Liquidity as Social Expertise

• Author: PABLO KURLAT
• Date: 01 April 2018
• DOI: http://doi.org/10.1111/jofi.12606
• Published date: 01 April 2018

THE JOURNAL OF FINANCE •VOL. LXXIII, NO. 2 •APRIL 2018

Liquidity as Social Expertise

PABLO KURLAT∗

ABSTRACT

This paper proposes a theory of liquidity dynamics. Illiquidity results from asymmet-

ric information. Observing the historical track record teaches agents how to interpret

public information and helps overcome information asymmetry.However, an illiquid-

ity trap can arise: too much asymmetric information leads to the breakdown of trade,

which interrupts learning and perpetuates illiquidity. Liquidity falls in response to

unexpected events that lead agents to question their valuation models (especially

in newer markets) may be slow to recover after a crisis, and is higher in periods of

stability.

THIS PAPER PROPOSES A THEORY OF MARKET LIQUIDITY and its evolution over time.

This theory is based on the interaction between information asymmetry and

social learning.

Liquidity is an elusive concept, with the literature on it plagued by challenges

of both definition and measurement.1In this paper, I refer to the following

notion of liquidity: The liquidity of an asset class is the fraction of the potential

gains from trade in that asset class that are realized in equilibrium. If the

potential gains from trade are large, it is tautologically true that asset liquidity

has important consequences for social welfare.

The theory in this paper is based on a minor modification of an otherwise

standard model of trade under asymmetric information in the spirit of Akerlof

(1970). The assumption is that rather than being purely uninformed, asset

buyers have access to some information, but their ability to make use of that

information depends on their collective experience, which I refer to as “social

expertise.” The model abstracts from the specific features of the assets that are

traded. However, to fix ideas, it is useful to think about the model as applying

to markets such as the market for initial public offerings (IPOs), the primary

market for asset-backed securities, or the primary market for sovereign bonds.

∗Pablo Kurlat is with the Department of Economics, Stanford University. An early version

of parts of this paper circulated under the title ‘Lemons, Market Shutdowns and Learning’. I

am grateful to Manuel Amador, Marios Angeletos, Ricardo Caballero, Bengt Holmstr¨

om, Peter

Kondor, Juan Pablo Nicolini, Fabrizio Perri, Victor Rios-Rull, Jean Tirole, Iv´

an Werning, and

various seminar participants for helpful comments. I have read the Journal of Finance’s disclosure

policy and have no conflicts of interest to disclose.

1See Brunnermeier and Pedersen (2009) for a discussion of the conceptual issues and Goyenko,

Holden, and Trzcinka (2009) for a discussion of measurement.

DOI: 10.1111/jofi.12606

619

620 The Journal of Finance R

The model works as follows. Each period, “sellers” own assets of heteroge-

neous quality that are independent over time. These assets are more productive

if held by “buyers” rather than sellers, so there are gains from trade. Sellers

know the asset qualities but buyers cannot observe them, so a classic lemons

problem arises.

Buyers can alleviate their information disadvantage by observing public,

asset-specific signals. However, for these signals to be useful, buyers need to

know how signals correlate with asset qualities. The key assumption in the

model is that the joint distribution of signals and asset qualities, summarized

by the single parameter μ, is not known exactly. Agents can learn about μfrom

commonly observed data, based on a sample of past assets. For each of them,

the data contain the signal it produced when it was created and an indicator

for how it turned out. The idea is that agents turn to the historical track record

in making sense of the information available on any given asset, which is stan-

dard practice among financial analysts. For instance, the popular textbook by

Damodaran (2008) provides guidance on the use of “comparables” to determine

which pieces of information about a particular asset one should focus on and

shows how to use them in valuation. One of the challenges, in practice, is find-

ing a sufficiently large and sufficiently similar sample of historical precedents.

In terms of the model, the question is at what rate are observations added to

the historical sample. I assume that this rate depends positively on the volume

of trade, a form of learning-by-doing.

The model delivers several predictions about the relationship between infor-

mation, trading, and liquidity.The first result is that if information asymmetry

is sufficiently severe, liquidity is increasing in the precision of agents’ estimates

of μ. Knowing μincreases traders’ ability to extract information from signals,

reducing the degree of information asymmetry and increasing liquidity. Hence,

liquidity is a function of traders’ expertise. Asset prices, the level of investment,

and the volume of trade are increasing in liquidity.

Depending on parameters, the model may feature an “illiquidity trap.” If at

any point in time, estimates of μare sufficiently imprecise, assets will be com-

pletely illiquid and trade will break down. If the learning process is such that

data only are generated by trading, markets will generate no data for agents

to learn from, which will perpetuate the illiquidity. Whether the economy falls

into an illiquidity trap depends on the sample realizations during the first few

periods. If the first observations lead to precise and correct estimates of μ,this

will increase liquidity and reinforce the learning process, which becomes self-

sustaining. If instead the first observations lead to imprecise estimates of μ

because they conflict with each other or with agents’ prior, signals will become

uninformative, which will lead to the illiquidity trap. Even under parameters

such that there is no permanent illiquidity trap, market liquidity can be slow

to recover after a disruption.

The model also predicts that markets will tend to become more liquid over

time, as traders accumulate more observations with which to estimate μ.Inthe

short run, unexpected events disrupt liquidity because they increase buyers’

uncertainty as to whether they are using the correct model (i.e., the correct

Liquidity as Social Expertise 621

value of μ) to evaluate assets. This increases information asymmetry and low-

ers liquidity. Furthermore, unexpected events will be more disruptive in newer

markets, where traders have not had time to accumulate a long track record of

observations and thus revise their beliefs more strongly in light of new infor-

mation.

I next extend the model to allow the structure of the economy, as captured by

μ, to change over time. In this case, unexpected events will be more common

in times of structural change and therefore liquidity will be higher when the

underlying economy is more stable. Even in the long run, liquidity will remain

fragile.

The dynamics of the model follow from positive feedback between learning

and trading. The fact that trading generates observations to learn from is sim-

ply an assumption, though I provide examples of underlying models that could

give rise to it. The main contribution of the paper is to show how,despite the fact

that asset payoffs are independent over time, past observations can generate

useful expertise that alleviates information asymmetry and increases liquidity.

This paper relates to several strands of literature. First, it builds on the

literature following Akerlof (1970) on how asymmetric information can cre-

ate barriers to trade (Wilson (1980), Kyle (1985), Glosten and Milgrom (1985),

Levin (2001), Attar, Mariotti, and Salani´

e(2011)). The main contribution rela-

tive to this literature is to explore the effect of more precise knowledge about

the information structure on trade in a simple model that is quite close to

Akerlof’s (1970) basic setup.

Second, the paper relates to a large literature on social learning. The form of

learning of interest here, namely, learning that is dependent on economic ac-

tivity, has been studied by Veldkamp (2005), van Nieuwerburgh and Veldkamp

(2006), Ordo˜

nez (2009), and Fajgelbaum, Schaal, and Taschereau-Dumouchel

(2014). These stories focus on agents who need to learn the level of aggregate

productivity to make production and investment decisions. This paper focuses

instead on agents who learn the right way to evaluate assets, and studies the

consequences this has for market liquidity. The mechanics of the illiquidity

trap are also related to the informational cascades studied by Banerjee (1992)

and Caplin and Leahy (1994).

Finally, the paper contributes to the literature on the sources of liquidity

fluctuations. These include theories based on asymmetric information (Daley

and Green (2012), Cespa and Foucault (2014), Dang, Gorton, and Holmstr¨

om

(2015), Daley and Green (2016)), theories based on Knightian uncertainty or

non-Bayesian learning (Hong, Stein, and Yu (2007), Caballero and Krishna-

murthy (2008), Routledge and Zin (2009), Uhlig (2010)), and theories based

on balance sheet effects (Shleifer and Vishny (1992), Holmstr¨

om and Tirole

(1997), Kiyotaki and Moore (1997), Brunnermeier and Pedersen (2009)). In my

model, the liquidity of an asset class can fluctuate even though all agents are

Bayesian, none are financially constrained, and the distribution of payoffs for

the asset class remains unchanged.

The rest of the paper is organized as follows. Section Idescribes the model.

Section II characterizes the static outcomes of the model. Section III presents