Total Cost of PreservationCost Modeling for Sustainable Services

Stephen AbramsPatricia Cruse

John KunzeUniversity of California Curation Center

California Digital Library

CNI Spring 2012 Membership MeetingBaltimore, April 1-3, 2012

Outline Goals Prior work Modeling preservation activity Total cost of preservation

Pay-as-you-go price model Paid-up price model

Conclusions Questions and discussion

http://wiki.ucop.edu/display/Curation/Cost+Modeling

Source: Getty Images

Goals Understand costs in order to plan for and implement

sustainable preservation services Investigate the possibility of paid-up pricing in order

to address Boom-or-bust budget cycles Fixed-term, grant funded projects

Source: www.sharedidiz.com/

End date!

Prior work Nationaal Archief (2005)

http://www.nationaalarchief.nl/sites/default/files/docs/kennisbank/codpv1.pdf

LIFE (2008)http://www.life.ac.uk/

KRDS (2010)http://www.beagrie.com/krds.php

DataSpace (2010)http://arks.princeton.edu/ark:/88435/dsp01w6634361k

Jean-Daniel Zeller (2010)“Cost of digital archiving: Is there a universal model?”8th European Conference on Digital Archiving, Geneva, April 28-30, 2010 http://regarddejanus.files.wordpress.com/2010/05/costsdigitalarchiving-_jdz_eca2010.pdf

Rosenthal (2011)http://blog.dshr.org/2011/09/modeling-economics-of-long-term-storage.html

}Identification of granular cost components

}Assumption of annual decrease in aggregate cost, i.e., discounted cash flow (DCF)

Critique of DCF approach

Key assumptions Consider only the costs incurred by the preservation

service provider Costs of content creation by collection managers are out

of scope

Costs can be categorized unambiguously as fixed or marginal, and one-time or recurring One-time costs can be annualized over the effective

lifespan of the activity or system component

Cost model components

System, composed of various

Services for necessary/desirable functions, running on

Servers, deployed by

Staff, in support of content

Producers, who use

Workflows to submit instances of

Content Types, which occupy

Storage, and are subject to ongoing

Monitoring and periodic

Interventions

Number and unit cost of Producers

Total cost of preservation

ViMjSkCWmPnATCP

Fixed cost of System

Number and unit cost of Workflows

Unit cost and number of

Content Types

Number and unit cost of

Storage

Number and unit cost of Monitoring

Number and unit cost of

Interventions

System component subsumes Services

and Servers

Staff costs are subsumed by other

components

Total cost to service

provider

Total cost of preservation

ViMjSkCWmPnATCP

Model is rich enough to represent the full economic cost of preservation

Implemented by a spreadsheet that captures all subsidiary costs

Total cost of preservation

ViMjSkCWmPnATCP

Model is rich enough to represent the full economic cost or preservation

But service providers can customize the model to exclude components whose costs are not recoverable or are subsidized as a matter of local policy

Assumption: Cost allocation Cost of the Archive, Workflows, Content Types,

Monitoring, and Interventions are “common goods” Equally beneficial to all Providers Properly apportioned across all Providers

Cost of a single Producer

SkPn

ViMjCWmAG P

Number of Storage units attributable to

Producer

Number of Producers

Unit cost of a Producer

Total cost attributable to a given Producer

Assumptions: Billing Costs are billed for at the end of the period of

service The cost model should be revenue neutral

Pay-as-you-go cash flow

Expense

Income

GGt = 0 1 2 3

GCash flow diagram

G G G

TGGTGT

t

1

0

)(

Cost of a single Producer

Cumulative pay-as-you-go price over time period T

Pay-as-you-go price for a single

Producer

Cumulative pay-as-you-go price

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

$16,000$14,000$12,000

$10,000$ 8,000$ 6,000$ 4,000$ 2,000$ 0

Year (T)

Cost

($)

TGGTGT

t

1

0

)( )(G

Cumulative pay-as-you-goG (T )

Cumulative pay-as-you-go price over time period T … for “forever”

as a function of time T

Assumptions: Costs over time

Moore’s law, 1971 – 2011Source: Wikipedia

Kryder’s law, 1980 – 2012Source: Wikipedia

The aggregate cost of providing preservation service decreases over time; and that decrease is uniform Moore’s and Kryder’s laws

$$

Assumptions: Costs over time The aggregate cost of providing preservation service

decreases over time; and that decrease is uniform Moore’s and Kryder’s laws State-of-the-art tools and understanding Productivity increases

Discounted pay-as-you-go cash flow

(1–d )2·GG (1–d )·GDiscounting

factor

t = 0 1 2 3

Expense

Income

Discounted cash flow

(DCF) diagram

G (1–d )·G (1–d )2·G

tT

t

dGdTG )1(),(1

0

Discounted pay-as-you-go price over time period T

Cost of a single Producer

Pay-as-you-go price for a single

Producer

as a function of time TDiscounted pay-as-you-go price

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

$16,000$14,000$12,000

$10,000$ 8,000$ 6,000$ 4,000$ 2,000$ 0

Year (T)

Cost

($)

dd T

GdTG 11),(dGdG ),(

(1-d)t discount factor

Discounted pay-as-you-goG (T,d )Discounted pay-as-you-goG (,d )

Cumulative pay-as-you-goG (T )

… for “forever”Discounted pay-as-you-go price over time period T

Discount factor d is the weighted sum of the expected changes in

number and unit cost of individual components

Weighting factors ω are the proportion that a particular component contributes to the aggregate cost G, e.g.

)()( CCWmWAA dddddd )()()( SkSPPViVMjM ddddddd

GnA

A

GnWm

W

Drawbacks to pay-as-you-go pricing Only viable for Producers with reliable annual

funding sources Boom-or-bust budgeting or the termination of

funded project work can interrupt this funding Any interruption in proactive preservation care can

lead to irretrievable data loss

Assumptions: Investment return Preservation service providers can carry forward

budgetary surpluses across fiscal years Surplus funds can be invested with the return

supplementing the surplus

Paid-up cash flow

t = 0 1 2 3

Expense

Income

(1–d )2·G(1–d )·G

F r ·F

F Surplus (1+r )·F –G

r ·[(1+r )· F –G ]

(1+r )· [(1+r )·F –G ]–(1–d )·G

r ·[(1+r )·[(1+r) ·F–G ]–(1–d )·G ]–(1–d )2·G

(1+r )·[(1+r )· [(1+r )· F –G ]–(1–d )·G ]–(1–d )2·G 1

1

0 )1(1),,(

t

tT

t rdGrdTF

G

Paid-up price for time period T

Paid-up price Investment return

Cost of a single Producer

as a function of time TPaid-up price

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30

$16,000$14,000$12,000$10,000$ 8,000$ 6,000$ 4,000$ 2,000$ 0

Year (T)

Cost

($)

drrdr

T

TT

GrdTF

)1()1()1(),,(

drGrdF

),,(

(1–d)t discount factor

(1+r)t investment return

Paid-up price, for TF (T,d ,r)Paid-up price, for F (,d ,r)

Discounted pay-as-you-goG (T,d )Discounted pay-as-you-goG (,d )

Cumulative pay-as-you-goG (T )

… for “forever”Paid-up price for time period T

Paid-up example

Pay-as-you-go price, G $ 650 (1 TB) Discount factor, d 5% Investment return, r 2% Term, T 10 years Paid-up price, F $ 4,725

Year Income Expense Surplus0 $ 4,725.00 – $ 4,725.001 $ 94.32 $ 650.00 $ 4.285.322 $ 83.39 $ 617.50 $ 3,764.213 $ 72.70 $ 586.63 $ 3,262.294 $ 62.43 $ 557.29 $ 2,778.425 $ 52.53 $ 529.43 $ 2,311.526 $ 42.99 $ 502.96 $ 1,859.55

7 $ 33.79 $ 477.81 $ 1,422.538 $ 24.91 $ 453.92 $ 816.529 $ 16.33 $ 431.22 $ 401.63

10 $ 8.03 $ 409.66 $ 0.00

< $ 5,216 < $ 6,500dr

Coefficient of permanence It is useful to be able to transition from a pay-as-

you-go to a paid-up price basis If you’re currently paying G on a pay-as-you-go

basis, you can upgrade to a paid-up basis with a

one-time payment of F = G ·φ , where Princeton DataSpace, φ ≈ 30 (T = ) USC digital repository, φ ≈ 1.2 (T = 20)

dr

1

Problems with R&D TCP modeling is dependent on the predicative

reliability of r and d For d, extrapolate from Moore’s and Kryder’s laws?

Moore’s law, 1971 – 2011Source: Wikipedia

Kryder’s law, 1980 – 2012Source: Wikipedia

??

Problems with R&D TCP modeling is dependent on the predicative

reliability of r and d For d, extrapolate from Moore’s and Kryder’s laws? For r, extrapolate from 30 year Treasury bonds?

30 year treasuries, 2007 – 2012Source: http://ycharts.com/indicators/30_year_treasury_rate

30 year treasuries, 1882 – 2012Source: Robert Schiller

?

Model the risk Round up r and d, i.e., adding a fixed “risk premium” Add an additional risk component R to the formula

for G Its influence on the price can grow over time, reflecting

increasing uncertainty, by setting a negative discount factor dR so that 1–dR > 1 Note, however, that if the weighted sum d becomes less

than 0 and |d | > r then G (T ) will not converge to a limit

SkPn

ViMjCWmAG P

+ R

Recalibrate the model G and F do not have to be fixed values over time

Periodically recalculate based on current conditions (actual costs for G ) and predictions (r and d ), and apply prospectively

Retrospective service contracts remain “locked-in”

Hybrid price model Distinguish between costs that are (relatively) easy

to quantify and forecast, and those that aren’t Use the paid-up model for the former and pay-as-you-go

for the latter

Easy Difficult

Archive Intervention

Producer

Workflow

Content Type

Monitoring

Storage

Hybrid price model Distinguish between costs that are (relatively) easy

to quantify and forecast, and those that aren’t Use the paid-up model for the former and pay-as-you-go

for the latter

Bit preservation only

Easy Difficult

Archive Content Type

Producer Workflow

Storage Monitoring

Intervention

Preservation forever Some things are intended to last forever…

Source: John Church Company Source: United Artists

Preservation forever

?

Some things are intended to last forever…

Preservation for … A fixed term – 10 years? 20 years? – may be

appropriate for much content Give content an opportunity to prove its worth, as

evidenced by someone’s commitment to pay for its subsequent preservation

Embrace uncertainty The discounted cash flow (DCF) approach is

problematic on practical and theoretical grounds Difficulty in the setting fixed values for r and d that

realistically represent financial and technological trends over time

Stochastic modeling to determine the probability distribution of possible outcomes C.f., David Rosenthal, FAST ‘12

http://blog.dshr.org/2012/02/fast-2012.html

CNI Fall 2011http://www.youtube.com/watch?v=_5lQxmyz3xY

Understand the risks Possible outcomes…

We overestimate our costs and collect too much● Fund a higher level of service● Refund some portion

We underestimate● Ask for additional funds● Lower service levels● De-accession content – but at least it was preserved up to that

point and had a chance to prove its value, and gain an advocate

Conclusions Different customers have different funding capabilities

Flexibility in price models is important

Any price model is based on an idealization of the real world Assumptions matter

Knowing all of the costs is distinct from a policy decision to recover all of those costs

If investment return and discount factor can be reliably projected, then standard DCF/NPV methods provide a reasonable prediction of long-term costs

Looking for feedback

Thanks to our reviewers Lisa Baird, UCOP Raym Crow, SPARC Todd Grappone, UCLA Cliff Lynch, CNI David Minor, UCSD/SDSC Richard Moore, SDSC Michael Mundrane, UC Berkeley Jake Nadal, UCLA David Rosenthal, LOCKSS Mackenzie Smith UC Council of University Librarians /

Systemwide Operations and Planning Group Candace Yano, UC Berkeley

Have we missed something in our analysis, logic, assumptions, math, etc?

Are the objections to the DCF-based analysis substantial enough to invalidate this approach?

Are there better forecasting methodologies?

Even if we don’t have a perfect model, we need to move forward now with a “good enough” model

For more information

Total Cost of Preservation: Cost Modeling for Sustainable Serviceshttp://wiki.ucop.edu/display/Curation/Cost+Modeling

UC Curation Centerhttp://www.cdlib.org/uc3uc3@ucop.edu

Stephen Abrams Mark ReyesPatricia Cruse Abhishek SalveScott Fisher Joan StarrErik Hetzner Tracy SenecaGreg Janée Carly StrasserJohn Kunze Marisa StrongMargaret Low Adrian TurnerDavid Loy Perry Willett