Data and the Aggregate Economy · 2019. 6. 5. · Data as a by-product of economic activity Veldkamp (2005), Ordonez (2013), Fajgelbaum et al. (2017): data is only about aggregate

Data and the Aggregate Economy

Laura Veldkamp

Columbia Business School, NBER, and CEPR

[email protected]

2019

Laura Veldkamp (Columbia) Data and the Aggregate Economy 2019 1 / 27

Data in the Aggregate Economy

Data is information that can be encoded as binary sequence ofzeroes and ones

This includes a spectrum of knowledge, from music, topoetry, and patents

Big data revolution focuses on statistics and records oftransactions.

Production of information and communications tech (ICT) goodsand services ≡ 6.5% of global GDP, ∼ 100 million workers (UN

(2017))

What's dierent about an aggregate data econonmy?


Agenda

1 How data makes contact with many facets ofmacroeconomics

GDP measurement? Price exibility? Labor?

2 Existing theoretical tools to understand data

How data is similar to and dierent from technology?

Relevant tools from information frictions literature

3 Framework to explore eect of data on aggregateeconomic activity


Data is transforming the economy

GDP Measurement

Most data is bartered

Consumers give data in exchange for a service. How tovalue?

Gain in well-being not captured by GDP

Losing access to search engines or emails ≈ earning$500-$1,000 less per year (Brynjolfsson et al (2018))

Possibility for alternative measurements, such aswelfare-based GDP-B (Brynjolfsson (2019)) or cost-saving-basedEGDP (Hulten and Nakamura (2017))

Might explain productivity slowdown?

Yes: Brynjolfsson et al. (2017)

No, eect too small: Syverson (2017)


Data in International Economics

Will the data economy negate geography?... in Rwanda we are determined to take full advantage of the digital revolution.This revolution is summed up by the fact that it no longer is of utmost importancewhere you are but rather what you can do ..."

President Paul Kagame of Rwanda(Government of Rwanda (2010))

Data facilitates international trade

Machine translation increases exports by 17.5% through reduction insearch costs (Brynjolfsson et al. (2018))

Control of data is a signicant issue in trade negotiationsAre we exporting data assets at zero price to foreign entities?


Price Flexibility

Digital economy changes how rms adjust prices and hencemonetary policy eectiveness

Price changes occur more frequently in online stores than inregular stores (Gorodichenko and Talavera (2017))

Every 3 weeks or less for online stores

Every 4-5 months or more for regular stores


Firm Dynamics

Investment in IT boosts rms' eciency (Aral and Weill (2007)),especially large rms (Tambe and Hitt (2012), Brynjolfsson and McElheran

(2016))

Aects rm size distribution

Use of big data in nancial markets can lower the cost ofcapital for large rms (Begenau et al. (2018))

Market power?


Labor Demand

Will articial intelligence displace labor?

The Great Decoupling:" divergence between GDP and employment(Brynjolfsson and McAfee (2013)), (Furman and Seamans (2018))

Displacement eect vs. Productivity eect (Acemoglu and Restrepo

(2018))

AI ↓es labor, but productivity ↑ labor in non-automated tasksNet eect: A reduction in labor's income share.

Industry/ task heterogeneity: Some more suitable for machinelearning (Brynjolfsson and Mitchell (2017))


Tools to think sensibly about data

Is Data Like Technology?

Models of data as tech (Jones and Tonetti (2018), Agrawal et al. (2018), Lu

(2019), Aghion et al. (2017))

Although data and technology are both non-rival, they also dier

Creating technology require resources, while data is a by-productof economic activity

No incentives required to produce

While technologies leak (Easterly (2002)), data is not embodied inhuman capital. Less likely to diuse

Data can be easily monetized

Can be described without revealing information contentCan be split (999 data points vs. 1,000)

Can be sold directly or indirectly (data services, ad placement)


Ideas From Information Frictions

Rational inattention dictates what data a rm should use.Tools originally developed for data allocation.

Sims (2003), Cover and Thomas (1991), Mackowiak and Wiederholt(2009), Kacperczyk et. al. (2016), FMVV (2019)

Data as a by-product of economic activityVeldkamp (2005), Ordonez (2013), Fajgelbaum et al. (2017):data is only about aggregate productivity

Direct and indirect sale of informationAdmati and Peiderer (1990): Data service best because data usagehard to control

ML is learning about joint distributions: Unusual events havepersistent eects (Kozlowski, Veldkamp, Venkateswaran, 2019)

Information selection / bias: Nimark (2011), Gentzkow (2010, 2006)


With So Much Data, Why Use Theory?

Structural economic changes are changing covariances.

Answer policy questions.

Understand why changes might logically arise.

What features does a macro-data theory need?

Data is a by-product of economic activity

Firms with data benet from better forecasts

Data is a tradeable asset


A Macro Model of Data

Builds on Farboodi et al. (AEA P& P, 2019)

A continuum of competitive rms i

Each rm produces kαi ,t units of goods with ki ,t units of capital

These goods have quality Ai ,t

Firms take equilibrium price Pt as given and their quality-adjusted

outputs are perfect substitutes

Pt = PY−γt ,

Yt =

∫iAi ,tk

αi ,tdi


Model: Goods Quality and Information

Quality depends on chosen production technique ai ,t

Firm has one optimal technique: θi ,t + εa,i ,t

θi ,t is an AR(1), with innovation ηi ,t ∼ N(µ, σ2θ).εa,i ,t ∼ N(0, σ2a) is unlearnable and i.i.d.

Ai ,t = A− (ai ,t − θi ,t − εa,i ,t)2

Number of data points depends on t − 1 production:ni ,t = zik

αi ,t−1 z is data-savvyness.

Each data point m ∈ [1 : ni ,t ] revealssi ,t,m = θi ,t + εi ,t,m where εi ,t,m ∼ N(0, σ2ε ),


A Simple Recursive Solution

Firm Problem:

maxki,t ,ai,t

E0

∞∑t=0

βt(PtAi ,tk

αi ,t − rki ,t

)Solution:

Ωi ,t ≡ E[(E[θi ,t |Ii ,t ]− θi ,t

)2]−1(Posterior precision)

E[Ai ,t

]= A− Ω−1i ,t − σ

2a (Expected quality)

Lemma 1:

The optimal sequence of capital investment choices ki,t solves the recursive problem:

V (Ωi,t) = maxki,t

Pt

(A− Ω−1i,t − σ2

a

)kαi,t − rki,t + βV

(Ωi,t+1

)where ni,t+1 = zik

αi,t and

Ωi,t+1 =[ρ2(Ωi,t + σ−2a )−1 + σ2

θ

]−1+ ni,t+1 σ−2ε


Pricing and Valuing Data

Ωi ,t can be interpreted as the amount of data a rm has

∂Vt/∂Ωi ,t = Marginal value of one additional price of data

= price of data

The optimal price of good may be zero (data barter)

To generate transactions and accumulate data


Data Growth, Solow-Style

Inows: zikαitσ−2θ (# of data points × precision)

Outows (data depreciation): Ωit

ρ2(1+σ−2a Ω−1it )−1+σ2θΩit− 1

Steady state: inows = outows

0 0.5 1 1.5 2

Firm i data stock ( it)

0

0.5

1

1.5

data inflowdata depreciation


How Firm's Data Stock Grows Over Time

Takeaway: Data has increasing, then diminishing returns (Bajari

et al. (2018)).Data → more output → more data, until forecast error → 0.

0 0.5 1 1.5 2

time

0

0.5

1

1.5

Firm

i da

ta s

tock

(it)

Data convergence path to steady state


How Firms' Data Stock Grows Over Time

Three rms with dierent initial data stocks.

0 0.5 1 1.5 2time

0

0.5

1

1.5

Firm

i da

ta s

tock

(it)

Data convergence, with different initial data stocks

Divergence, then convergence. Inequality is transitory?


Data Growth, Aggregate Equilibrium

Inows and outows the same. Pt now adjusts.

0 0.5 1 1.5 2

Firm i data stock (it)

-0.2

0

0.2

0.4

0.6

0.8

1

1.2

1.4

data inflows and outflows, Pss

data inflowdata depreciation

Eqbm → Faster early growth bc initial goods price is high.


Tradeable Data (one rm, exclusive use)

Use precision ω at price π:ω < (>) n is selling (buying) data.

V (Ωi,t) = maxki,t ,ωit

Pt

(A− Ω−1i,t − σ

2

a

)kαi,t + π(ni,t+1 − ωit)

− rki,t + βV(Ωi,t+1

)Ωi,t+1 =

[ρ2(Ωi,t + σ−2a )−1 + σ2θ

]−1+ ωit

Immediate convergence:

1 2 3 4 5 6 7 8 9 10

time

0

0.5

1

1.5

Qua

ntity

of D

ata

data convergence path to st state

What are the frictions? Capital adjustment? Tech adjustment?


Data as a Business Stealing Technology

It is not clear that data is always productive. Advertising?

What if data used for business stealing? Suppose data processinghelps the rm that uses it, but has no social value:(Morris-Shin (2002))

Ai ,t = A−(ai ,t − θi ,t − εa,i ,t

)2+

∫ 1

j=0

(aj ,t − θj ,t − εa,j ,t

)2dj

This doesn't change rms' choices, or the relative rm dynamics,just the quality of goods and the welfare consequences.


Articial Intelligence (AI) and Knowledge Production

Big data is about new ways of analyzing data.How does that t in a model? (Abis Veldkamp `19)

Data and labor combine to make usable knowledge (nit signals).

AI changes the data/labor mix.

Models were artisanal. Now, modeling is automated.

nit = dαit lβit

AI increased α. Less diminishing returns to data d .(Like industrialization changed the capital-labor mix.)

Changes income shares, wages, returns to size.


Conclusions

Data is an important asset in the modern economy

Has some features of tech: non-rival

Key dierence: As forecast errors get smaller, most information hasdecreasing returns

Data is changing the allocation of rents in the economy ... butnot its fundamental economics

Lots of open questions for theory and data:How much of the superstar rm phenomenon is due to data economics?Has data changed the returns to capital?Policy questions: If rms have to pay customers for data, what then?


Questions & Comments


Documents

Data and the Aggregate Economy · 2019. 6. 5. · Data as a by-product of economic activity Veldkamp (2005), Ordonez (2013), Fajgelbaum et al. (2017): data is only about aggregate