When Analysts Get It Wrong: Expectation Bias, Market Inefficiency, and What It Means for Investors

For decades, the efficient market hypothesis (EMH) has been the central pillar of modern finance, asserting that asset prices fully and immediately incorporate all available information. If true, you cannot reliably beat the market using publicly known data—because that information is already priced in. But what if the problem isn’t the information itself, but how people process it?

Cheng Gao, Siyuan Ma, and Peixuan Yuan, authors of the February 2026 study “Expectation Bias and Short-term Momentum,” tackled this question. Rather than asking whether the market has access to good information, they ask whether investors and analysts systematically misinterpret it — and whether that misinterpretation is predictable enough to generate real investment returns.

Their answer, backed by almost four decades of data spanning multiple countries, is a clear yes.

What Did the Researchers Set Out to Do?

The authors set out to measure something simple: analyst forecast bias. Every quarter, Wall Street analysts publish earnings estimates for public companies. Those estimates are later compared against actual reported earnings. The gap between the two—the forecast error—signals how well analysts are processing information.

The challenge is that previous attempts to measure this bias have been methodologically flawed. Prior studies typically defined bias as the difference between an analyst’s forecast and a statistical model of “rational” earnings expectations. This sounds sensible, but it creates a contamination problem: the resulting bias measure gets tangled up with analysts’ genuine private information advantages and with random forecast noise. The result is a muddy signal that doesn’t cleanly capture belief distortion.

Gao, Ma, and Yuan take a different approach. Instead of asking “how does the analyst’s forecast compare to a rational benchmark?” they ask: “Can we predict the analyst’s forecast error using only publicly available information?” If yes, then analysts are making systematic, avoidable mistakes—and those mistakes are the true measure of expectation bias.

The Method: Machine Learning Meets Behavioral Finance

To implement this idea, the researchers trained an XGBoost machine learning model (an open-source, highly efficient, and scalable machine learning library that implements gradient boosted decision trees) to predict analyst forecast errors. The model was fed 70 firm characteristics—such as past returns, earnings history, analyst revision history, firm size, profitability ratios, and more—all drawn from publicly available sources.

The model was trained on a rolling basis, using only data that would have been available before each prediction, ensuring no accidental use of future information. The result is what they call the Expectation Bias (EB) measure: the predictable component of analyst forecast error, estimated purely from public information.

A high EB means the model predicts the analyst is being too optimistic—overestimating future earnings. A low (negative) EB means the analyst is too pessimistic relative to what the data would suggest.

Crucially, this framework has an elegant extension: once trained, the model can be applied to firms with no analyst coverage at all. This capability becomes central to their most important conclusions.

Their full sample covers US stocks and spans the period from January 1986 through December 2022, comprising 2,116,852 firm-month observations across 19,915 unique firms.

Key Findings

1. Expectation Bias Strongly Predicts Stock Returns

The headline result is striking. Sorting stocks into ten groups by their EB score and constructing a long-short portfolio—buying the lowest-bias stocks and shorting the highest-bias ones—generated a return of -0.92% per month (roughly -11% annualized) for the high-bias (0.08%) minus low-bias (1.00%) spread. The t-statistic of -4.13 makes this statistically robust.

After accounting for standard risk factors (the Fama-French three- and five-factor models), the strategy continued to generate significant alpha—returns that cannot be explained by known risk exposures. The Fama-French five-factor alpha was 0.88% per month, with a t-statistic of 4.03.

Importantly, this predictability survived the post-2000 era, a period where many well-known anomalies have weakened or disappeared as markets have become more efficient. The EB anomaly persisted with a five-factor alpha of 0.70% per month after 2000, suggesting it reflects something fundamental rather than a historical artifact.

In contrast, the older “conditional bias” approach used by prior research—which conflates genuine analyst information with the bias signal—failed to predict returns at all, with a monthly spread of just -0.15% and a t-statistic of -0.78.

2. The Bias Comes from Underreaction, Not Just Optimism

A natural explanation for analyst forecast errors is simple optimism—analysts are paid by investment banks and have incentives to be upbeat. But the data tells a more nuanced story.

When the researchers looked at when high expectation bias appears, they found it is strongly concentrated in stocks that have recently received bad news: negative past returns; downward analyst forecast revisions; and negative earnings announcement reactions. High EB stocks had an average past twelve-month return of -3.95%, versus +31.40% for low-EB stocks—a massive gap.

This is the sign of underreaction, not unconditional optimism. After bad news, analysts fail to revise their forecasts downward enough. After good news, analysts fail to revise upward enough. The bias is not a constant tilt—it’s a systematic failure to update beliefs fully when new information arrives.

Event study evidence supports this. Stocks where analysts underreacted to bad news (high EB) subsequently drifted downward over the following months as prices gradually caught up to reality. Stocks where analysts underreacted to good news (low EB) drifted upward. This post-earnings announcement drift is exactly what you’d expect if prices were slow to fully incorporate fundamental information.

3. Overconfidence and Sticky Beliefs Drive the Effect

The researchers dug into why analysts underreact, examining several behavioral explanations.

The most powerful was investor overconfidence, proxied by share turnover. High-turnover stocks—where investors trade frequently, suggesting excessive confidence in their private views—showed dramatically stronger expectation bias effects. The difference in monthly alpha between high- and low-turnover stocks was -1.46% (t-statistic of -5.38). The interpretation is that in overconfident environments, investors place too much weight on their prior beliefs and too little on new information, causing sluggish belief updating.

The second key driver was belief stickiness—the tendency of analysts to anchor on past forecasts rather than fully incorporating new data. Measured using a regression approach from prior academic literature, stocks with highly sticky analyst beliefs showed a monthly alpha of -0.98% for the EB strategy, compared to an economically negligible -0.07% for stocks with low stickiness. That difference was itself statistically significant.

Interestingly, several other explanations failed to pan out. The EB anomaly showed no meaningful relationship with limited investor attention (proxied by media coverage), gradual information diffusion (proxied by analyst coverage), or the disposition effect (loss aversion leads investors to sell assets that have increased in value too early, while holding onto assets that have decreased in value for too long). The effect was also largely insensitive to stock illiquidity, suggesting cognitive mechanisms rather than trading frictions are to blame.

Information uncertainty did amplify the effect—younger firms, firms with higher return volatility and greater forecast dispersion, and financially distressed firms all showed stronger EB anomalies. This is consistent with the idea that when information is genuinely harder to interpret, cognitive biases have more room to distort beliefs.

4. Expectation Bias Explains Short-Term Momentum

One of the paper’s important contributions is its resolution of a puzzle in the momentum literature.

The standard story of stock market momentum is well-known: stocks that have performed well over the past 6-12 months tend to keep outperforming over the following months. But at shorter horizons — looking back just one month—the conventional wisdom is reversal, not continuation. Recent winners tend to underperform and recent losers tend to bounce back, likely due to market microstructure effects and temporary price pressure.

A 2021 paper by Medhat and Schmeling (read my review here) documented an exception to this: among high-turnover stocks, one-month returns actually continue rather than reverse. This short-term momentum effect was puzzling precisely because it appeared in large, liquid stocks—the stocks where markets are supposed to be most efficient.

Gao, Ma, and Yuan argued this short-term momentum is just expectation bias wearing a different mask. Recent winners are stocks where investors underreacted to good news (low EB); their prices haven’t fully caught up yet, so they keep drifting upward. Recent losers are stocks where investors underreacted to bad news (high EB); their prices haven’t fully adjusted downward, so they keep drifting lower.

The test is compelling. A winner-minus-loser momentum strategy among high-turnover stocks earns approximately 1.01% per month and is highly statistically significant. After controlling for expectation bias—adding the EB factor to the regression—the momentum spread falls to 0.27% with a t-statistic of just 1.13. The anomaly effectively disappears.

Double-sort tests confirm the story: once you hold expectation bias constant, past one-month return has no additional predictive power for future returns as the “short-term momentum effect is largely subsumed once these belief distortions are accounted for.”

5. The Bias Is Universal, Not an Analyst Industry Artifact

Perhaps the most theoretically important finding is what happens when the model is applied outside the analyst-covered universe.

If expectation bias were simply a product of biased analyst reports—a supply-side distortion created by conflicted sell-side researchers—it should only matter in stocks that analysts cover. But the researchers applied their U.S.-trained model to firms with no analyst coverage, using industry averages to impute analyst-related features.

The result: the EB anomaly was actually stronger in uncovered firms, with a monthly return spread of -1.25% (versus -0.92% for covered stocks) and a five-factor alpha of 0.88% per month. Since there are no analyst reports to distort investor beliefs in uncovered firms, the bias must reflect something more fundamental: a general human tendency to underreact to fundamental information.

The researchers further validated this by applying the U.S.-trained model to international equity markets. Across virtually all developed markets tested—including the UK, Germany, France, Japan, and Australia—the EB strategy generated negative CAPM alphas, with many reaching statistical significance. A model trained entirely on U.S. data, applied without any re-calibration to markets in Europe, Asia, and beyond, continued to work. This is strong evidence of a universal cognitive constraint rather than a quirk of U.S. market structure.

Their findings led the authors to conclude: “We find that this measure robustly predicts the cross-section of stock returns and is not explained by existing return predictors.”

What This Means for the Efficient Market Hypothesis

These findings put meaningful pressure on strong-form and semi-strong-form versions of the EMH.

The EMH’s semi-strong form holds that prices immediately incorporate all publicly available information. If that were true, you could not use publicly observable characteristics—past returns, financial ratios, analyst forecast data—to systematically predict which stocks will outperform. Yet the EB measure does exactly this, and it does so robustly, across subperiods, across countries, and even among firms with no analyst coverage.

The authors’ evidence suggests the source of this inefficiency is not a lack of information, but a systematic failure in how that information is processed. Markets underreact to fundamental news because investors and analysts are cognitively constrained: they are overconfident in prior beliefs, slow to update those beliefs when evidence contradicts them, and especially prone to these errors when information is uncertain or complex.

Several details are worth noting for a balanced view:

The anomaly does face real transaction costs. The strategy has high portfolio turnover—nearly 97% per month for analyst-covered stocks—and after realistic trading costs, the net alpha shrinks to around 0.49% per month. This is still positive and statistically significant (t-stat = 2.30), but it is smaller. Importantly, the Fama-French 5-factor alpha calculated on net returns remained economically significant statistically significant at 0.46% per month (t-stat = 2.27). Whether the net alpha is large enough to justify implementation after all real-world frictions (bid-ask spreads, market impact, management fees) depends on the scale and sophistication of the investor.

The anomaly also involves real risk. Like momentum strategies, it can suffer sharp drawdowns, notably around market rebounds from crisis periods. A volatility-managed version of the strategy performed considerably better on a risk-adjusted basis, suggesting the raw anomaly is partly compensation for crash risk rather than pure mispricing.

Finally, the finding that expectation bias is distinct from—rather than subsumed by—the 70 other known anomalies examined is important. It is not simply a repackaging of existing signals. But like all anomalies, it may face gradual attenuation as it becomes more widely known and arbitraged away.

Key Takeaways