Two Biases, One Model: Barberis, Shleifer and Vishny on Investor Sentiment

By the late 1990s, empirical finance had produced two facts that appeared to be flatly contradictory, and it was becoming embarrassing.

Fact one: markets underreact. After a company announces surprisingly good earnings, its stock drifts upward for months afterward. Jegadeesh and Titman had shown that winners over the past 3 to 12 months keep winning. The market seems to absorb news too slowly.

Fact two: markets overreact. De Bondt and Thaler had shown that over 3 to 5 years, the biggest losers beat the biggest winners. Companies with long strings of good news end up overpriced and subsequently disappoint. The market seems to absorb news too eagerly.

So which is it? Does the market react too little, or too much? Saying "both" sounds like a cop-out, an unfalsifiable dodge where you get to explain everything after the fact.

Barberis, Shleifer and Vishny's answer was to build a single, simple investor with a single, fixed psychology who does both automatically, and to derive both facts from one set of assumptions. That is the whole trick, and it is elegant.

The problem: you need one brain that makes both errors

Any model that just adds a fudge factor for underreaction and a separate fudge factor for overreaction has explained nothing. The bar is high: you need one consistent mental model, drawn from real psychology, that spits out short-horizon momentum and long-horizon reversal without being told to.

BSV reached for two biases that Tversky and Kahneman's literature had already documented in the lab:

• Conservatism. People update their beliefs too slowly when new evidence arrives. Confronted with one surprising data point, they cling to their prior view and adjust only a little. This is a well-documented laboratory finding, and it obviously produces underreaction.
• Representativeness. People see patterns in small samples. Give them a short run of similar outcomes and they conclude a trend exists and will continue, ignoring how easily a short run could happen by chance. This obviously produces overreaction.

The two biases pull in opposite directions. The genius of the paper is arranging them so each one dominates at a different horizon.

The key idea via analogy: the stubborn friend who suddenly becomes a fanatic

Everyone knows this person.

You tell them something that contradicts their worldview. Once. They shrug it off. "One data point, doesn't mean anything." That's conservatism. They barely update.

You tell them again. They budge slightly, grumbling.

But somewhere around the fourth or fifth time, something snaps. They flip completely and become a total convert. "It's obvious, it's a whole trend, it was always going to be this way, everyone can see it." That's representativeness. They have now seen a short run and concluded a permanent regime has arrived, and they promptly overshoot far past what the evidence supports.

Now here is the modelling move. BSV assume the investor believes that earnings follow one of two regimes, and he is always trying to figure out which one he is in:

• A mean-reverting regime, where good earnings are likely to be followed by bad ones. Random noise, nothing to see.
• A trending regime, where good earnings are likely to be followed by more good ones. A real streak.

Crucially, neither regime actually exists. In the model's true world, earnings follow a random walk: they are unpredictable, and past earnings tell you nothing about future ones. The investor is chasing a structure that isn't there. That is his whole error, and it is a beautifully realistic one, because human beings are pattern-detection machines running in an environment that is mostly noise.

And now watch the two facts fall out:

• A single earnings surprise arrives. The investor is mostly convinced he is in the mean-reverting regime, so he thinks this will probably reverse, and he barely moves the price. Underreaction. The good news is not fully in the price, so as more results come in and confirm the earnings level, the price keeps drifting up. That is post-earnings-announcement drift and momentum, produced from nothing but conservatism.
• A long string of good surprises arrives. Now the investor becomes convinced he has identified a trending regime. This company is a growth machine. He extrapolates, bids the price far above fair value. Overreaction. But the earnings were a random walk all along, so the streak inevitably breaks, the "trend" evaporates, and the price collapses back. That is long-horizon reversal, produced from nothing but representativeness.

One investor. One fixed psychology. Both anomalies. The horizon does all the work: a little news triggers conservatism, a lot of news triggers pattern-seeking.

Why it mattered

• It made behavioral finance a modelling discipline, not just a critique. Before this, behavioral finance was mostly a collection of embarrassing facts about the efficient market hypothesis. BSV showed you could take psychology, put it in a formal asset pricing model, and generate quantitative, testable predictions. That earns respect from people who otherwise dismiss the field.
• It answered the "you can explain anything" objection head-on. The most common attack on behavioral finance is that with enough biases in the drawer you can rationalize any pattern. BSV pre-empted this by using exactly two biases, both documented independently in psychology labs long before, and deriving both anomalies from a single specification. That is a much harder thing to do than it sounds.
• It was one of three simultaneous shots. 1998 and 1999 produced three competing behavioral models of the same puzzle: BSV (conservatism and representativeness), Daniel, Hirshleifer and Subrahmanyam (overconfidence and self-attribution), and Hong and Stein (slow information diffusion and momentum traders). They disagree about the mechanism but agree on the target. Reading all three together is the best short course in behavioral asset pricing there is.
• It reframes momentum as a bug in your own head. If BSV are right, momentum profits exist because everyone (including you) is too slow to update on a single piece of news. That is a much more useful mental model than "momentum works because it works".

The honest limitations

• The two biases were chosen to produce the answer. This is the sharpest criticism and it is fair. Conservatism and representativeness are real, but so are dozens of other documented biases, and BSV selected the pair that generates the known facts. It is not exactly reverse-engineering, but it is not far off. The defence is that the biases were documented first, in a different field, by people not trying to explain stock returns.
• The two biases are arguably in tension. Conservatism says you under-update on evidence. Representativeness says you over-infer from small samples. Those are close to opposite claims about how much weight people put on new data, and the model needs a fairly specific structure to make them coexist without contradiction.
• It is a single-asset, representative-agent model. There is one stock and one investor. Real markets have thousands of assets and a distribution of investor types, some of whom would be happy to trade against the sentiment. The model has no arbitrageurs at all, which is a large omission for a paper co-written by the author of The Limits of Arbitrage. The implicit assumption is that limits to arbitrage let the sentiment survive, and it is left to the reader to supply that.
• The regime parameters are not pinned down by anything. The model needs the investor to hold particular beliefs about how likely each regime is and how persistent it is. Different choices give different strengths of momentum and reversal. The theory tells you the shape of the effect, not its size.

The one-line takeaway

Barberis, Shleifer and Vishny built an investor who under-updates on single pieces of news (conservatism) but sees permanent trends in short streaks (representativeness), and showed that this one stubborn, pattern-hungry brain generates both short-run momentum and long-run reversal without any further tinkering.