Asymmetric Learning from Financial Information

• DOI: http://doi.org/10.1111/jofi.12223
• Date: 01 October 2015
• Published date: 01 October 2015
• Author: CAMELIA M. KUHNEN

THE JOURNAL OF FINANCE •VOL. LXX, NO. 5 •OCTOBER 2015

Asymmetric Learning from Financial

Information

CAMELIA M. KUHNEN∗

ABSTRACT

This study asks whether investors learn differently from gains versus losses. I find

experimental evidence that indicates that being in the negative domain leads indi-

viduals to form overly pessimistic beliefs about available investment options. This

pessimism bias is driven by people reacting more to low outcomes in the negative

domain relative to the positive domain. Such asymmetric learning may help explain

documented empirical patterns regarding the differential role of poor versus good

economic conditions on investment behavior and household economic choices.

DO INVESTORS LEARN THE same way when they face positive outcomes as when

they face negative outcomes? Do economic agents form beliefs using the same

learning rules during recessions as during booms? Converging findings from

finance and neuroscience suggest that this may not be the case.

Recent empirical finance work indicates that learning by market partici-

pants may differ depending on whether the economic conditions are good or

bad. Economic downturns are characterized by stronger reactions to negative

news by equity markets, higher risk premia, and more pessimistic expectations

by corporate executives (Andersen et al. (2007), Bollerslev and Todorov (2011),

Ben-David, Graham, and Harvey (2013)). Poor stock market outcomes receive

disproportionately pessimistic press coverage (Garcia (2012)). Households that

witness bad economic times become reluctant to invest in equities and have

pessimistic beliefs about future stock returns (Malmendier and Nagel (2011)).

After floods or earthquakes, people are more likely to buy insurance against

∗Kuhnen is with the Kenan-Flagler Business School at the University of North Carolina

at Chapel Hill. I thank Bruno Biais (the Editor), two anonymous referees, Nicholas Barberis,

Marianne Bertrand, Cary Frydman, Cam Harvey, David Hofmann, Eric Hughson, Jonathan

Parker, and Richard Todd, as well as seminar participants at Northwestern University, Univer-

sity of Washington, University of Utah, Santa Clara University, University of Oregon, Caltech,

University of Southern California, Massachusetts Institute of Technology, Stanford University,

Arizona State University, UCLA, University of Minnesota, UC Davis, Duke University, Univer-

sity of North Carolina, New York University, and the Consumer Financial Protection Bureau and

participants at the 2011 Society for Neuroeconomics annual meeting, the 2012 Western Finance

Association meeting, the 2012 Boulder Consumer Finance meeting, the 2012 Miami Behavioral

Finance conference, the 2013 UNC Jackson Hole Winter Finance conference, the 2013 Behav-

ioral Economics annual meeting, and the 2013 NBER Household Finance Summer Institute for

helpful comments and discussion. Alexandra Baleanu provided excellent research assistance. All

remaining errors are mine.

DOI: 10.1111/jofi.12223

2029

2030 The Journal of Finance R

such events, even though the probability of their occurrence does not change

(Froot (2001)). This empirical evidence suggests that bad times, characterized

by a preponderance of negative outcomes, may have a particularly strong in-

fluence on people’s beliefs about the future.

Moreover, neuroscience evidence indicates that the brain processes deployed

when people learn from their environment differ depending on whether they are

faced with positive or negative outcomes (Kuhnen and Knutson (2005), Knutson

and Bossaerts (2007)). Memory processes are different for details related to

positive contexts than for those related to negative contexts (Eppinger, Herbert,

and Kray (2010), Mather and Schoeke (2011)), in that negative contexts lead

to a more narrow focus than positive ones. People’s emotional reactions are

stronger in the face of losses, relative to gains, and this is particularly true

when the stakes are higher (Sokol-Hessner, Camerer, and Phelps (2013)). This

biology-based evidence suggests that people perceive and incorporate negative

outcomes differently than positive ones.

Here, I use an experimental setting to examine whether people indeed learn

differently from gains or positive news relative to losses or negative news. I

find that, when they are in the negative domain, people form overly pessimistic

beliefs about the available financial assets, particularly if they are actively

investing. This pessimism bias is driven by an overreaction to low outcomes in

the negative domain relative to the positive domain. These results are robust

to alternative explanations and they replicate out of sample.

The idea that learning may be different in the gain and loss domains is dif-

ferent from and complementary to the well-documented phenomenon of loss-

aversion suggested by Kahneman and Tversky (1979), whereby the disutility

of losing an amount of money is greater, in absolute terms, than the utility

of winning that amount. A large body of work has provided evidence for this

difference in preferences in the gain and loss domains. The findings that I doc-

ument here suggest that gains and losses are different not only in terms of how

they shape the value function, but also in terms of how they are incorporated

in the formation of beliefs.

To investigate whether learning is different when people face negative out-

comes relative to when they face positive ones, adult participants from a U.S.

university were invited to a study that required the completion of two financial

decision making tasks. In the Active task, subjects made 60 decisions, split into

10 separate blocks of six trials each, to invest in one of two securities: a stock

with risky payoffs coming from one of two distributions (good and bad), one

that was better than the other in the sense of first-order stochastic dominance,

and a bond with a known payoff. In each trial, participants observed the div-

idend paid by the stock after making their asset choice, and then were asked

to provide an estimate of the probability that the stock was paying from the

good distribution. In the Passive task, subjects were only asked to provide the

probability estimate that the stock was paying from the good distribution, after

observing its payoff in each of 60 trials, which were also split into 10 separate

learning blocks of six trials each. In either task, two types of conditions—gain or

loss—were possible. In the gain condition, the two securities provided positive

Asymmetric Learning from Financial Information 2031

payoffs only. In the loss condition, the two securities provided negative payoffs

only. Subjects were paid based on their investment payoffs and the accuracy of

the probability estimates provided.

Importantly, the learning problem faced by subjects was exactly the same

in gain condition blocks as in loss condition blocks. The only difference was

that the two possible stock payoffs had a minus sign in front of them in the

loss condition relative to the gain condition (i.e., −$10 or −$2 in the former

vs. +$2 or +$10 in the latter condition). Hence, people’s estimate regarding

the probability that the stock was paying from the good dividend distribution,

namely that distribution where the high outcome for that condition was more

likely to occur than the low outcome, should not depend on whether they are in

a block where they learn from negative outcomes, or in one where they learn

from positive outcomes.

However, I find that subjects learn differently in the gain condition relative to

the loss condition. Subjective probability estimates that the stock is paying from

the good dividend distribution are 3% to 5% lower in the loss condition than in

the gain condition, controlling for the objective Bayesian posterior probability

that the stock is the good one given the dividends observed by participants.

That is, subjective beliefs about the risky asset are overly pessimistic in the

loss condition. Moreover, the deviation of subjective probability estimates from

the objective Bayesian posterior that the stock is the good one is 2% to 4%

larger in the loss condition relative to the gain condition. In other words, belief

errors are on average larger when people learn from negative payoffs than from

positive ones. The pessimism bias and resulting larger deviation in subjective

posteriors from Bayesian beliefs in the loss relative to the gain condition are

generated by the fact that people update more from a low outcome in the loss

condition (i.e., a −$10 dividend) than in the gain condition (i.e., a +$2 dividend).

There is no difference between the two conditions in terms of updating beliefs

from high outcomes (i.e., either −$2 or a +$10 dividend, in the loss and gain

conditions, respectively).

I then conduct several robustness tests and find that the loss versus gain

condition effect on subjective beliefs is robust in-sample, whether I analyze data

from early or late learning blocks, or, within each learning block of six trials,

from early or late trials, or whether I conduct my analysis with or without

subject fixed effects.

Moreover, I show that the effect replicates out of sample, in a population

more than twice as large as in the original group of subjects, and in a different

country (Romania). There, too, I find that the loss condition induces larger

errors in subjective beliefs, and an overreaction to low outcomes, just as was

found in the U.S. sample.

Finally,I examine several alternative explanations for the documented learn-

ing effects induced by the loss versus gain context. I test whether in the loss

condition, relative to the gain condition, people may start with different pri-

ors that the stock is good, whether they may have different risk attitudes and

whether their beliefs may have a different impact on asset choices across the

two contexts. Finally, I test whether the experimental task that I use in fact