 has an observed rent at times 𝑎 and 𝑏 but not at time 𝑐. Where 𝑎 < 𝑐 < 𝑏,

𝐸𝑠𝑡𝑅𝑒𝑛𝑡𝐵,𝑐,1 =𝐼𝑛𝑑𝑒𝑥0,𝑐

𝐼𝑛𝑑𝑒𝑥0,𝑎 (

𝑅𝑒𝑛𝑡𝐵,𝑐

𝑅𝑒𝑛𝑡𝐵,𝑎)

𝑐−𝑎

𝑏−𝑎

Where 𝐸𝑠𝑡𝑅𝑒𝑛𝑡𝐵,𝑐,1 denotes the estimated rent in building 𝐵 at time 𝑐.

Now we repeat this process 𝑛 times, where growth is based on both the actual rents and the estimated rents from the equation above, maintaining the condition that, for the

6/30/2003 0:00

9/30/2003 0:00

12/31/2003 0:00

3/31/2004 0:00

6/30/2004 0:00

9/30/2004 0:00

12/31/2004 0:00

3/31/2005 0:00

6/30/2005 0:00

9/30/2005 0:00

12/31/2005 0:00

3/31/2006 0:00

6/30/2006 0:00

9/30/2006 0:00

12/31/2006 0:00

3/31/2007 0:00

Source: CoStar

£40

£45

£50

£55

£60

£65

£70

£75

£80

£85

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16

Rent Level

36

purposes of creating a rent index, we exclude flat rents to maximize the volatility of the series:

𝐺𝑟𝑜𝑤𝑡ℎ0,𝑡 =𝑅𝑒𝑛𝑡𝐵,𝑡,𝑛

𝑅𝑒𝑛𝑡𝐵,𝑡−1,𝑛 where 𝑅𝑒𝑛𝑡𝐵,𝑡,𝑛 ≠ 𝑅𝑒𝑛𝑡𝐵,𝑡−1,𝑛

Then recalculate the index: 𝐼𝑛𝑑𝑒𝑥𝑛,𝑡 = 100 𝛱𝑛

𝑡 (1 + 𝐺𝑟𝑜𝑤𝑡ℎ𝑛,𝑡)

And reestimate the individual building rents:

𝐸𝑠𝑡𝑅𝑒𝑛𝑡𝐵,𝑐,𝑛 =𝐼𝑛𝑑𝑒𝑥𝑛−1,𝑐

𝐼𝑛𝑑𝑒𝑥𝑛−1,𝑎 (

𝑅𝑒𝑛𝑡𝐵,𝑐

𝑅𝑒𝑛𝑡𝐵,𝑎)

𝑐−𝑎

𝑏−𝑎

4.3.4 Estimating Rents in Properties Without Rent Information

This process produces a full time series for every property for which we have a rent. But what about the properties without rents? Of the 680,000 office properties in the CoStar dataset, we have at least one rent observation for about 300,000. Properties for which CoStar has at least one rent observation tend to be larger (26,000 SF on average, compared with 14,000 for properties without a rent), newer (44 years old on average, compared with 55 years for properties without a rent), and higher rated (averaging 2.3 stars, compared with 2.1 for properties without a rent). Due to these differences, we can assume that properties that have a reported a rent tend to be more expensive, and as a result, omitting the sample properties without rents will likely cause us to overstate the true rent in a particular market. To correct this bias, we can estimate rents in properties for which we have never had a rent. To do so, we calculate the average current rent by submarket, space type, and quality. In submarkets with limited data, we draw from larger geographies. We then apply the corresponding submarket trend line to this estimate of current rent to create a full time series. Exhibit 26 presents national series for properties with rents, those without, and the aggregate series.

37

Exhibit 26: Aggregate Same-Store Office Rent Series Comparison

We see from Exhibit 26 that properties for which CoStar does not have a rent do likely have lower rent levels, such that reporting only the known rents likely exaggerates the aggregate rent figures. This difference is stark is some markets: For example, Exhibit 27 shows the same analysis for New York City office properties, where properties that have reported a rent are about $10 more expensive than properties that have never reported a rent to CoStar.

Exhibit 27: Aggregate Same-Store New York Office Rent Series Comparison

QuarterEndDate

3/31/2000 0:00

6/30/2000 0:00

9/30/2000 0:00

12/31/2000 0:00

3/31/2001 0:00

6/30/2001 0:00

9/30/2001 0:00

12/31/2001 0:00

3/31/2002 0:00

6/30/2002 0:00

9/30/2002 0:00

12/31/2002 0:00

3/31/2003 0:00

6/30/2003 0:00

9/30/2003 0:00

12/31/2003 0:00

3/31/2004 0:00

6/30/2004 0:00

9/30/2004 0:00

12/31/2004 0:00

3/31/2005 0:00

6/30/2005 0:00

9/30/2005 0:00

12/31/2005 0:00

3/31/2006 0:00

6/30/2006 0:00

9/30/2006 0:00

12/31/2006 0:00

3/31/2007 0:00

6/30/2007 0:00

Source: CoStar As of March 2018

$0

$5

$10

$15

$20

$25

$30

$35

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Properties with Real Rents Properties without Real Rents TotalRent

Full Service Asking Rent

QuarterEndDate

3/31/2000 0:00

6/30/2000 0:00

9/30/2000 0:00

12/31/2000 0:00

3/31/2001 0:00

6/30/2001 0:00

9/30/2001 0:00

12/31/2001 0:00

3/31/2002 0:00

6/30/2002 0:00

9/30/2002 0:00

12/31/2002 0:00

3/31/2003 0:00

6/30/2003 0:00

9/30/2003 0:00

12/31/2003 0:00

3/31/2004 0:00

6/30/2004 0:00

9/30/2004 0:00

12/31/2004 0:00

3/31/2005 0:00

6/30/2005 0:00

9/30/2005 0:00

12/31/2005 0:00

3/31/2006 0:00

6/30/2006 0:00

9/30/2006 0:00

12/31/2006 0:00

3/31/2007 0:00

6/30/2007 0:00


$0

$10

$20

$30

$40

$50

$60

$70

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Properties with Real Rents Properties without Real Rents TotalRent


38

4.3.5 Comparing Same-Store Rents to Available Space-Weighted Rents

At this point, we have created a full time series for every property in the dataset. For properties with some actual rent information, we have filled in gaps using market trends. For properties without any real rent information, we have estimated the rent level and followed the market trend. Aggregating the property-level data produces series that fully control for the changing composition of the marketplace, as well as the bias toward properties that report rents. Let us examine how these aggregate same-store series compare with the available space-weighted series presented above, starting with Boston’s Financial District Submarket in Exhibit 28:

Exhibit 28: Same-Store v. Available Space-Weighted Rents, Boston Financial District

The same-store approach nicely smooths out the quarter-to-quarter volatility, presenting a more useful view of rent trends in the submarket. In East Cambridge/Kendall Square, the same-store series shows a different picture a steady increase dating back to 2010—much more in line with expectations, given the strength of this market, as seen below in Exhibit 29.

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675

675


$0

$10

$20

$30

$40

$50

$60

$70

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Space-Weighted Rent Same-Store Rent


39

Exhibit 29: Same-Store v. Available Space-Weighted Rents, E Cambridge/Kendall Sq. Submarket

Applying the same-store approach to Houston also produces a different picture, one of falling rents over the past several years in contrast to the flat trend depicted from the available-space weighted rents, as shown below in Exhibit 30.

Exhibit 30: Same-Store v. Available Space-Weighted Rents, Houston Office Market

In New York’s Plaza District, the same-store approach eliminates the sharp decline in 2009 that resulted from many large office landlords no longer listing asking rents. Same-store rents also show steady increases from 2010 to 2015, and a more recent flattening of rents, as shown in Exhibit 31:

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519

519


$0

$10

$20

$30

$40

$50

$60

$70

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17




$0

$5

$10

$15

$20

$25

$30

$35

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17



40

Exhibit 31: Same-Store v. Available Space-Weighted Rents, New York Plaza District Submarket

Finally, the same-store series for the Scranton industrial market show a much smoother series and a more accurate portrayal of the rent level in Exhibit 32, by virtue of including all properties in the market:

Exhibit 32: Same-Store v. Available Space-Weighted Rents, Scranton Industrial


$0

$20

$40

$60

$80

$100

$120

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17




$0.00

$0.50

$1.00

$1.50

$2.00

$2.50

$3.00

$3.50

$4.00

$4.50

$5.00

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17



41

4.4 MULTIFAMILY RENTS CoStar’s multifamily rent data differs both in degree and kind from the commercial rent data. Whereas CoStar’s dataset contains some 4 million asking rent observations for commercial properties, the multifamily rent dataset exceeds two billion data points, collected via various channels:

- CoStar’s traditional research teams today focus on collecting vacancy and concession data, but historically have collected rent data since the inauguration of CoStar’s apartment service in 2012

- Direct feeds and user entered data from clients of CoStar’s ILS platforms, including Apartments.com, ApartmentFinder.com, and ForRent.com

- Automated data collection algorithms that “scrape” data from property websites

- CoStar acquired the RealFacts dataset of building-level rent and vacancy data dating back to the mid-1990s for more than 12,000 properties. We use this information to estimate a longer time series back to 2000.

Exhibit 33 shows the rent data for one-bedroom units within a single property. The different shapes indicate the various data source, while the colors represent different one-bedroom models, varying by size and layout.

Exhibit 33: Multifamily Rent Data

The high-frequency, detail, and sheer size of CoStar’s multifamily rent data provide us with all the information necessary to present the most accurate and timely view of apartment rent trends. In fact, we produce a monthly series of multifamily rents back to 2000—and a daily series dating back to the beginning of 2015. Exhibit 34 presents daily rents from the CoStar product for the same property presented above:

select top 1000 *

from EnterpriseSub.dbo.ApartmentRent

where ObservationDate < '12/31/2007'

select

from

left join

left join

where

order by

Source: CoStar (PropertyID 1337935) As of January 2018

$0

$200

$400

$600

$800

$1,000

$1,200

$1,400

$1,600

$1,800

02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Asking Rent per Unit

RealFacts Data

Apartments.com Data

42

Exhibit 34: Multifamily Daily Rents

The multifamily rent data also present challenges, however. As seen in Exhibit 33, not all models have rent information at all points in time; property managers don’t advertise unavailable units. Moreover, CoStar may collect rent observations for the same one-bedroom model at the same time from several different sources—from Apartments.com, from an automated scrape of the community website, and from a CoStar researcher or analyst—and may receive different responses. Which observation is “correct”? Finally, CoStar may have never collected rent data for particular units within a community; for example, a property may have a handful of four-bedroom units that have never become available. What is the rent in these units?

4.4.1 Same-Store Multifamily Rents

To solve these problems, we create a full time series back to 2000 for every model, in every property for which we have ever collected a rent observation. Creating full time series for every property ensures that rent trends only reflect market-driven movements, rather than changes in the set of properties and units available for rent. Creating the multifamily same-store rent series involves three steps: First, we select the set of rents to use. Next, we fill in rents between non-adjacent observations. Then, we extend the rent series prior to the first observation and after the most recent observation. Finally, we estimate rents for models that do not have any rent information. For the headline asking rent series, we give top preference to rent data supplied via Apartments.com direct feeds. These data match the rents advertised on CoStar ILS platforms, are updated frequently (typically daily), and comprise the majority of the CoStar rent dataset (more than 2 billion observations at the time of writing). Data providers can report a rent range; we use the minimum rent reported. The minimum rent is far more stable than the maximum, which can vary widely depending on the

select top 1000 *

from EnterpriseSub.dbo.ApartmentRent

where ObservationDate < '12/31/2007'

select

from

left join

left join

where

order by

Source: CoStar (PropertyID 1337935) As of January 2018

$0

$200

$400

$600

$800

$1,000

$1,200

$1,400

$1,600

$1,800

15 16 17 18

Asking Rent per Unit

43

specific terms of the lease (short-term, furnished units can have very high rents that do not reflect the broad market). In general, the minimum rent reflects the typical 12-month lease. Using the midpoint of the minimum and maximum would still allow half of the artificial volatility inherent in the maximum values to affect the series. About 50,000 properties provide data via automated feeds. For the remaining 150,000 or so properties for which we have rent data, CoStar gives preference to to data collected by CoStar researchers or analysts, then to user-entered rents on CoStar ILS sites, and finally to rents collected via automated data collection. Should a model begin to receive data of a higher priority, we reset the level. Properties on a feed do not revert to lower-priority data sources until they have not received a feed update for at least three months; in these cases we assume the property is no longer on a feed. At this point we reset the history for all time, using the old feed history to inform the trend. To fill in rents between non-adjacent observations, we use simple linear interpolation—simply drawing a straight line between the two observations. Out-of-sample testing showed simple interpolation to be as good as or better than more sophisticated approaches, and because most gaps in the multifamily data are short, the straight-line interpolation doesn’t confound broad trends. Where we are missing data, we assume that rents would have followed the same trend as rents for other models in the same property. In the absence of other property-level data, we assume that rents would have followed the same trend as the submarket and quality slice. Exhibit 35 shows the real and estimated data for an actual multifamily property.

Exhibit 35: Estimating Missing Rents

At this point, every property for which CoStar has ever collected any rent information has a full rent series. However, a small percentage of models in these properties do not have any rent data; typically these are larger three- or four-bedroom units that have

2014m122014m122014m122014m122014m122015m12015m12015m12015m12015m12015m22015m22015m22015m22015m22015m32015m32015m32015m32015m32015m42015m42015m42015m42015m42015m52015m52015m52015m52015m52015m6


$0.00

$0.20

$0.40

$0.60

$0.80

$1.00

$1.20

$1.40

12/14 3/15 6/15 9/15 1/16 4/16

Model entirely based on rent curvesMedian error: 5.9%

Rent / SF

Interpolated rent between two pointsMedian error: 2.1%

Extended Rent series, based on same-store chain-linked rent seriesMedian Error: 2.7%

44

never been available for rent. Thanks to the depth of the CoStar data, we can estimate these rents with a high degree of accuracy given the observed relationships between per-square-foot rents for various configurations. For example, we may observe that three-bedroom units rent for 25% less than one-bedroom units on a per square foot basis; if we know the rents for a one-bedroom in the same property, we can assume that the three-bedroom will be 25% less. Exhibit 36 shows how per-square-foot rents vary by configuration type.

Exhibit 36: National Rent Curve Matrix

The underlying assumption of this approach is that rents in different units in the same building will move parallel to each other, all else equal. This is equivalent to saying that the ratio of rents between unit types will hold constant. If true, the rents of a given configuration for which CoStar does not have data can be estimated using the known rents for other configurations in the same building, submarket, or market. For example, as shown in Exhibit 36, studio apartments are about 20% more expensive per square foot than one-bedroom apartments on average for 3 Star-rated buildings across the nation. Thus, given rents for a one-bedroom unit in a given building, one could simply assume that the studios in the same building are 20% higher. In reality, the situation is more complex. The relationship between different configurations might be drastically different in different metros, or even between different submarkets in the same metro. A building might have five, ten, or even more different layouts, and each layout might include a variety of rents related to the specific characteristics of each unit. Moreover, some combinations occur more often than others, and some combinations are rare. For example, we might observe the ratio of studio to one-bedroom rents, and the ratio of one-bedroom to two-bedroom rents, but might never observe the ratio of studios to two-bedrooms simply because those data points do not appear in the data set at the same time. Exhibit 37 presents a stylized view of all 55 potential combinations.

National

Source: CoStar As of 2015Q4

0.4

0.5

0.6

0.7

0.8

0.9

1.0

1.1

0/1

1/1

1/2

2/1

2/2

2/3

3/1

3/2

3/3

4/2

4/3

National: 3 Star National: 4 & 5 Star

45

Exhibit 37: Diagram of Possible Combinations of Configurations

Even worse, the ratios across different unit combinations might conflict. We might observe that studios are 20% more expensive than one-bedroom apartments, and that one-bedroom apartments are 5% more expensive than two-bedroom apartments, but that studios are 40% more expensive than two-bedroom apartments—a mathematical impossibility, but nonetheless present in the data because the observations come from different time periods and across thousands of buildings. Exhibit 38 illustrates this problem, using an actual property in the dataset.

Exhibit 38: Rent Average in a Building with Changing Unit Type History


$0.60

$0.70

$0.80

$0.90

$1.00

$1.10

$1.20

$1.30

$1.40

Dec-14 Mar-15 May-15 Jul-15 Sep-15

Studio 1 Bed 1 Bath 2 Bed 1 Bath

Rent in $/SF

20% 40%

5%

46

Resolving these complexities and contradictions requires a single reference point from which the rent for any configuration can be estimated: Let us call this variable µ. The actual level of µ is arbitrary, but most closely represents the average rent of a given building. For each configuration within the building, we estimate a coefficient that relates that configuration to µ. Thus, via µ, any unit type can be related to any other unit type. For instance, if the coefficient of µ for the one-bedroom, one-bathroom unit is 0.9, and the coefficient of µ for the studio is 1.2, then the studio rent must be 1.2/0.9 = 1.333 times the one-bedroom rent. Exhibit 39 illustrates the interactions between µ and the various configurations.

Exhibit 39: Illustration of Relationships Between Unit Types and µ

To arrive at µ, start with a random value for each ci. Given a set of coefficients ci and rents xi for all possible configurations i, we define: µ = mean(xi /ci) Our hypothesis for what xi should be is: hi = ci*µ The error for a set of coefficients is defined as:

J(c1,c2,…) = ∑ (hi-xi)2

i

47

and the partial derivatives of the cost function are:

Where m is the number of configurations for a particular geography. The cost function measures how closely a set of coefficients estimates the data. The initial random estimate is generally poor, but a gradient descent minimizes the error and determines the coefficients that are the best predictors. Run at the building level, submarket-quality level, metro-quality level, and national-quality level, the process allows users to make a prediction for rents in any building based on the most granular, comparable set possible, depending on the sample size. These coefficients can predict the rent of any unit within a building based on the rents of the other units in that building. There are a number of different potential applications of that information. First, we can estimate rents for units for which CoStar has never collected a rent observation, as shown in Exhibit 40. In this example, we know that the building does have two-bedroom, three-bathroom units—but we’ve never collected a rent observation for this unit type. Using the relationships observed between two-bedroom, three-bathroom units and all other units across similar buildings, µ allows us to estimate what the missing rents would be.

Exhibit 40: Rent Estimation Based on Other Unit Types

Second, it makes it possible to compare rents for buildings with different unit mixes. For instance, consider a building A, which consists primarily of two-bedroom units with some three-bedroom units, and next-door building B, primarily three-bedroom units with some two-bedroom units. Based on simple averages, building B would appear to have


$0.60

$0.70

$0.80

$0.90

$1.00

$1.10

$1.20

$1.30

$1.40

$1.50

$1.60

Dec-14 Mar-15 May-15 Jul-15 Sep-15 Nov-15 Jan-16

Studio 1 Bed 1 Bath 2 Bed 1 Bath 2 Bed 2 Bath 2 Bed 3 Bath Estimate

Rent in $/SF

Estimate for model without data

48

higher per-unit rents, while building A would have higher per-square-foot rents. But the differences are due to differences in unit types rather than actual market forces. We can use µ to estimate an equivalent rent, permitting direct comparisons between buildings (or markets) with systematically different unit mixes.

4.4.2 Concessions and Effective Rents

Of course, asking rents do not necessarily reflect that actual rent paid by a new tenant. Properties frequently offer incentives and concessions to attract new renters, particularly properties in lease-up. CoStar’s multi-family research team focuses in particular on collecting detailed information around various concessions offered by landlords: waived fees, gift cards, cash incentives, and months of free rent, to name a few. CoStar uses this rich dataset, including more than 10 million observations at the time of writing, to estimate concession rates for all properties over time. CoStar’s definition of a concession includes free or reduced rent, as well as cash payouts and gift cards. We do not include waived fees, as the data suggests that many fees exist so that property managers may use them in negotiations, offering to “waive” the fee in lieu of free rent or outright cash incentives. Including waived fees would not alter the estimate significantly in any case, as fees represent a very small fraction of typical rent. We apply an estimate of expenses to properties for which we do not have any expense data based on the age of a property and its vacancy rate. Newer and higher-vacancy properties will have higher estimated concessions, reflecting the tendency of newer properties, particularly those in lease-up, to offer concessions. The vast majority of properties, however, do not offer concessions. By our estimation, only a small fraction of properties offer some form of discount to the asking rent. As a result, in the aggregate the effective rent series is very close to the asking rent series.

49

5 PERFORMANCE METRICS

CoStar’s collection of transaction information has amassed a dataset of more than 4 million commercial real estate transactions, totaling more than $9 trillion and covering the entire universe from small offices selling for $10,000 to the largest portfolio deals. This vast dataset, coupled with the building-level data, offers unprecedented insight into the performance of commercial real estate investments. Unlike stocks or bonds, or other frequently-traded asset classes, no one really knows the value of a commercial real estate asset until it trades—and then only briefly. Owners rely on appraisers, comps, and gut instinct to value their assets and set prices. As a result, commercial real estate values are necessarily estimates. Given CoStar’s unmatched set of transaction data, no firm is better positioned to make estimates about individual property values and trends than CoStar. Our goal is to provide our clients with the best estimates of values and cap rates, both current and over time, for any market, submarket, or set of properties. Ultimately, achieving this goal necessitates estimating a value for every property in our dataset. We can then aggregate these property-level estimates to arrive at time series for a market, a submarket, the nation overall, or any custom set of properties that interest our clients. We understand that property-level estimates will be imperfect. We know that our clients will—and should—question them. But we also believe that, our approach, laid out in this paper, produces the most accurate, comprehensive, and unbiased view of levels and trends in prices and cap rates available to the industry. What follows is a detailed description of our approach for estimate values and cap rates for the more than 4 million properties in our dataset. We continue to refine the approach by systematic out-of-sample testing, and as a result the estimates and the historical time series will change; we will notify clients of any major changes to our methodology and will update this document accordingly.

5.1 APPROACHES TO TRACKING PROPERTY VALUES Land, and by extension the real estate on that land, constitutes the essential unit of capital. Given real estate’s centrality to the economy, the questions of how to value real estate, and track changes in value over time, have occupied economists for centuries. They are not the only interested party: Governments wishing to tax real estate need assessments of value, banks underwriting loans need estimates of the collateral’s worth, and property owners and investors consider valuations as they manage their portfolio and make buy/sell decisions. Such needs have created the $8 billion appraisal industry in the United States. In the United Kingdom, the industry has even earned the Queen’s imprimatur: the Royal Institution of Chartered Surveyors.

50

The ubiquity of appraisal-based valuations, compared to the relative scarcity of actual transaction prices, led to the creation of NCREIF (National Council of Real Estate Investment Fiduciaries) in the United States and IPD (Investment Property Databank, now part of MSCI) in the United Kingdom. These organizations collect property-level valuation data, based on periodic appraisals, from member investment funds, and report the information back to the industry in the form of anonymized indices. These indices have become essential benchmarks to measuring commercial real estate performance in the United States and the United Kingdom. Their datasets have also informed much academic work. Government statistical agencies have also produced valuable price indices. Perhaps most famously, Japan’s Urban Land Price Index depicted in real-time the 80% collapse in Tokyo land prices in the early 1990s. Both Singapore and Hong Kong also have maintained long time series of commercial and residential property prices, based on stamp tax reporting of real estate deals. The academic work around commercial real estate pricing has broadly followed two tracks: Approaches to aggregate price indices, and approaches to automated valuation models (AVMs). Formal studies of long-term price series likely began with Dr. Homer Hoyt’s seminal 1933 study, “One Hundred Years of Land Value in Chicago.” Dr. Hoyt drew on a wide variety of sources including assessments, transactions, and appraisals to build one of the first long-term series of commercial property prices in the United States. That he did so without the benefit of modern econometric concepts or computers makes his achievement all the more remarkable. In 1963, Drs. Martin Baily, Richard Muth, and Hugh Nourse laid out an econometric approach to constructing repeat-sale indices in their paper, “A Regression Method for Real Estate Price Indices.” Many others have built on this framework, including Dr. Daniel McMillen in his 1996 homage to Dr. Hoyt, “One Hundred and Fifty Years of Land Values in Chicago: A Nonparametric Approach.” Dr. William Wheaton, Mark Baranski, and Cesarina Templeton applied these ideas to New York in their 2007 paper, “100 Years of Commercial Real Estate Prices in Manhattan,” in which they applied a regression approach to 89 repeat trade pairs to construct a long-term time series for Manhattan office properties. Dr. David Geltner, along with Drs. Jeffrey Fisher and R. Brian Webb, wrote the first of many studies on real estate price indices in 1994 in their “Value Indices of Commercial Real Estate: A Comparison of Index Construction Methods,” which introduced a hedonic approach to constructing price indices using NCREIF data. Jason Barr, Fred Smith, and Sayali Kulkarni also drew on CoStar’s rich dataset on Manhattan land prices in their 2015 paper, “What’s Manhattan Worth: A Land Values Index from 1950 to 2013.”

51

Tom Nicholas and Anna Scherbina apply a hedonic approach to hand-collected transaction data from a weekly trade publication to analyze multifamily prices in New York City in the 1920s and 1930s in their 2013 paper, “Real Estate Prices During the Roaring Twenties and Great Depression.” Joseph Nichols, Stephen Oliner, and Michael Mulhall, all of the Federal Reserve Board, also applied a hedonic approach to CoStar data on land transactions to construct land price indices for 23 MSAs in the United States. Automated valuation models have received less attention from academics, due to the difficulty of obtaining quality information on real estate transactions and valuations. Much of the literature evaluates the accuracy of appraisals for properties that subsequently sold, using the NCREIF dataset, including Rebel Cole, David Guilkey, and Mike Miles’ 1986 study “Toward an Assessment of the Reliability of Commercial Appraisals”; Jeffrey Fisher, Mike Miles, and R Brian Webb’s 1999 study, “How Reliable Are Commercial Appraisals? Another Look”; and Susanne Cannon and Rebel Cole’s 2011 study, “How Accurate Are Commercial Real Estate Appraisals? Evidence from 25 Years of NCREIF Sales Data.” Only recently have academics focused on the prospects of automated valuation models. Nils Kok, Eija-Leena Koponen, and Camen Adriana Martinez-Barbosa apply a variety of “big data” techniques to a combination of CMBS and RCA data. We agree with these authors that such approaches could lend efficiency and transparency to this notoriously opaque asset class. Working at CoStar, we have at our disposal an unmatched dataset of transactions as well as property characteristics and conditions. Dr. Ruijue Peng, Dr. Ozlem Yanmaz-Tuzel, and Dr. Xiaojing Li built the first large-scale repeat-sale indices for commercial real estate out this dataset: the CoStar Commercial Repeat-Sales Indices (“CCRSI”). Monthly updates to these series, as well as a description of their methodology, may be found at http://www.costargroup.com/costar-news/ccrsi. The various attempts at constructing commercial real estate price series have focused on aggregate series, either for the entire nation, a property type, or a particular market. Our aim, however, extends beyond broad market trends. In keeping with our philosophy of providing property-level information, we wish to estimate a full time series for price and cap rate for every commercial and multifamily property in the CoStar dataset. If we can provide reasonable estimates of a particular property’s current values and how that value has changed over time, we believe we can also provide the truest picture of broad market trends by aggregating together the price and cap rate estimates for the hundreds of thousands of individual properties.

http://www.costargroup.com/costar-news/ccrsi

52

5.2 THE COSTAR TRANSACTION DATA CoStar its most talented and experienced researchers to its Comps team. This group is committed to collecting the highest quality information about every commercial real estate transaction that occurs, including the terms; the reported and the true buyer and seller; other deal participants; relevant sale conditions; and financing arrangements. Exhibit 41 presents the sales transaction information available on the CoStar product for 33 Arch Street, CoStar’s Boston office location.

Exhibit 41: CoStar Transaction Data for 33 Arch Street, Boston

CoStar’s comps researchers recorded more than 330,000 trades totaling more than $800 billion in 2019. Exhibit 42 presents CoStar’s comp data by year.

Exhibit 42: USD-Denominated Total Deal Volume, By Year

Year

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2018

Source: CoStar. As of March 2018

0

20,000

40,000

60,000

80,000

100,000

120,000

140,000

160,000

$0B

$100B

$200B

$300B

$400B

$500B

$600B

$700B

$800B

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Volume Number of Trades

Total USD Deal Volume Number of USD-Denominated Trades

53

Other firms, most notably Real Capital Analytics, track the commercial real estate transaction market. However, no other firm so thoroughly covers the entire market: In addition to capturing a census of large trades (more than 17,000 trades of more than $50 million), 90% of the trades in CoStar’s dataset are for less than $5 million. Exhibit 43 captures the scale of the CoStar data. Each dot represents a non-portfolio trade for an office asset in the United States of at least $1 million—more than 80,000 transactions in all. Each dot is shaded by its distance from the 180-day trailing average to highlight a general trend.

Exhibit 43: Office Trades in the United States

This vast dataset offers unmatched analytic potential for divining trends in commercial real estate pricing. Even a simple average of transaction prices provides a useful national series, as shown in Exhibit 44:

Exhibit 44: Average Annual Office Price per Square Foot

Source: CoStar. Prices are shaded by distance from trailing average to identify trend. As of March 2018Includes all non-portfolio sales of office assets selling for at least $1 million.

$0

$200

$400

$600

$800

$1,000

$1,200

$1,400

$1,600

$1,800

$2,000

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Transaction Price per Square Foot

200020012002200320042005200620072008200920102011201220132014201520162017


$0

$50

$100

$150

$200

$250

$300

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Quarterly Series Annual Series

Average Transaction Price per Square Foot

54

This simple approach works well enough in depicting broad, long-term trends at high levels of aggregation. However, as shown above, even a national series shows stochastic movements at higher frequencies, as in the quarterly series shown in the lighter gray. These movements result from changes in the composition of properties and markets that trade period to period. At lower levels of aggregation, these problems render the series unusable. For example, Exhibit 45 shows the simple average transaction price for the heavily traded New York office market. While the annual series (shown in the darker gray) shows a sensible long-term trend, the quarterly series exhibits extreme period-to-period volatility.

Exhibit 45: Average Annual Office Price per Square Foot, New York

The question of how to weight each trade further complicates the analysis of transaction data. The graphics shown above show simple averages, such that each transaction counts the same regardless of the size of the property or of the deal: A small dentist office in Queens has the same weight as a Park Avenue tower. Value-weighting, typically according to the size of the deal or the size of the property, usually results in higher average prices, but also adds volatility since it effectively reduces the sample size by making small trades irrelevant. (Weighting by transaction value means that a $1 billion trade will count 1,000 times more than a $1 million trade). Value-weighting also exacerbates the problem of changing composition, with periods in which “blockbuster” trades occur showing much higher average prices than in periods without a “blockbuster” trade. Exhibit 46 shows the value-weighted average transaction price for the New York office market.

200020012002200320042005200620072008200920102011201220132014201520162017


$0

$50

$100

$150

$200

$250

$300

$350

$400

$450

$500

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17


Simple Average Transaction Price per Square Foot

55

Exhibit 46: Value-Weighted Average Annual Office Price per Square Foot, New York

In this value-weighted series, even the annual trend becomes unstable, while the spikes in the quarterly series highlight the “blockbuster” trade problem. Repeat-sale models offer a solution to the changing composition problem (of which the “blockbuster” trade problem is a special case). Repeat-sale models only include properties with multiple trades, thereby isolating price movements to market forces rather than changes in the sample. However, restricting the set of deals to repeat trade pairs means that repeat-sale models only use 10-20% of the transaction data, and require a minimum number of trades in each period to produce statistically meaningful price estimates. As a result, repeat-sale models only work at a national level and for large, heavily traded markets. Repeat-sale models also require trades before and after each period to produce statistically valid results. The model will restate, often significantly, the current period (and the recent past) as researchers collect additional datapoints. Repeat-trades may also exhibit a bias, as we observe that properties that retrade almost always do so at a profit. Owners may only sell the properties on which they can realize the highest returns, and may choose to hold properties rather than incur a loss by selling. This bias affects the entire sample of transactions, but may be more pronounced for repeat-trades. These concerns notwithstanding, repeat-trade models offer a valid alternative to simple or weighted-average prices. In 2012, CoStar released the CoStar Commercial Repeat-Sales Indices (CCRSI), the first-large scale repeat-sales series for the sector. The firm continues to report results each month, found at http://www.costargroup.com/ costar-news/ccrsi. Exhibit 47 presents the national equal- and value-weighted composite repeat-sales indices through January 2018.

200020012002200320042005200620072008200920102011201220132014201520162017


$0

$100

$200

$300

$400

$500

$600

$700

$800

$900

$1,000

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17


Value-Weighted Average Transaction Price per Square Foot

http://www.costargroup.com/%20costar-news/ccrsi

http://www.costargroup.com/%20costar-news/ccrsi

56

Exhibit 47: CoStar Office Repeat-Sales Indices

We believe, however, that the CoStar transaction dataset proffers a more intriguing opportunity. Given the amount of data and computing power at our disposal, we believe we can produce a reasonable—and empirically testable—estimate of price and cap rate for every property, over time. In doing so, we can fully control for the composition problem by aggregating all the properties in a particular geography, rather than only those that traded (or repeat-traded). Moreover, we believe we can capture high-frequency movements by including peer property trades in our estimate: If the property across the street trades at a higher price, we can safely assume that the price for our property also rose. Property-level price and cap rates series also fill another need. CoStar’s new forecasting models produce forecasts at a property-level. In order to produce a forecast, we require a history, and therefore require full time series for all variables for every property.

5.3 ESTIMATING PROPERTY VALUES What is the value of a particular real estate asset? Real estate assets trade infrequently, and the unique characteristics and idiosyncrasies of any particular property make applying market average or comp trades problematic. The appraisal industry exists to answer this question, and to do so relies on a synthesis of three different approaches: first, a comps-based approach, where appraisers review and adjust nearby and recent comparable trades to arrive at a likely value; second, a discounted cash flow (or DCF) analysis approach, where appraisers attempt to forecast future income and eventual sale proceeds discounted to today; and third, a replacement cost analysis, where appraisers attempt to estimate the cost of building the asset in question.

http://www.costargroup.com/costar-news/ccrsi

Period

3/31/1996

6/30/1996

9/30/1996

12/31/1996

3/31/1997

6/30/1997

9/30/1997

12/31/1997

3/31/1998

6/30/1998

9/30/1998

12/31/1998

3/31/1999

6/30/1999

9/30/1999

12/31/1999

3/31/2000

6/30/2000

9/30/2000

12/31/2000

3/31/2001

6/30/2001

9/30/2001

12/31/2001

3/31/2002

6/30/2002

9/30/2002

12/31/2002

3/31/2003

Source: CoStar. http://www.costargroup.com/costar-news/ccrsi As of January 2018

0

50

100

150

200

250

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Equal-Weighted Value-Weighted

Repeat-Sales Index (100 = Minimum Value after 2007)

57

The DCF approach requires assumptions involving future cash flows, future resale value, and discount rate, and different firms may arrive at different answers depending on their outlook for rents, their exit cap rate assumptions, and their cost of capital. CoStar does forecast rent growth as well as value (the next section describes CoStar’s forecasting methodology), and could, in theory, use this approach to estimate current values after making an assumption about discount rates. However, our aim is not only a current value estimate but a full time series, which greatly complicates the deployment of the DCF approach, as we would require rent and value forecasts made at points in the past, as well as a system to modulate discount rates over time. Given these complexities, we have chosen not to employ a DCF approach—at least not yet. Replacement cost analyses also poses challenges, due largely to a lack of quality data around construction costs and land costs. Furthermore, estimates of replacement costs for older, outdated assets become problematic, as a replacement of the asset would likely look a lot different. For example, what is the replacement cost of a warehouse on Atlantic Avenue in downtown Brooklyn, or a gas station in the South of Market district in San Francisco? A comps-based approach to value properties, on the other hand, maximizes the value of CoStar’s unmatched comps dataset. Moreover, a comps-based approach arguably produces the most objective outcome, relying on market activity rather than forecasts or assumptions about costs of construction or capital. We begin with certainty: The value of a property on the day the property transacts is equal to the transaction price. Moving forward from the transaction date, the value of the property becomes less and less certain, and will vary based on the prices at which comparable properties trade, the local rent trends, and broad capital market conditions. We have different approaches for valuing properties that have traded from those that did not. We also handle large properties differently from smaller properties; testing results show that the comps-based approach performs much worse for smaller properties, and that less sophisticated and more efficient results perform equally well. We define small properties as the square footage cut off at which no more than 35% of total RBA for the property type was classified as “small”; Exhibit 48 shows the specific cut-offs for each property type.

Exhibit 48: Cut-Offs For Large and Small Properties

Property Type Description

Office 37,000 SF

Industrial 50,000 SF

Multifamily 37,000 SF

Retail 12,000 SF

58

The sections below lay out how we identify comp properties; how we use comps to value large properties for which we have a transaction price; how we value large properties that did not trade; and how we handle smaller properties.

5.3.1 Identifying Comp Properties

To implement a pricing model based on comps, we must first identify comp properties. To do so, we first identify all properties within 10 miles of the subject property for which we have a market-rate trade3. We then calculated a set of weights, each between 0 and 1, based on the distance to the subject property and the density of properties around it, the difference in sizes, the difference in the lot sizes, and the different in the ages of the comp property relative to the subject property. We use the product of these weights as a judge of the relevance of a particular property in valuing the subject property. On average, each traded office property has about 50 comp properties. As expected, properties in large markets have many more comps, while more remote properties may have only a few comps, or even none at all. Exhibit 49 presents the comps for 33 Arch Street in Boston, the home of CoStar’s analytic team, shaded by the quality of each comp based on distance from the subject property and similarity in size and age.

3 We exclude all non-arm’s length trades, as well as condo, partial interest, portfolio, ground lease, and other non-standard trades for the purposes of limiting the comp set to “market-rate” trades.

59

Exhibit 49: Sale Comps for 33 Arch Street, Boston

Note that in Boston, New York, Washington, and San Francisco we do not allow properties to use sale comps in other counties. This prohibition ensures that we do not use, for example, Oakland trades as comps for San Francisco properties, or Queens for Manhattan, or Cambridge for Boston (and vice versa).

5.3.2 Estimating Values in Large Traded Properties

As noted previously, we begin with the assumption that the value of a property on the day that it trades is equal to the sale price. Going forward, three factors will cause our price estimates to change: actual sales of peer properties (or the subject property); estimated trends in NOI; and broad national cap rate trends. Sales of peer properties will affect the price estimate. At each point in time after the subject property trades, the price estimate will equal the weighted average of the subject property’s initial trade and the sale comps, where the weight of each comp (including the subject property) diminishes over time using an exponential decay function. Exhibit 50 shows the estimate of value, based solely on comp trades, for 33 Arch Street. The size of each dot indicates the weight of that comp, with higher weights indicating a comp more similar in terms of proximity, size, lot size, and age.

60

Exhibit 50: Sale Comps for 33 Arch Street, Boston

Readers will note the large red dot, indicating the subject property trade. Assigning a weight to this trade poses a problem. By default, the weight on the subject property’s trade would equal 1; however, in practice, a weight of 1 may not properly reflect the value of the information of the initial trade. Too low of a weight will decline too quickly as a result of time, or may be overwhelmed by comp trades in heavily traded markets, such that the subject property trade ceases to matter almost immediately. In general, we would prefer to assign the subject property a higher weight to ensure that the future estimated values do not deviate too far from the initial value. To achieve some permanence of the subject property’s first trade, we set the weight of the subject property trade equal to the sum of the all the time-adjusted weights on comp trades that have preceded the trade, multiplied by five. This construct has the effect of giving the subject property a weight more or less proportional to the amount of trading activity in the market, and ensures that comp trades do not overwhelm the initial value. Moreover, a large weight on the subject property helps maintain the premium or discount a particular property may have to its comps: For example, we would expect a property that trades at 20% above a large set of comps to maintain a 20% premium to the market going forward. Thus far, we can denote the methodology for valuing properties based on comps as:

𝑉𝑝𝑡=

∑ 𝑉𝑐 ∗ 𝑊𝑃,𝑐 ∗ 𝑒−(𝑡−𝑡𝑐) 𝑐

∑ 𝑊𝑃,𝑐 ∗ 𝑒−(𝑡−𝑡𝑐) 𝑐

Where 𝑉𝑝𝑡

denotes the value of a given property 𝑝 at time 𝑡, 𝑉𝑐 is the value of comp 𝑐,

𝑊𝑃,𝑐 is the time-invariant weight of comp 𝑐 on property 𝑝, and 𝑒−(𝑡−𝑡𝑐) is the time weight

PropertyID

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

Source: CoStar Comps As of December 2017

$0

$100

$200

$300

$400

$500

$600

$700

$800

$900

$1,000

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Price per SF

61

of comp 𝑐 based on the difference between time 𝑡 and the date of comp 𝑐’s trade at time 𝑡𝑐. Or, in plain English, the value of a property at a particular point in time equals the weighted average of all the comps since and including the subject property trade, where the weight of each trade depends on proximity, similarity in size and lot size, age, and time. Exhibit 51 adds a line as the estimate of pricing over time for the same office asset based solely on its comps.

Exhibit 51: Price Estimates for 33 Arch Street Using Comps

5.3.2.1 Adjustments to Comp Prices

We have identified peer properties based on their proximity and physical similarity to the subject property, and weighted each peer based on its similarity to the subject property. In doing so, we hope to down weight peer comps where the price per square foot may not be reflective of the subject property. Ideally, however, we would like to adjust peer prices to bring them more in line with the likely price of the subject property. We can arrive at these adjustments via repeat trades of peer properties. For example, let us assume that the subject property trades for $100 per square foot at the same time that a peer property trades for $200 per square foot. Thus, we know that the subject property traded for half the price of the peer property. Five years later, the peer property trades again, this time for $400 per square foot. For the purposes of estimating the subject property price, we can assume a value of $200 per square foot based on the ratio established by the initial trade. Repeat trades, however, happen infrequently, and so have limited usefulness in making price adjustments. But, we can take advantage of the transitive property from mathematics to greatly expand the set of properties for which we can make these price

PropertyID

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932


$0

$200

$400

$600

$800

$1,000

$1,200

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Price per SF

62

adjustments. The transitive property declares that if a is equivalent to b, and b is equivalent to c, then a must also be equivalent to c. We refer to our application of the transitive property as the “Kevin Bacon model,” in honor of the cinephile pastime of linking Kevin Bacon to other actors via the films in which they have performed together. The Kevin Bacon model works as follows. We wish to estimate the price of property A, which trades for $100 at t0. At the same time, property B trades for $200. As in the example above, property B trades again five years later at t5, for $400. Also at t5, property C trades for $300, giving us the ratio between B and C, and by extension, from A to C. Property C had also traded two years prior, at t3, for $200, allowing us to estimate the value of property A then. And why not include property D, which also trades at t3 as well as in t1, for $500 and $400 respectively, allowing us to estimate the value of A then as well. Exhibit 52 presents a stylized example of how this model works.

Exhibit 52: The Kevin Bacon Model

In practice, we first triplicate all the repeat trades, such that we assume each trade could have occurred in the prior or next year. The triplication greatly increases the number of matches, and thus the set of adjustment factors. We prefer to err of the side of more matches, believing that a rough price adjustment will be better than no adjustment. We standardize all prices by estimating a year-end value in each of the three years using rent growth as a proxy of NOI and the national cap rate trend (more on the effects of rent trends and the national cap rate series below). We then join this initial dataset onto itself repeatedly until we exhaust the set of possible combinations. In total, we can establish an adjustment factor for more than a third of all comps. Using these adjustment factors significantly reduces the error around our estimates of the price per square foot in large traded properties. Exhibit 53 presents the estimates for 33 Arch Street, incorporating the price adjustments. The darker dots indicate the adjusted prices for each comp.

0

1

2

3

4

5

Source: CoStar

A AA

AB

BD

D

C

C

$0

$100

$200

$300

$400

$500

$600

0 1 2 3 4 5

Price per SF

Period

A and B trade at time 0.A = B * (100/200) = 0.5

B and C trade at time 5.B = C * (400/300) thereforeA = C * (100/200) * (400/300) = 0.67

C and D trade at time 3.C = D * (200/500) thereforeA = D * (100/200) * (400/500) * (200/500) = 0.27

At time 1, we can assume A = D * 0.27

At time 3, we can assume A = C * 0.67

At time 5, we can assume A = B* 0.5

63

Exhibit 53: Price Estimates for 33 Arch Street Using Adjusted Comps

We can represent this mathematically as:

𝑉𝑝𝑡=

∑ 𝑉𝑐 ∗ 𝑊𝑃,𝑐 ∗ 𝑒−(𝑡−𝑡𝑐) ∗ 𝐾𝐵𝑐 𝑐


Where 𝐾𝐵𝑐 denotes the adjustment factor for the comp in question. While the comps-based estimation includes the effect of new comps, it also includes all prior comps, including the subject property’s initial trade, at the original prices. A comp from five or ten years ago continues to inform our current price estimate, albeit at a lower weight due to the passage of time. Moreover, in the absence of any comps, the value of a property today would equal the value of its last trade, no matter how long ago the trade occurred. As part of the calculation, we would like to trend old prices forward. But how? Consider this thought experiment: Imagine a property in a small market where no other properties trade. The property trades for $10 million in January of 2000, and we wish to estimate the value of the property at the end of each quarter since then. We reject the naïve assumption that the price has remained flat over time. Instead, we will draw on the classic real estate equation that cap rates equal income divided by value, which we can rearrange to solve for change in value:

Δ𝑉𝑎𝑙𝑢𝑒 =𝛥𝑅𝑒𝑛𝑡𝑂𝑐𝑐

Δ𝑁𝑎𝑡𝐶𝑎𝑝

PropertyID

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932


$0

$200

$400

$600

$800

$1,000

$1,200

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Price per SF

64

Where Δ𝑉𝑎𝑙𝑢𝑒 denotes the percentage change in a property’s value, Δ𝑅𝑒𝑛𝑡𝑂𝑐𝑐 denotes the percentage change in rent multiplied by occupancy, a proxy for the effect of NOI on value, and Δ𝑁𝑎𝑡𝐶𝑎𝑝 denotes the percentage change in national yield trends. With this equation in mind, we will allow our estimate of the property’s value to change with trends in submarket rent multiplied by occupancy, as a proxy for NOI, and with broad national cap rate trends.

5.3.2.2 The Effect of Rent Movements on Price Estimates

All else equal, we will assume that a property’s price will change with trends in rents (using the same-store rent series) and occupancy at a submarket level. At a national level, we observe some correlation between this measure and the national repeat-sale series, as shown in Exhibit 54.

Exhibit 54: Correlation Between Rents and Values

We choose to use submarket rent trends rather than the particular property’s rent, which can in some cases show erratic patterns or excessive smoothness, either of which could produce value trends out of step with the overall market. Submarket rent trends create differentiation within a market while producing broadly consistent series for nearby properties. We can denote the inclusion of the rent and occupancy data as:

𝑉𝑝𝑡=

∑ 𝑉𝑐 ∗ 𝑊𝑃,𝑐 ∗ 𝑒−(𝑡−𝑡𝑐) ∗ 𝐾𝐵𝑐 ∗ Δ(𝑅𝑒𝑛𝑡𝑂𝑐𝑐)𝑐𝑡 𝑐


3/31/1996

6/30/1996

9/30/1996

########

3/31/1997

6/30/1997

9/30/1997

########

3/31/1998

6/30/1998

9/30/1998

########

3/31/1999

6/30/1999

9/30/1999

########

3/31/2000

6/30/2000

9/30/2000

########

3/31/2001

6/30/2001

9/30/2001

########

3/31/2002

6/30/2002

9/30/2002

########

3/31/2003

6/30/2003

Source: CoStar As of December 2017

$10

$15

$20

$25

$30

$35

0

20

40

60

80

100

120

140

160

180

200

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16

Equal-Weighted Office Repeat Sale Index Same-Store Rent

Equal-Weighted Office Repeat Sale Index National Office Same-Store Rent

65

Where Δ(𝑅𝑒𝑛𝑡𝑂𝑐𝑐)𝑐𝑡 is the percent change in rent times occupancy in comp 𝑐 between time 𝑡𝑐 (the date of the comp trade) to time 𝑡 (the valuation date). Testing results show that including submarket rent and occupancy trends decrease the error of property-level price estimates, as presented in Exhibit 55.

Exhibit 55: Price Estimates for 33 Arch Street Using Comps and Rent Trends

5.3.2.3 The Effect of National Cap Rate Trends on Price Estimates

Broad cap rates trends serve as a proxy for the capital markets environment and overall pricing trends. Whereas including rent and occupancy captures local trends in performance, the national cap rate trend ensures that, all else equal, property prices around the country move together, all subject to powerful macro forces that allocate capital. To include a national cap rate series in our price estimates, we must first create one. CoStar’s dataset contains more than 350,000 cap rates. Not every transaction reports a cap rate, and cap rate definitions in the United States are not necessarily consistent—in some cases the seller reports the cap rate, in others the buyer. The reported cap rate may be based on stabilized, in-place NOI, or it may be based on pro forma cash flow expectations. Despite these vagaries, we believe that CoStar’s cap rate data provides the best information to determine broad trends in cap rates. Exhibit 56 graphically depicts CoStar’s cap rate data for 3 Star and higher office properties selling for at least $5 million. The size of each dot indicates the size of the trade, and the shading indicates the trend: Cap rates nearer to the average show as black, and the lightest shading indicates outliers.

PropertyID

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932


$0

$200

$400

$600

$800

$1,000

$1,200

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Price per SF

66

Exhibit 56: National Office Cap Rates

From this discrete transaction cap rate data we create value-weighted average national series to serve as the broad trend lines. These trends represent a center-smoothed value-weighted monthly average of the transaction cap rates. The center smoothing will result in some minor changes to history in the near term; we try to minimize the extent of the restatement by assuming cap rates will remain flat into the future. Exhibit 57 presents the national cap rate trends for the four major property types.

Exhibit 57: National Cap Rate Trend Lines by Property Type

We will assume that prices follow this national trend in cap rates, after adjusting the level of the national series to match the average cap rate level for each cluster. For example, we will shift the national series lower in low cap rate markets, magnifying the percentage change in cap rates. On the other hand, high cap rate markets will have

Source: CoStar Portfolio Strategy. Cap rates are shaded by distance from trailing average to identify trend. As of December 2017Includes 3+ star-rated office assets selling for at least $5 million.

2%

3%

4%

5%

6%

7%

8%

9%

10%

11%

12%

99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Office Transaction Cap Rate

2

2

2

2

2

2

2

2

2

2

2

2

2

2

2

2


4%

5%

6%

7%

8%

9%

10%

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Industrial Office Retail Apartment

National Cap Rate Trends

67

less volatility due to cap rate movements. Mathematically, we can represent the inclusion of the cap rate trends as

𝑉𝑝𝑡=

∑ 𝑉𝑐 ∗ 𝑊𝑃,𝑐 ∗ 𝑒−(𝑡−𝑡𝑐) ∗ 𝐾𝐵𝑐 ∗ Δ(𝑅𝑒𝑛𝑡𝑂𝑐𝑐)𝑐𝑡 ∗1

Δ(𝑁𝑎𝑡𝐶𝑎𝑝)𝑐𝑡

𝑐


Where Δ(𝑁𝑎𝑡𝐶𝑎𝑝)𝑐𝑡 is the percent change in the national cap rate series between time 𝑡𝑐 (the date of the comp trade) to time 𝑡 (the valuation date), where the national cap rate is shifted up or down by the difference between the typical cap rate in the cluster and the national average. This adjustment has the effect of increasing the volatility of price movements in low cap rate markets, and vice versa. Testing results show that including submarket rent and occupancy trends decrease the error of property-level price estimates, as presented in Exhibit 58. Exhibit 58: Price Estimates for 33 Arch Street Using Comps, Rents, and National Cap Rate Trends

The inclusion of the cap rate trends results in the expected decrease in values beginning in 2008, an effect not otherwise realized given the scarcity of comp rates and the slower reaction from rent trends.

5.3.3 Estimating Values in Large Untraded Properties

About one in five large properties has a recorded market-rate sale price from which we can extrapolate full time series using the methodology described in the preceding section. For the remaining 80%, we can draw on the full time series for values created for every traded property by the previous step, provided we make some adjustments to account for differences between the subject property and traded peer properties.

PropertyID

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932

9932


$0

$200

$400

$600

$800

$1,000

$1,200

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18

Price per SF

68

The set of peer properties once again depends on proximity, and similarities in age, size, and rent. We can also use the differences in size, age, and rent to estimate an adjusted value for the peer properties, in much the same way appraisers make adjustments to comp properties. Our adjustments are based on results of regressions by property type that seek to explain the differences in price based on differences in rent, building size, lot size, and age. We specify this equation as:

log(𝑝𝑟𝑖𝑐𝑒𝑑𝑖𝑓𝑓) = 𝑓𝑛( log(𝑠𝑖𝑧𝑒𝑑𝑖𝑓𝑓) , log(𝑙𝑜𝑡𝑑𝑖𝑓𝑓) , log(𝑎𝑔𝑒𝑑𝑖𝑓𝑓) , log(𝑟𝑒𝑛𝑡𝑑𝑖𝑓𝑓) ) Where log (𝑝𝑟𝑖𝑐𝑒𝑑𝑖𝑓𝑓) denotes the natural log of the difference in price per square-foot (units for multifamily assets) between two peer properties which traded within 12 months of each other in arms-length, market-rate trades. As independent variables, log (𝑠𝑖𝑧𝑒𝑑𝑖𝑓𝑓) denotes the natural log of the difference in the physical size in square

footage between the two assets; log (𝑙𝑜𝑡𝑑𝑖𝑓𝑓) denotes the natural log of the difference in the lot size in square footage between the two the assets; log (𝑎𝑔𝑒𝑑𝑖𝑓𝑓) denotes the natural log of the difference in age between the two assets; and log (𝑟𝑒𝑛𝑡𝑑𝑖𝑓𝑓) denotes the natural log of the difference in rent between the two assets. The specific equations and coefficients are presented below for reference in Exhibit 59:

Exhibit 59: Peer Price Adjustment Regressions

Property Type Constant rentdiff agediff rbadiff lotdiff storiesdiff unitsdiff

Industrial -0.1300 0.3230 -0.4300 -0.1870 0.1460 NA NA

Office -0.0970 1.0020 -0.2180 NA NA NA NA

Retail -0.2150 0.6190 -0.4430 -0.3240 0.1240 NA NA

Multifamily -0.0021 0.8540 -0.3120 NA 0.0944 0.1450 -0.1370

We use these coefficients to systematically adjust the values of peer properties to make the peer valuations more applicable to a particular untraded asset. For example, the Inter-American Development Bank Building at 1300 New York Avenue NW in Washington, D.C. has a large number of nearby peer properties, all built at different times, and of varying sizes and rents. Exhibit 60 presents information for the subject asset and a set of nearby peer properties, as well as two measures of value for each peer asset: Our unadjusted estimated value in the current period (2019Q1 at the time of writing), and an adjusted price which we calculate by applying the coefficients above to the differences in size, lot size, age, and rent between the subject property and each peer property. For example, peer property 129671 (1299 Pennsylvania Avenue) is older (built in 1924) but has higher rents ($61 compared with $58). Applying the coefficients from above to these differences results in a $18 adjustment to the peer price for the purposes of estimating a value for the subject property.

69

Exhibit 60: Adjusted and Unadjusted Peer Prices for 1300 New York Avenue, Washington

Peer ID Distance Rent Diff.

Age Diff.

Rent Adj.

Age Adj.

Constant Adj.

Raw Peer Price

Adj. Peer Price

129671 0.21 ($3.37) -60 ($34) $125 ($57) $618 $636 129211 0.13 $1.04 5 $10 ($13) ($49) $532 $479 121296 0.52 $7.40 0 $82 ($0) ($51) $554 $576 129241 0.34 ($1.36) 2 ($13) ($6) ($52) $565 $499 129210 0.12 ($2.23) 5 ($21) ($14) ($51) $553 $469 129118 0.76 $4.20 -6 $59 $21 ($69) $750 $771 129855 0.56 $0.45 7 $4 ($19) ($48) $520 $459 129231 0.51 ($0.99) 6 ($8) ($15) ($46) $497 $429 129110 0.87 ($1.41) -16 ($10) $29 ($39) $417 $395 129720 0.37 $0.33 17 $3 ($53) ($49) $535 $440 129029 1.02 $7.68 3 $54 ($5) ($33) $352 $362 129748 0.35 ($0.35) 17 ($5) ($89) ($84) $908 $738 129426 0.74 $8.17 -11 $70 $21 ($39) $422 $468 130483 0.33 ($1.79) 16 ($15) ($46) ($46) $499 $398 130648 0.10 ($2.11) 19 ($23) ($74) ($60) $653 $507 130011 0.27 ($2.96) 17 ($37) ($75) ($70) $759 $591 129266 0.16 $10.30 -56 $101 $89 ($43) $463 $607 129089 0.68 ($6.58) -12 ($85) $44 ($77) $829 $710 130089 1.36 $8.68 7 $114 ($23) ($59) $642 $657 753262 1.06 $7.61 21 $66 ($55) ($40) $433 $392 129861 0.24 ($9.99) 11 ($96) ($38) ($60) $647 $469 129386 1.17 $12.05 -10 $113 $19 ($39) $426 $508 817883 0.22 ($13.38) 22 ($191) ($138) ($94) $1,016 $645 478492 1.12 $14.36 -15 $159 $31 ($44) $479 $614

6389751 1.49 $4.26 26 $47 ($102) ($55) $594 $479 129704 1.29 $9.92 -55 $111 $101 ($49) $532 $690

To arrive at our estimate for the subject property value, we will take an average of the adjusted peer values weighted by how similar the subject and peer asset are, as measured by the size of the adjustment made to the peer asset. In this case, the subject asset is generally newer, and charges higher rents, than its peers, and as a result we increase the values of the peer properties upward. In addition to the peer property prices, we also include in the calculation a geographic average with a low weight. In heavily traded markets, this geographic average has no real impact on the estimate price. But, in lightly traded markets, the price estimates depend heavily on the geographic average, pro rata to the number of peer properties for which we can estimate a price based on an actual trade.

70

5.3.4 Estimating Values for Small Properties

Applying the comp-based approach described above to smaller properties results in relatively poor estimates. Error measures are much higher for small properties, for which the idiosyncrasies of a particular trade represent a larger share of the trade price. For example, trades can include business value, land, furniture or furnishings, or any other non-real estate value. Moreover, deals for smaller properties often involve non-institutional investors who may have different motivations for acquiring properties beyond earning a return. Complicating matters further, small properties represent more than 80% of all trades, and a corresponding demand on data storage and computation time. Thus, the comps-based valuation algorithm would spend the majority of its computational resources making relatively poor estimates for small properties, the peers of which comprise a significant portion of the very large peer property dataset. Given these realities, we have explored other, more efficient means of estimating the values of smaller properties. We have found that a gross rent multiplier approach yields somewhat better estimates, while requiring a fraction of the storage and computation time. We estimate the gross rent multiplier for each quality cut as the average sale price divided by the rent for more than 600,000 trades for smaller properties. We then trend the gross rent multiplier using the reference cap rate series and an income trend based on lagged rents and vacancies for the local market. The gross rent multiplier approach assumes that values relate directly to rents, according to the quality of the asset. Since we have a rent, albeit estimated in many cases, for every commercial property, we can apply a multiplier to the rent and very quickly arrive at an estimate of value. This value estimate will change with the property’s rents, and with modulations in the gross rent multiplier.

5.3.5 Accuracy of Value Estimates

The quality of CoStar’s building-level price model ultimately depends on how accurately we can predict a property’s trade value. Minimizing the error of our estimates will results in the best depiction of aggregate prices and trends. We use different testing regimes for the different types of estimates. For repeat trades, we calculate the error in our estimate as the difference between the repeat trade price and our estimated price in the preceding quarter (which will incorporate information from the first trade). For first trades, we compare the trade price to our estimate based on peer properties and the geographic averages. Exhibits 61 through 64 below present the median absolute error by property type for all trades, a property’s first trade, and a property’s repeat trades. The overall error is lowest for apartment properties and highest for retail properties. In general, our error is

71

lower for higher-value trades than for lower-value trades. We believe the higher error results from the inherent idiosyncrasies of small trades, which may include business value, land value, other non-real estate assets, or unusual relationships between buyers and sellers. Our error on repeat trades is also significantly lower than our error on first trades, reflecting the large amount of information contained in a property’s prior trades, an effect especially apparent for retail properties.

Exhibit 61: Office Price Estimate Median Absolute Error

Office Median Absolute Error Total Observations Total First Trade Repeat Trades

Total 136,342 29.5% 31.1% 20.9%

Value: $100,000 - $1,000,000 83,693 31.7% 32.9% 22.3%

Value: $1,000,000 - $10,000,000 43,463 28.6% 30.1% 22.2%

Value: $10,000,000 - $50,000,000 6,945 19.4% 21.3% 13.4%

Value: $50,000,000 - $100,000,000 1,221 16.0% 19.1% 11.0%

Value: $100,000,000 - $250,000,000 763 15.3% 18.4% 11.5%

Value: $250,000,000 - $500,000,000 199 16.6% 17.6% 10.7%

Value: $500,000,000+ 58 20.6% 23.0% 12.5%

Exhibit 62: Apartment Price Estimate Median Absolute Error

Apartment Median Absolute Error Total Observations Total First Trade Repeat Trades

Total 175,911 23.3% 25.6% 18.6%

Value: $100,000 - $1,000,000 125,272 28.5% 30.2% 22.9%

Value: $1,000,000 - $10,000,000 36,436 21.7% 24.4% 17.8%

Value: $10,000,000 - $50,000,000 10,446 13.4% 13.9% 12.5%

Value: $50,000,000 - $100,000,000 2,829 11.7% 13.1% 8.7%

Value: $100,000,000 - $250,000,000 705 14.4% 16.5% 9.6%

Value: $250,000,000 - $500,000,000 169 22.9% 17.8% 40.0%

Value: $500,000,000+ 36 19.2% 19.2% NA

72

Exhibit 63: Industrial Price Estimate Median Absolute Error

Industrial Median Absolute Error Total Observations Total First Trade Repeat Trades

Total 166,421 27.8% 28.9% 21.9%

Value: $100,000 - $1,000,000 94,453 31.3% 32.3% 24.0%

Value: $1,000,000 - $10,000,000 67,518 24.0% 24.9% 20.7%

Value: $10,000,000 - $50,000,000 4,208 23.2% 25.0% 17.6%

Value: $50,000,000 - $100,000,000 213 26.6% 29.5% 16.6%

Value: $100,000,000 - $250,000,000 27 50.8% 50.8% 56.8%

Value: $250,000,000 - $500,000,000 1 66.3% 66.3% NA

Value: $500,000,000+ 1 69.4% 69.4% NA

Exhibit 64: Retail Price Estimate Median Absolute Error

Retail Median Absolute Error Total Observations Total First Trade Repeat Trades

Total 254,846 39.2% 41.6% 23.8%

Value: $100,000 - $1,000,000 157,869 44.6% 47.9% 24.4%

Value: $1,000,000 - $10,000,000 93,311 33.4% 35.1% 23.2%

Value: $10,000,000 - $50,000,000 3,514 34.8% 36.2% 21.7%

Value: $50,000,000 - $100,000,000 115 48.8% 48.1% 60.7%

Value: $100,000,000 - $250,000,000 32 45.6% 46.6% 28.2%

Value: $250,000,000 - $500,000,000 4 58.6% 58.6% NA

Value: $500,000,000+ 1 94.6% 94.6% NA

Overall, these error metrics highlight the difficulties in estimating the values of highly idiosyncratic assets such as commercial real estate properties. Property values are highly dependent on characteristics that are difficult to observe and buildings that appear similar in the data can trade for very different prices. For example, the CoStar dataset includes two office properties in Portland that traded three days apart in January 2017. Both are 4-star office buildings in the CBD submarket, each over 250,000 SF. However, one property traded for $325 per SF, while the other traded for $221, a difference of over 50%. We will continue to test and refine our methodology to incrementally lower the error in our estimates. Key areas of investigation include the selection and weighting of peer properties; the estimation of prices for smaller properties; the use of rent and cap rate trends; and the optimal weighting of a property’s prior trades.

73

5.4 ESTIMATING PROPERTY CAP RATES Cap rates represent another important pricing metric for commercial real estate. In keeping with our philosophy of providing property-level data across all key commercial real estate metrics, we estimate current and historical cap rates for all properties in the CoStar data set as well. As with prices, we believe that accurate property-level cap rate estimates will also result in the most accurate portrayal of aggregate trends. Additionally, property-level cap rate estimates allow our clients to define their own custom sets of properties and view cap rate series most applicable to their investments. This section outlines our approach to estimating property-level cap rates, and presents our aggregate national series.

5.4.1 Spot Cap Rate Estimates

We start by establishing a spot estimate of a property’s cap rate at a point in time. For the 130,000 or so traded properties that report a cap rate, we use this transaction cap rate as of the date of the trade. For all properties for which we have not collected a cap rate, we estimate a cap rate as of 2015Q1 using a regression of actual observed yields between 2011 and 2017 on the natural logs estimated price per square foot, total deal size, and cap rates on peer trades (over the same 2011 to 2017 time period, and weighted by similarity to the subject property, as well as proximity). The specific equations and coefficients we use to estimate cap rate are presented below for reference: Office Cap Rate Model: log(𝐶𝑎𝑝𝑅𝑎𝑡𝑒) = 1.492 −0.124 log(𝑝𝑝𝑠𝑓) − 0.0889 𝑙𝑜𝑔(𝑝𝑟𝑖𝑐𝑒) + 0.636 log (𝑝𝑒𝑒𝑟𝑐𝑎𝑝𝑟𝑎𝑡𝑒) Industrial Cap Rate Model: log(𝐶𝑎𝑝𝑅𝑎𝑡𝑒) = 1.016 −0.0723 log(𝑝𝑝𝑠𝑓) − 0.0268 𝑙𝑜𝑔(𝑝𝑟𝑖𝑐𝑒) + 0.845 log (𝑝𝑒𝑒𝑟𝑐𝑎𝑝𝑟𝑎𝑡𝑒) Retail Cap Rate Model: log(𝐶𝑎𝑝𝑅𝑎𝑡𝑒) = 1.601 −0.137 log(𝑝𝑝𝑠𝑓) − 0.012 𝑙𝑜𝑔(𝑝𝑟𝑖𝑐𝑒) + 0.624 log (𝑝𝑒𝑒𝑟𝑐𝑎𝑝𝑟𝑎𝑡𝑒) Apartment Cap Rate Model: log(𝐶𝑎𝑝𝑅𝑎𝑡𝑒) = 1.806 −0.142 log(𝑝𝑝𝑠𝑓) − 0.0266 𝑙𝑜𝑔(𝑝𝑟𝑖𝑐𝑒) + 0.598 log (𝑝𝑒𝑒𝑟𝑐𝑎𝑝𝑟𝑎𝑡𝑒)

5.4.2 Cap Rate Trends

To create full cap rate time series back to 2000 for every property, we modulate the national cap rate series based on the cap rate level for the particular property and the relative price performance of that property to the national trend. Properties that have outperformed the national price series will see cap rates compress by more than the national cap rate series, and vice versa, based on the equation presented below:

74

𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡 = 𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡−1 ∗ 𝑁𝑎𝑡𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡

𝑁𝑎𝑡𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡−1∗ √

(𝑁𝑎𝑡𝑃𝑟𝑖𝑐𝑒𝑡

𝑁𝑎𝑡𝑃𝑟𝑖𝑐𝑒𝑡−1)

(𝑃𝑟𝑜𝑝𝑃𝑟𝑖𝑐𝑒𝑡

𝑃𝑟𝑜𝑝𝑃𝑟𝑖𝑐𝑒𝑡−1)

Where 𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡 denotes the estimate of the property cap rate at time t, 𝑁𝑎𝑡𝐶𝑎𝑝𝑅𝑎𝑡𝑒 denotes the national cap rate, 𝑃𝑟𝑜𝑝𝑃𝑟𝑖𝑐𝑒 denotes the estimate price per square foot for

the property, and 𝑁𝑎𝑡𝑃𝑟𝑖𝑐𝑒 denotes the national price per square foot.

75

6 FORECASTING

The commercial real estate sector boasts a long and respected tradition of forecasting, dating back to the seminal work of Drs. Ray Torto and William Wheaton, who teamed up to form Torto Wheaton Research in 1982, and of Susan Hudson-Wilson, who founded Property & Portfolio Research in 1994. Scraping together the little data that existed at the time, these two Boston-based firms developed the basic formulation for forecasting commercial real estate still used by practitioners today. The leadership of CoStar’s analytics group consists of veterans of Property & Portfolio Research, which became part of the CoStar Group in 2009. Together, CoStar’s Boston-based analytics group has more than 40 years of commercial real estate forecasting experience, across multiple periods, continents, and datasets. The commercial real estate sector’s particular interest in forecasts likely relates to the widespread use of DCF models to underwrite and evaluate deals. These models require a forecast of rents and occupancies, cash flows, and exit cap rates. Forecasts also help investors allocate capital by providing metrics they can use to rank markets. More recently, lenders have employed statistical models to evaluate loans that also rely on forecasts of the performance of the underlying collateral. And Wall Street investment firms believe they can gain an edge in trading commercial real estate assets like REITs or CMBS pools by tying the underlying assets to a forecast and generating an aggregate expected return. Traditionally, commercial real estate forecasters have only had market-level data at their disposal, either from brokerage firms or industry organizations like NCREIF. Consequently, they have produced market-level forecasts of the key variables of supply, demand, vacancy, rent, NOI, cap rates, and price change. With the increasing availability of submarket-level fundamentals and rent data, forecasters have developed share-down models to produce submarket forecasts. But no one has seriously considered bringing forecasts to individual properties at scale…until now. The processes and methods laid out in the preceding sections to create same-store rents and price and cap rate estimates for every property not only produce the most accurate view of historical trends, but also provide the necessary raw inputs for creating building-level forecasts for every commercial and multifamily property CoStar tracks—more than 3.3 million properties. Property-level forecasts confer many benefits. They free our clients from the tyranny of geography: No longer limited to market or submarket forecasts, clients may view a forecast for any set of properties—their own portfolio, a REIT, or a CMBS pool. Property-level forecasts also allow real-time forecasting: As CoStar researchers update data for a property, the forecast will update as well to reflect the new information. The property-level model provides a structure for handling specific types of properties differently, such as affordable-rate housing, or owner-occupied office properties, for example. These nuances produce rich, textured forecasts that can take into account many more variables than traditional market-level models. Finally, and most obviously, property-level forecasts mean that analysts and underwriters can use a forecast for the

76

subject property in question, rather than relying on a market- or submarket-level forecast that may or may not capture the particulars of the asset. This section begins with a review of the various approaches academics and practitioners have used to forecast commercial real estate, and then describes how CoStar produces property-level forecasts, starting with the exogenous economic and capital markets inputs. Next, we will discuss the market model, which produces the guidelines that individual properties will follow. Finally, we explain the model that creates the property-level forecasts. The section also includes an overview of CoStar’s backtesting and model validation work around the forecasting model.

6.1 APPROACHES TO COMMERCIAL REAL ESTATE FORECASTING The topic of forecasting has busied economists for decades now, earning an extensive library of academic work. However, scare commercial real estate data limited the application of econometric techniques to the sector; those interested often had to construct their own data series. Hugh Kelly’s 1983 study, “Forecasting Office Space Demand in Urban Areas,” used annual supply and vacancy data for New York City to estimate demand, and then linked demand to office-using employment trends. This formative work established the basic framework for space market analysis and forecasting still used today. John Clapp formalized Kelly’s demand models by directly linking employment growth and demand in his 1989 study, “Absorption Forecasts Using Employment and Population Growths. Dr. Clapp’s 1993 monograph, “Dynamics of Office Markets,” summarizes his approach as well as other early research on commercial real estate. David Birch adopted a different approach to demand in his 1989 study, “America’s Office Needs, 1985-1995.” Dr. Birch accurately predicted the overbuilding of the late 1980s, and the subsequent deep recession in the early 1990s. He pioneered the concept of usage factors—the amount of space each employee uses—across different industries, and used his estimates of usage factors in conjunction with employment forecasts by sector to forecast demand. William Wheaton and Raymond Torto, the co-founders of Torto Wheaton Research, led the investigation into real estate rent dynamics. Their 1988 paper, “Vacancy Rates and Future of Office Rents,” continues to serve as the basis for rent forecasting today. They also proposed a simultaneous equation system of econometric models and identities to forecast the space market, a system they used commercially to produce forecasts for their industry clients. Quantitative forecasting models since, including those used by Property & Portfolio Research and used today by CoStar, follow the basic structure established by Drs. Wheaton and Torto. Patric Hendershott applied these ideas to international markets, while introducing new econometric approaches. He used interlinked regressions in his 1996 study of the Sydney office market, “Rental Adjustment and Valuation In Overbuilt Markets: Evidence

77

from the Sydney Office Market,” and applied similar ideas to London in his 1999 paper, “The Workings of the London Office Market” (with Colin Lizieri and George Matysiak). He pioneered the use of panel models to rents in his 2002 paper, “Explaining Real Commercial Rents Using an Error Correction Model with Panel Data” (with Brian Macgregor and Michael White). Sotiris Tsolacos (the Director of European Research for CoStar from 2006 to 2013) also advanced the study of European commercial real estate markets. His (along with Tony McGough and Éamonn D’Arcy) 1997 paper “National Economic Trends, Market Size and City Growth Effects on European Office Rents” used a panel approach to modeling rent trends. The trio followed up with their 1999 paper, “An Econometric Analysis and Forecasts of the Office Rental Cycle in the Dublin Area.” Dr. Tsolacos and Dr. Chris Brooks also published a textbook dedicated to real estate modeling and forecasting. Commercial real estate pricing and cap rates have received somewhat less attention from academics, perhaps due to the relative paucity of data. Dr. Hugh Norse used ACLI data to measure the impact of changes to income tax laws on cap rates in his 1987 study, “1966-1984: A Test of the Impact of Income Tax Changes on Income Property.” Jeffrey Fisher outlined the first structure for jointly analyzing commercial real estate space and capital markets in his 1992 paper, “Integrating Research on Markets for Space and Capital.” Dr. Fisher collaborated with Susan Hudson-Wilson and Charles Wurtzebach in 1993 on a related study, “Equilibrium in Commercial Real Estate Markets: Linking Space and Capital Markets.” Dr. William Wheaton and Serguei Chervachidze used NCREIF data to study the “great compression” and subsequent reversal of U.S. cap rates from 2000 to 2009 in their 2010 paper, “What Determined the Great Cap Rate Compression of 2000-2007, and the Dramatic Reversal During the 2008-2009 Financial Crisis?” They found that macro forces, much more than local real estate fundamentals, drive cap rate movements—a finding that CoStar has built into our forecasting models. John Affleck (an author of this paper and CoStar’s Vice President of Market Analytics) found similar conclusions in his investigations of the predictability of commercial real estate returns. Using NCREIF data, he used current capital markets pricing indicators to forecast excess returns over the next five years, and found cap rate levels, spreads, as well as risk spreads and the S&P 500 yield, to be significant predictors in his 2012 paper, “Cap Rates vs. Spreads: Is CRE Under or Overvalued.” He also explicitly tested the statistical relationship between the 1-year treasury rate and cap rates in his 2013 paper, “Do Interest Rates Matter? CRE in a Rising Rate Environment,” finding interest rates to be a marginal predictor, but less important than risk spreads and S&P 500 earnings yields.

78

6.2 ECONOMIC AND CAPITAL MARKETS DATA Commercial real estate forecasts start with exogenous economic and capital markets inputs. Traditionally, demand models have used job growth, often specific to the particular property type, as the primary independent variable in the demand model on the theory that jobs translate into real estate demand. Capital markets inputs like interest rates and risk spreads drive price and cap rate models. We have designed our forecasting systems to run a wide range of scenarios, potentially including custom economic forecasts. This section discusses the economic and capital markets inputs used, as well as the economic scenarios that CoStar runs each quarter.

6.2.1 Economic and Capital Markets Inputs

CoStar’s forecasting models operate at a metro (or regional) level, and as such require metro-level economic inputs. CoStar receives these inputs from Oxford Economics. Headquartered in London with offices around the world, Oxford Economics employs more than 250 economists and analysts serving leading multinational companies, financial groups, real estate organizations, governments, central banks and academic institutions worldwide. Long regarded as the industry leader in city, regional and other sub-national economic forecasting globally, Oxford Economics recently expanded its US forecasting coverage to include more than 3,500 sub-national economies, spanning all 50 states, 382 metros and 3,142 counties.

6.2.2 Scenarios

Oxford Economics produces a range of economic forecasts, including a Baseline scenario, Moderate Up and Downside scenarios, and a Severe Downside case. In addition, CoStar also creates an Interest Rate Shock scenario, which uses the Moderate Downside inputs, coupled with the highest interest rate recorded across all the Oxford scenarios. CoStar also runs an in-house scenario that we call the Trend Growth scenario. In the Trend Growth scenario, we make two simple assumptions. First, we assume that the labor force in each metro will continue to grow at the average rate of the past three years. Second, we assume that the unemployment rate in each metro will revert to the post-2000 average over the next five years, and then level off. From these assumptions we can calculate job growth and offer clients a scenario in which recent trends largely persist.

79

6.3 FORECAST “GUIDELINES”: THE MARKET MODELS CoStar’s market-level forecasting models, described in detail below, differ from other forecasting approaches in that they do not produce the final forecast results. Rather, the outputs of the market models provide guidelines that, all else equal, all properties in the market will follow. The market models employ a series of interlinked regression models addressing supply, vacancy, rent, and cap rates. From the results of these models, we mathematically derive demand, NOI, and price change. Exhibit 65 presents the market-level forecasting process.

Exhibit 65: Market Guideline Forecasting System

We run these regression models for the 54 largest metros. We also run panel regressions using all the metros. The panel models serve as back-up models in situations where the metro model does not produce useful results, either due to incorrect relationships between variables, low statistical significance, or poor r-squared measures. The sections below discuss each model in turn.

80

We group together the remaining, smaller metros into regional groupings and forecast the entire region together. These regional forecasts serve as the guidelines for the properties in these smaller, lower-tier metros.

6.3.1 The Supply Model

The market-level forecasting process begins with a forecast of supply, defined for this purpose as the change in total stock. Two steps make up the supply model: First, an estimate of construction starts, and second, a translation of those starts into completions. These steps correspond to the Construction Starts Model and the Deliveries Equation in Exhibit 65. Since buildings take some time to complete—many years for urban towers—the near-term supply forecast can rely on known projects already underway, rather than a modelled outcome. Thus, the supply forecast consists of two types of supply: Known supply—the properties currently underway and for which research has collected a completion date, and the econometrically-generated modeled supply.

6.3.1.1 Known Supply

As of Spring 2019, CoStar was tracking some 12,000 properties currently under construction for which we know size and the expected completion date. CoStar’s research teams regularly review under-construction projects as to their status and the likelihood of meeting the delivery date. Our forecasts use this information directly to produce the near- and medium-term supply outlook. However, we note that the supply, in the aggregate, can appear optimistic in the near term. Exhibit 66 shows industrial deliveries for the entire United States back to 2000 and the supply currently underway. Based on the expected completion dates, the next quarter (2018Q1) would see a significant supply shock relative to the recent past. However, the near-term spike results from delays in the completion of properties with delivery dates in the past; these properties default to deliver in the next quarter.

81

Exhibit 66: Industrial Construction, Historical and Expected

In reality, however, some of these projects will be delayed. We recognize the likelihood of a delay by moving forward by one month the completion dates for every property under construction. Since the forecast uses a quarterly frequency, the one-month shift has the effect of moving roughly one-third of the supply into the next quarter (since December deliveries will shift to January and a new quarter, but October and November will remain in the fourth quarter). Exhibit 67 shows the industrial supply outlook after this shift:

Exhibit 67: Industrial Construction, Historical and Expected With Date Shift


0M

20M

40M

60M

80M

100M

120M

140M

160M

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Quarterly Deliveries (millions of SF)

Known projects underway


0M

20M

40M

60M

80M

100M

120M

140M

160M

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23



82

The shift brings near-term deliveries more in line with historical norms, removes the unlikely supply spike in the next quarter, and results in a more accurate outlook.

6.3.1.2 Multifamily Supply

At the time of writing in April 2019, a large number of multifamily projects are underway, and a large number of these projects are expected to deliver in the next quarter, as shown in Exhibit 68:

Exhibit 68: Multifamily Construction, Historical and Expected

As shown, the expected deliveries for next quarter, based on researched and curated delivery dates, represent about four times the typical quarterly deliveries. This “wall of supply” has increased every quarter during this period of heightened multifamily construction. Over-optimistic developers, delays of nearly complete projects, and the reality that some percentage of delivery dates, however accurate on an individual basis, will slip, account for this unrealistic outlook. History has shown that only a fraction of these scheduled deliveries actually complete in the next quarter as expected. Those that do not deliver as expected are assumed to deliver in the next quarter—exacerbating the supply spike. As a result, this “wall of supply” has steadily risen throughout this cycle. The one-month shift solution applied to commercial property completion dates does not alleviate the multifamily supply problem, as it only spreads the “wall” over two quarters; an ideal solution will spread out the deliveries over the next year or even longer. To effect such a distribution, we first rank the scheduled deliveries into a “queue” for each market (or region, for smaller markets), ordering them by size, scheduled delivery date, by time under construction in descending order.


0M

50M

100M

150M

200M

250M

300M

350M

400M

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23



83

We then assume that these deliveries will follow the shape produced by the supply model, described in the section above. This model estimates quarterly deliveries as a share of the space underway. Given the large amount of construction in progress now, the model shows a “bump” in modeled deliveries in the near-term. In effect, the model assumes that only a share—typically about one eighth—of the supply underway will deliver each quarter. We allocate the queued properties across this distribution such that the last property allocated to each quarter causes the total space to exceed the modeled deliveries. Exhibit 69 graphically shows this dynamic for the New York apartment market, where the line represents the modeled supply, the superimposed bars show the scheduled deliveries, and the blue bars show the redistributed supply.

Exhibit 69: Adjusted Multifamily Construction

6.3.1.3 Modeled Supply

A supply forecast that relies only on known supply will systematically understate total new supply, since over the course of the next several years developers will plan for, propose, start, and ultimately complete new projects. Thus, an accurate forecast of supply must make estimates as to how much new space will deliver beyond the projects tracked by CoStar research. Traditionally, supply models have used completions of new supply as the independent variable, since most datasets for commercial real estate track completions. These models typically use real estate performance metrics like rent growth, vacancy, or price change, and capital market variables as independent variables. However, such models must necessarily include lags in the independent variables to account for the construction time between the decision to start a new project and when the project actually delivers. These lags in the independent variables result in problematic models, for several reasons: First, imprecision around the construction times—and as a result


0

5,000

10,000

15,000

20,000

25,000

30,000

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Historical Deliveries Scheduled Future Deliveries Redistributed Deliveries Modeled Deliveries


84

the length of the lag between completions and the independent variables—lowers the statistical power of the model. Second, the defined lag can create breaks and jumps in the data. For example, if deliveries today depend on interest rates two years ago, and interest rates jumped in that quarter and then returned to normal levels in the next quarter, then the modeled supply forecast today would show a corresponding movement. Finally, the build time in different markets, for different property types, and at different times, can vary widely, and models of completions fail to account for such nuance. CoStar’s property-level data not only includes the completion date for every property, but also the start date for most. Therefore, CoStar can directly model the decision to start construction projects, using contemporaneous variables instead of multi-year lags, resulting in much stronger models. The starts model includes lagged change in rent and the lagged natural log of the vacancy, represented as:

Equation 1: 𝑆𝑡𝑎𝑟𝑡𝑠𝑡 = 𝑓(Δ𝑅𝑒𝑛𝑡𝑡−1, 𝑙𝑛𝑉𝑎𝑐𝑡−1)

Where 𝑆𝑡𝑎𝑟𝑡𝑠𝑡 is construction starts, Δ𝑅𝑒𝑛𝑡𝑡−1 is the percentage change in rents, and 𝑉𝑎𝑐𝑡−1 is the lagged natural log of the vacancy rate. We restrict the model to the period since 2008, to capture the trends during the most recent expansion. We can then model the amount of construction underway by starting with the space underway today (the last historical date) and adding modeled starts and subtracting modeled deliveries, as in:

Equation 2: 𝐶𝑜𝑛𝑠𝑡𝑡 = 𝐶𝑜𝑛𝑠𝑡𝑡−1 + 𝑆𝑡𝑎𝑟𝑡𝑠𝑡 − 𝐷𝑒𝑙𝑖𝑣𝑡 Where 𝐶𝑜𝑛𝑠𝑡𝑡 is the amount of space underway at time t, 𝐶𝑜𝑛𝑠𝑡𝑡−1 is the amount of space underway in the prior period, 𝑆𝑡𝑎𝑟𝑡𝑠𝑡 is construction starts (from Equation 1), and 𝐷𝑒𝑙𝑖𝑣𝑡 is modeled deliveries. We estimate deliveries by applying the historical average of deliveries as a share of space underway in each market, as denoted in Equation 3:

Equation 3:

𝐷𝑒𝑙𝑖𝑣𝑡 = 𝐶𝑜𝑛𝑠𝑡𝑡

1

𝑛(∑

𝐷𝑒𝑙𝑖𝑣ℎ

𝐶𝑜𝑛𝑠𝑡ℎ

𝑛

ℎ=0

)

Where 𝐷𝑒𝑙𝑖𝑣𝑡 is the deliveries of new space at time t, 𝐶𝑜𝑛𝑠𝑡𝑡 is the amount of space

underway at time t, and h denotes historical time periods. The final supply forecast includes both the known and modeled supply. Exhibit 67 presents the supply outlook for U.S. industrial, showing both the known supply and modeled supply. Typically, the forecast calls for very little modeled supply in the first few quarters of the forecast; in the case of industrial, the very short build times result in some modeled supply even in the first forecast quarter.

85

Note that the supply model separately forecasts demolitions, also based on the average amount of space demolished each quarter.

Exhibit 70: Industrial Construction, Historical and Forecast, Including Modeled Supply

6.3.2 The Vacancy Model

With the supply forecast established, we now turn to demand, defined for this purpose as the net change in occupied space. We have two options by which we can produce a demand forecast: We can model demand directly and derive the vacancy rate, or we can model the vacancy rate and derive demand. Commercial real estate forecasts have traditionally modeled demand directly as a function of jobs. This approach makes complete sense, and indeed CoStar’s forecasting group has often employed such models. However, demand models suffer from some deficiencies. First, demand models tend to feature large constants. In other words, demand doesn’t change very much quarter to quarter, regardless of the economic inputs. This is because demand historically does not vary much quarter to quarter, and because demand is almost always positive. As a result, a forecast based on a demand model tends to call for very steady demand right around the long-term average, with only minor variations due to the economy. The “large constant” problem inherent in demand models makes scenario analysis challenging. Even in severely adverse economic scenarios, demand tends to remain steady, since the coefficients on the economic inputs have a relatively small effect and cannot overcome the large constant built into the model.


0M

20M

40M

60M

80M

100M

120M

140M

160M

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Historical Deliveries Modeled Supply Known Deliveries



86

Second, demand models do not necessarily take into account supply; the models assume steady demand more or less regardless of the amount of space delivering, or the vacancy rate. Thus, the vacancy forecast (derived from the supply and demand forecast) can be quite volatile in situations where supply and demand differs in a particular period. Moreover, a forecast based on a demand model does not automatically constrain the vacancy forecast between 0% and 100%. We have designed our forecasting process to run automatically with minimal human intervention, and to produce out-of-the-box scenarios and custom forecasts. With these goals in mind, we have chosen not to use a traditional demand model. Instead, we model the vacancy rate directly, and derive the demand. Our vacancy model begins with establishing a long-term steady-state vacancy rate, using a regression of the vacancy level against total job growth in the market (or region, for smaller markets). This steady-state forecast will serve as an independent variable in the second step of the vacancy forecast. Mathematically, this model is represented as

Equation 1: 𝑉𝑎𝑐𝑡 = 𝑓(Δ𝐸𝑚𝑝𝑡)

Where 𝑉𝑎𝑐𝑡 is log vacancy and Δ𝐸𝑚𝑝𝑡is the first difference of log employment. With the long-term steady-state vacancy rate established, the second step of the vacancy forecast model estimates the quarterly change in the log vacancy rate as a function of the trailing four-quarter employment level relative to a trailing 20-quarter median, the change in stock, and the lagged predicted residual from Equation 1 above, as denoted below:

Equation 2: Δ𝑉𝑎𝑐𝑡 = 𝑓(𝐸𝑚𝑝 , Δ𝑆𝑡𝑜𝑐𝑘𝑡−1, 𝑣𝑡−1) Where Δ𝑉𝑎𝑐𝑡 is the first difference in log vacancy, 𝐸𝑚𝑝 is the ratio of the trailing four-quarter average employment to the trailing 20-quarter median employment level, Δ𝑆𝑡𝑜𝑐𝑘𝑡−1, is the lagged first difference of log stock, and 𝑣𝑡−1 is the lagged predicted residual from Equation 1. The use of the trailing median in the vacancy model arises from the Covid period, when the second quarter of 2020 recorded job loses of 15 million—the worst since the end of the Second World War. The use of a trailing median in the denominator effectively removes such outliers from the trend, while the use of a trailing four-quarter average in the numerator smooths out, delays, and extends the effect of such outliers on vacancy. CoStar has tested the model specifications and found these variables to produce the low model error across all markets and also to have the high-quality backtesting results. The lagged predicted residual from Equation 1 ensures that the long-term forecast reverts to the steady-state vacancy rate established by Equation 1. This error-correction step also improves backtesting results.

87

To arrive at the demand forecast, we simply multiply the forecast supply by the forecast occupancy rate (equal to 1 – vacancy).

6.3.3 The Rent Model

Rents tend to move with vacancies, as illustrated in Exhibit 71 which presents the office same-store rent and the vacancy rate for the National Index of the 54 largest markets in the United States.

Exhibit 71: United States Office Same-Store Rent and Vacancy

Consequently, forecasters typically model rent change as a function of vacancy. Such models tend to produce good results; refinements include adding the recent trajectory of vacancy rates to capture momentum, and lagged rent levels such that rent growth slows as rents reach higher levels (and vice versa). As with vacancies, CoStar first establishes a long-term rent trend forecast, by estimating the rent level as a function of the occupancy rate and total employment, denoted as

Equation 1: 𝑅𝑒𝑛𝑡𝑡 = 𝑓(𝑂𝑐𝑐𝑡−1, 𝐸𝑚𝑝𝑡) Where 𝑅𝑒𝑛𝑡𝑡 is log rent, 𝑂𝑐𝑐𝑡−1 is lagged log occupancy, and 𝐸𝑚𝑝𝑡 is log employment. The rent change forecast includes this steady-state rent level from Equation 1, along with the first difference of log occupancy and the trailing four-quarter employment change, denoted as:

Equation 2: Δ𝑅𝑒𝑛𝑡𝑡 = 𝑓(Δ𝑂𝑐𝑐𝑡−1, Δ𝐸𝑚𝑝 , �̂�𝑡−1)

3/31/2000

6/30/2000

9/30/2000

12/31/2000

3/31/2001

6/30/2001

9/30/2001

12/31/2001

3/31/2002

6/30/2002

9/30/2002

12/31/2002

3/31/2003

6/30/2003

9/30/2003

12/31/2003

3/31/2004

6/30/2004

9/30/2004

12/31/2004

3/31/2005

6/30/2005

9/30/2005

12/31/2005

3/31/2006

6/30/2006

9/30/2006

12/31/2006

3/31/2007

6/30/2007

9/30/2007

Souce: CoStar As of December 2017

0%

2%

4%

6%

8%

10%

12%

14%

$0

$5

$10

$15

$20

$25

$30

$35

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Rent Vacancy

Same-Store Asking Rent Vacancy Rate

88

Where Δ𝑅𝑒𝑛𝑡𝑡 is the first difference in log rent, Δ𝑂𝑐𝑐𝑡−1 is the first difference in lagged log occupancy, Δ𝐸𝑚𝑝 is the trailing four-quarter change in employment, and �̂�𝑡−1 is the lagged predicted residual from Equation 1. The models for all property types are estimated using historical total employment. However, the forecast uses the change in retail employment as the driver for retail rents; the change in office-using employment for office rents; and the change in total employment for multifamily and industrial rents. The use of sector-specific employment forecasts for office and retail captures the divergence in outcomes between the property types, particularly during the Covid period. Note the CoStar’s rent change model does not include inflation. Many models estimate the change in real rents and then add a forecast of inflation to the outlook. Our analysis shows that, since 2000, inflation has had little impact on rents; only multifamily shows a statistically meaningful (but still low) relationship to inflation. Moreover, we have found that models trained on nominal rents have higher goodness-of-fit measures and produce better backtesting results since 2000. Finally, if we choose to forecast real rent growth, then we need a forecast for inflation in order to product nominal rents, introducing another source of forecast error.

6.3.4 Performance Models: NOI, Cap Rates, and Prices

From rents, we transition to forecasting market- and regional-level guidelines for the real estate “performance” variables: net operating income (“NOI”), cap rates, and prices. These three variables form the essential commercial real estate equation,

𝑌𝑖𝑒𝑙𝑑 =𝑁𝑂𝐼

𝑉𝑎𝑙𝑢𝑒

Where 𝑌𝑖𝑒𝑙𝑑 is the initial yield, 𝑁𝑂𝐼 is income, and 𝑉𝑎𝑙𝑢𝑒 is the estimated value of an asset. Based on this framework, to forecast commercial real estate performance we will model two of these variables, and then derive the third as the mathematical residual of this equation. Which two of these metrics variables to model, and which one to derive, ultimately is a philosophical choice. Over the years, CoStar’s analytics team has experimented with all the permutations of this equation and arrived at some guiding principles. First, we would like NOI to have a mathematical connection to our fundamentals and rent data. Therefore, we prefer to model NOI directly as some function of rent trends and occupancy for the market, submarket, or even property in question. This leaves us which the choice of cap rates or prices to model directly. Traditionally, the firm modeled price change, typically as a function of rent change, vacancy, and capital market indicators like interest rates and risk spreads. Such models tend to work

89

well, showing good statistical indicators and resulting in forecasts that broadly track fundamentals. However, in price change models, capital market variables like interest rates and risk spreads tend to have less influence in the model. As a result, price change models do not perform well in credit shock environments like 2008 and 2009: Instead of abrupt price changes brought on by an acute rise in risk spreads, prices tend to follow the more gradual decline in rents. Price change models also can create diverse outcomes across markets, since price trajectories in each market depend on the fundamentals and rent dynamics unique to that market. While some may see a diversity of performance outcomes as a laudable feature, such results also confound our ability to accurately model a 2008-style credit market recession, which affected all markets, largely without regard to fundamentals. With this in mind, we would prefer that markets have more unified outcomes, all else equal. Our other option is to model cap rates directly. Models of cap rates show strong statistical relationships to risk spreads and interest rates, while rent growth and fundamentals have less power in the model. These results satisfy our preference that capital market conditions drive our performance forecast. And, since capital market indicators do not vary by market (interest rates and risk spreads are both national indicators), cap rates also produce more unified outcomes across markets. With this in mind, our performance forecast starts with a forecast of NOI based on rent and vacancy. We then forecast cap rates, using a model dependent primarily on capital market conditions. We do not need to produce a market guideline for price change, since we will compute a price forecast for every individual property based on the NOI and cap rate forecasts for each property. The sections below outline in detail our approach to metro-level NOI and cap rate forecasts.

6.3.4.1 NOI Model

Net operating income refers to the cash flows generated by a property before taxes and depreciation. In other words, potential rental income, less vacancy, and less fixed and variable operating expenses. A typical multi-tenant asset will house occupiers who have signed leases at various points in the cycle at different rents. The rental income will depend on when these leases were signed, and at what rent levels, while the growth in income will depend on scheduled escalations on in-place leases, and on the rate at which existing leases expire and new leases are signed at higher rents. As with our historical price estimates, explained in section 5.3, we estimate NOI as the trailing four-quarter average rent multiplied by current occupancy. Note that in the past, we had used different trailing rent period depending on the property type. A longer period of average trailing rents, perhaps for office, delays the effect of rent movements

90

on NOI, and therefore on values. As a result, the use of varying periods for trailing rents produces different value trends across property types. During the Covid period, we made the decision to apply the same trailing four-quarter average to all property types to ensure that all property types would produce the same price change given the same rent, vacancy, and cap rate trends.

6.3.4.2 The Cap Rate Model

As noted in the introduction to this section, we choose to model cap rate levels as a function of secular capital market conditions, the local economy, and metro-level rent growth. The model specification for the cap rate model is:

𝐶𝑎𝑝𝑅𝑎𝑡𝑒𝑡 = 𝑓(𝑅𝑖𝑠𝑘𝑆𝑝𝑟𝑒𝑎𝑑𝑡, 𝑈𝑅𝑡, Δlog (𝑅𝑒𝑛𝑡𝑡) Where 𝑅𝑖𝑠𝑘𝑆𝑝𝑟𝑒𝑎𝑑𝑡 denotes the spread between the Baa yield rate and the 10-year Treasure yield at time t, 𝑈𝑅𝑡 denotes the local unemployment rate at time t, and

Δlog (𝑅𝑒𝑛𝑡𝑡) denotes the percentage change in rents at time t. Note that the use of the risk spread removes from the model any assumption about secular trends in interest rates; we are effectively assuming that the risk-free rate remains constant at the last historical level.

6.3.5 Interventions to the Market Model

As a matter of policy, CoStar will not make any direct interventions in the market- and regional-level models, either by altering input variables, entering add factors, or by adjusting constants or model coefficients. At the time of writing in Summer 2020, we believe the models to be producing reasonable and statistically sound outputs for use as guidelines in the property-level model. Should the model begin to produce results that do not appear reasonable to us, to the CoStar analysts and economists who cover markets, or to our clients, we will reexamine the models and potentially make adjustments to model specifications or parameters so as to produce results more in line with expectations. Depending on the degree to which we change the models, we may make a public announcement about these changes. During the Covid period in Spring 2020, we found that the extreme economic data resulting from the cessation of economic activity during the lockdown did produce unreasonable CRE forecasts. In response, we made several changes to our model, which we communicated to our clients via regular email updates. These changes are enumerated in section 1.1 of this paper.

91

6.4 THE BUILDING-LEVEL FORECAST MODEL Commercial real estate forecasters have traditionally provided forecasts for markets or submarkets. Forecasters lacked either the property-level information necessary to forecast individual buildings or the analytic wherewithal to develop, test, and deploy at scale property-level models. No longer: In January of 2018, CoStar began forecasting key real estate variables for all 6 million buildings across the apartment, office, industrial, flex, and retail property types. Several considerations led CoStar to develop a property-level forecasting apparatus. First, we believe property-level forecasts will be useful to our clients, whether to underwrite deals, value collateral, or manage their portfolio, and that a forecast that considers the conditions of an individual property is better than mapping a property to a market- or submarket-level outlook. Second, having a forecast for every property allows us and our clients to produce forecasts for any set of properties by aggregating the individual properties. In the past, any new level of aggregation, whether a submarket slice or custom grouping of submarkets, required a new model. Now, the property-level forecast serves as the single model by which we can produce any custom forecast. Finally, the property-level forecasting structure offers limitless avenues for refinement and sophistication. This initial release serves as a version 1.0, to which we can add modules to handle various types of properties and situations, including single-tenant properties, rent-controlled apartments, industry and tenant dynamics, granular economic and demographic trends, and much more. No market model, however sophisticated, can ever take into account all the possible building-level dynamics. As a result, aggregating individual property-level forecasts into a market or submarket forecast produces a deeply textured forecast that accounts for the idiosyncrasies of a market or submarket in a way that no market-level model possibly can. Exhibit 73 presents a schematic showing the property-level forecasting process. The process begins with the market (or regional, for smaller markets) guidelines produced by the market models, denoted by the gray and blue boxes on the left that correspond to the metro models shown in Exhibit 65. These guidelines include forecasts for supply (both known and modeled); for aggregate demand; for rent; and for NOI, cap rate, and price change. The property-level forecast process then shares the guideline down to the submarket slice level (denoted as the blue boxes), and finally to the individual properties (denoted by the orange boxes). The multiple layers ensure that all properties in a particular geography broadly follow the same trend, and also offer an efficient way to alter the forecasts for a large number of properties by intervening at a submarket cluster or submarket level.

92

Exhibit 73: The Property-Level Forecasting Process

6.4.1 Building-Level Supply Forecasts

The market-level model provides a forecast of aggregate modeled supply, which we will allocate across the submarket slices and then to individual properties. To allocate the supply across submarkets, we consider two factors: the amount of space currently under construction plus deliveries of space over the past two years, and the amount of proposed space in each submarket slice, where the proposed space counts one-tenth of the under-construction space for the purposes of allocating supply. This has the effect of putting the modeled supply into submarkets with significant recent construction and/or construction underway, with a secondary effect from proposed space. We assume all new space will be 4 & 5 Star. In the rare market with no construction underway, no recent deliveries, and nothing proposed, we distribute the modeled supply pro rata to the amount of existing stock in each submarket slice.

93

Allocating the modeled supply to individual buildings poses a particular challenge, since we do not know where or in what form the modeled supply will arrive. We solve this problem by adding a small amount of modeled supply to each property in each forecast quarter, in proportion to each building’s share of the submarket slice total stock. This solution effectively distributes the modeled supply across the submarket in relation to the current distribution of 4 & 5 star properties. Attaching the modeled supply to properties allows us to return an estimate of future supply for any custom geography or set of properties. The model also allocates the pool of demolitions across submarkets. We assume that only lower-quality properties are demolished, and that demolitions tend to occur in submarkets with the highest levels of construction.

6.4.2 Allocating Demand Across Buildings

As with supply, the market-level model returns a forecast of aggregate demand for the market. Our property-level forecast must allocate this aggregate demand in the same geographical cascade, first to submarket cluster slices, then to submarket slices, and finally to individual properties. The first allocation of demand goes to new construction, both projects currently underway and the modeled supply later in the forecast. We assume a standard lease-up pattern based on our analysis of CoStar’s historical data in the first two years of a property’s life. Existing properties can not only receive new demand—they can also lose their current occupancy. To capture the possibility of a property losing occupancy, we deduct from every property a small percentage of its occupancy, and add this “churn” to the aggregate pool of demand. The churn percentage varies by slice to recognize the relative inelasticity of some occupancy (in high-quality or owner-occupied offices, for example). Each existing property receives an allocation of the aggregate demand (including the churn) based on its vacancy rate and its rent relative to its peers. All else equal, higher vacancy properties receive more demand, as will lower rent properties. Properties that receive less demand than their churn will see vacancies rise. Exhibit 74 shows an example of vacancy trajectories for the ten 3 Star office properties in Chicago:

94

Exhibit 74: Property-Level Vacancy Forecast Example

Generally, higher-vacancy properties receive more demand, while lower-vacancy or completely rented properties lose some demand, depending on the rent in the property and its location within the market.

6.4.3 Building-Level Rent Forecasts

A particular building’s rent forecast depends on the following factors: the metro rent guideline, shared down to the submarket slice for the property’s location; the property’s vacancy forecast; the property’s rent level relative to its peers; the recent momentum in rent growth; and the property’s relative performance over the past two years. Vacancy affects the rent forecast in the expected way: Higher vacancy properties will have less rent growth, and vice-versa. Similarly, properties with higher rent levels than the submarket slice average will have lower rent growth. The final two factors have to do with the property’s recent history. Momentum maintains the recent performance, such that properties with rising rents will see rents continue to rise in the near-term. Relative performance works in the opposite way: Rent growth in properties that have outperformed the submarket slice average over the past two years will experience slowing rent growth in the outer years of the forecast. Exhibit 75 presents rent forecasts for the same set of Chicago office properties. Note the divergence in rent growth across properties, depending on each property’s vacancy, rent level, and recent performance.

PropertyID

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626


0%

5%

10%

15%

20%

25%

30%

10 11 12 13 14 15 16 17 18 19 20 21 22 23

Vacancy Rate

History Forecast

95

Exhibit 75: Property-Level Rent Forecast Example

6.4.4 Building-Level Performance

Property-level forecasts of NOI, price, and cap rate offer tantalizing possibilities. Underwriters no longer need to plug in a submarket or market forecast of cash flows or exit cap rates—they can simply use the forecasts for the building in question. Investors can quickly arrive at a forecast for a loan pool, portfolio, or REIT. Building-level forecasts of real estate performance data also fulfills our vision of comprehensive historical and forecast data for every property, allowing us to create custom forecasts for any set of properties without regard to history or forecast. This section details how we apply the market guidelines to individual properties to produce building-level forecasts for these variables.

6.4.4.1 Building-Level NOI forecasts

We create a building-level NOI forecast by modulating the market NOI guideline by the building’s rent forecast relative to the guideline. Thus, a property with a rent forecast that outperforms the guideline by 1% will also have an NOI forecast that outperforms the market by 1%. This may appear to be an oversimplification, as it omits any effect from occupancy or move-ins and move-outs. However, most investors apply a pro forma NOI when pricing a property, assuming that they will lease-up the property to full occupancy. Therefore, we believe the simple application of expected rent growth most closely approximates the NOI expectations an investor would use in underwriting a deal. Moreover, for the purposes of forecasting value changes, we prefer a more volatile measure, such as rent growth. We expect to refine these assumptions in upcoming revisions to the model, with the eventual goal that our property-level NOI forecasts might be useful initial inputs into underwriting models.

PropertyID

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626

626


$0

$10

$20

$30

$40

$50

$60

10 11 12 13 14 15 16 17 18 19 20 21 22 23

Same-Store Asking Rent

History Forecast

96

6.4.4.2 Building-Level Cap Rate Forecasts

We make a similarly simplified assumption about property-level cap rates: We simply attach the market guideline trend to each property by adjusting the level. This has the effect of producing different rates of percentage change in cap rates depending on the cap rate level of the property. Properties with lower cap rates will experience large percentage changes, on average, and higher cap rates properties will have lower volatility. We believe this dynamic to be consistent with historical performance trends suggesting that lower cap rate properties show more volatility, both on the downside and especially on the upside. Exhibit 76 presents cap rate forecasts for a set of office properties in Washington, D.C.

Exhibit 76: Property-Level Cap Rate Forecasts

6.4.4.3 Building-Level Price Forecasts

We can now derive a forecast of property-level price trends by dividing the change in NOI by the change in cap rates, per the essential commercial real estate equation:

Δ𝑉𝑎𝑙𝑢𝑒 =Δ(𝑁𝑂𝐼)

Δ𝑁𝑎𝑡𝐶𝑎𝑝

By modulating individual property NOI series based on rent and arriving at different percentage changes in cap rates by adjusting the level of the market guideline, we create broadly similar outlooks in prices, but with enough variation to produce interesting results. Exhibit 77 presents the price forecasts for the same set of office properties in Washington, D.C.

select

into

from

left join

left join

where

order by newid()

drop table #history

select

into #history

from


4%

5%

6%

7%

8%

9%

10%

11%

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Estimated Cap Rate

97

Exhibit 77: Property-Level Cap Price Forecasts

6.4.5 Interventions to the Building-Level Forecasts

By design, we have not implemented the ability to make adjustments to specific properties. As noted, we do incorporate all known move-ins and move-outs into the property-level forecasts. We have built the capability to make adjustments, using add factors, to cluster or submarket forecasts. We do not anticipate using these adjustments extensively; we added the capability to make interventions in order to recognize known future events that we cannot enter into the CoStar data in any other way. These events could include known future demolitions or conversions (like the former U.S. embassy in Grosvenor Square in London); the anticipated arrival of a large new tenant for which we don’t know the specific property into which they will move (for example, Amazon’s HQ2); or acts of God or nature, like the recent hurricanes in Houston and Florida. We document and systematically review all such interventions. Upon request, we will share such interventions with our clients.

select

into

from

left join

left join

where

order by newid()

drop table #history

select

into #history

from


$0

$100

$200

$300

$400

$500

$600

$700

$800

$900

$1,000

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Estimated Price per Square Foot

98

6.5 BACKTESTING AND MODEL VALIDATION A key measure of the usefulness of any forecasting model is how well the model would have predicted events in the past. We have designed both the market and property-level models around such tests. In particular, we recognize that many of our clients, particularly large financial organizations, use the 2009 Great Financial Crisis experience as a test of the resiliency of their investments, and we have designed our models around this period. CoStar’s quantitative team has extensive experience assisting large financial institutions with model validation. We have worked closely with model risk management groups, and understand the requirements many institutions have regarding the testing and validation of third-party models. To assist such groups as they work to validate our models, we have prepared, in addition to this White Paper, a suite of more technical documentation outlining the statistical performance and testing of our models. We have also engaged a third-party consulting firm to produce an independent validation of our forecasting models. We will provide these materials and related technical assistance to advisory clients with contract provisions related to model validation. The section below provides a synopsis of our model validation work.

6.5.1 Backtesting the Market Models

Model backtesting refers to the practice of statistically analyzing how well a model would have performed in an out-of-sample test: that is to say, using a model based on data to a point in history (say, 2008) and given real economic inputs, how well would the model have predicted the actual outcomes in the succeeding years (2009 and after). The goal of backtesting is to minimize the out-of-sample error; in doing so, we hope that we will also minimize the forecast error. To backtest the models, we use data from 2000 to 2007 to calibrate the models. We then increment the calibration period by one quarter, creating a total of 38 out-of-sample forecasts (at the time of writing). Exhibit 78 graphically shows the model performance for Miami apartment rents, where the heavy black line denotes the actual series and the thin gray lines denote the 38 quarterly out-of-sample forecast made in each quarter starting in 2007Q3, using actual economic data.

99

Exhibit 78: Property-Level Rent Forecasts

6.5.2 Backtesting the Building-Level Model

To backtest the building-level model, we assumed perfect inputs for employment, construction dates, and metro demand and rent, in order to see how much of the final error comes from the building forecast instead of from imperfect inputs (we did not include the lease information in the out of sample testing, as doing so would unreasonably bias the results toward extra accuracy). Using those inputs, we ran out of sample testing, from 2012-2017 and from 2008-2013, looking primarily at the accuracy of submarket aggregates rather than individual buildings. We focused on submarket error since much of the error at the building level comes from the inherent choppiness of building-level data, whereas the forecast is, by design, inherently smooth. Moreover, we will capture some of the choppiness in the building-level forecast by including the actual known move-ins and move-outs. Using this testing structure, we optimized the various inputs into the building-level models, including the effects of momentum and reversion; rent and vacancy levels; and churn. We will continue to use this testing structure to further refine our property-level models.


$1,000

$1,100

$1,200

$1,300

$1,400

$1,500

$1,600

$1,700

$1,800

00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17

Same-Store Asking Rent

100

7 THE FUTURE

At the outset of this millennium, CoStar began providing analytic data for millions of properties across more than 40 major markets. For the first time, clients could see vacancy and asking rents for any individual property, or set of properties. We believe the information and transparency CoStar brought to this infamously opaque asset class helped produce this golden age of commercial real estate investment. CoStar’s entrée into the multifamily space with its acquisition of Apartments.com in 2014 brought about a similar revolution. Drawing on a dataset of more than 4 billion rent data points and vacancy data for more than 150,000 properties, CoStar provides the most accurate, granular, and timely data for the multifamily sector. We believe the advances outlined in this paper—same-store estimated rents for every property, building-level price and cap rate estimates, powerful new market models, and building-level forecasts for all key commercial real estate variable—represent the next evolution of information and analytics for the asset class. Same-store rents, based on millions of individual rent data points, present the truest picture of market rent trends, from individual properties to markets large and small to the nation overall. Property-level price and cap rate trends not only provide an objective, comps-based estimate of market value for every property, but also allow us to quickly arrive at market capitalization estimates for any market or portfolio or REIT. And forecasting every property, rather than only forecasting markets and submarkets, frees us from the tyranny of geography and removes the wall between history and the future. Most of all, though, the models and frameworks outlined in this document provide a platform for ongoing and unlimited refinement. We look forward to hearing feedback from our clients, and incorporating your ideas to improve and enrich these models. Please feel free to contact us directly using the contact information on the final page. In addition to the quantitative group responsible for these processes, more than 70 analysts and economists make up the CoStar Market Analytics team. Together, they produce thousands of market and submarket reports using the data series produced by these analytic processes. And CoStar Portfolio Strategy, CoStar’s advisory and consulting arm, uses CoStar’s remarkable data to answer specific research questions and develop investment theses. Our collective dedication to the highest-quality data, analytics, and market coverage serves CoStar’s vision of a more transparent, efficient real estate marketplace.

101

8 APPENDIX

8.1 CHANGES IN THIS VERSION Version 3.0 of this paper, released in July 2020, updated figures, made minor rewrites, and the following changes: 1.1: New section outlining changes to our models made in response to the

Covid outbreak Exhibit 3: Updated with new data Exhibit 4: New exhibit Exhibit 13: Updated with new data 5.4.2 Added Exhibit 65 6.2.2 Updated to reflect current set of scenarios 6.3.1.3 Updated to reflect change in time period used in supply model 6.3.2 Updated to reflect change in vacancy model demand driver 6.3.3 Updated to reflect change in economic input to rent model 6.3.4.1 Updated to reflect change in trailing rent period in NOI model; deleted

Exhibit 72 6.3.4.2 Updated to reflect use of risk spread in cap rate model 6.3.5 Updated to address changes during Covid period

8.2 CHANGE IN PRIOR VERSIONS Version 2.1 of this paper, released in July 2019, corrected typos and grammatical errors. Version 2.0 of this paper, released in April 2019, made several additions, clarifications, changes, and corrections to the prior version. Exhibit 8: Clarified Slice definitions 4.2.2.3: Added section on backcasting of multifamily data 4.2.2.4: Added comment on why the multifamily time series becomes more volatile

over time 4.3.1: Added attributions of changes in available space-weighted rent series 4.3.3: Added equations outlining construction of same-store rents 4.3.4: Added section on how we estimate rents for commercial properties for

which we have never collected rent information 4.3.5: Added section comparing same-store rent series to available space-

weighted series 5.1: Added review of prior approaches to estimating commercial real estate

performance trends 5.2: Added section on CoStar transaction data and presentations of simple

averages of transaction prices

102

5.3.3: Updated estimation of large untraded properties to include adjustment coefficients

5.3.5 Updated accuracy results 5.4.1 Updated cap rate regression models 6.1 Added literature review of forecasting approaches 6.2.1 Updated to reflect use of Oxford Economics data and the CoStar Trend

Growth scenario

103

John Affleck, Vice President, Market Analytics [email protected]

Mike Taylor, Forecasting Director [email protected]

Rob Jennings, Senior Quantitative Analyst [email protected] These Costar Group, Inc. materials contain financial and other information from a variety of public and proprietary sources. CoStar Group, Inc. and its affiliates (collectively, “CoStar”) have assumed and relied upon, without independent verification, the accuracy and completeness of such third party information in preparing these materials. The modeling, calculations, forecasts, projections, evaluations, analyses, simulations, or other forward-looking information prepared by CoStar and presented herein (the “Materials”) are based on various assumptions concerning future events and circumstances, which are speculative, uncertain and subject to change without notice. You should not rely upon the Materials as predictions of future results or events, as actual results and events may differ materially. All Materials speak only as of the date referenced with respect to such data and may have materially changed since such date. CoStar has no obligation to update any of the Materials included in this document. You should not construe any of the data provided herein as investment, tax, accounting or legal advice. CoStar does not represent, warrant or guaranty the accuracy or completeness of the information provided herein and shall not be held responsible for any errors in such information. Any user of the information provided herein accepts the information “AS IS” without any warranties whatsoever. To the maximum extent permitted by law, CoStar disclaims any and all liability in the event any information provided herein proves to be inaccurate, incomplete or unreliable. © 2020 CoStar Realty Information, Inc. No reproduction or distribution without permission.

mailto:[email protected]



Documents

CoStar Data and Methodology · 2021. 1. 5. · CoStar’s data provides the raw material for a comprehensive view of U.S. commercial real estate. Per CoStar’s 10-K, as of January