[Special] What Survived in Behavioral Finance — The Alpha That Publication Killed Was No Victory for EMH but a Map Drawn by the Limits of Arbitrage

In the companion piece I put behavioral economics on the stand of the replication crisis—the failure of an experiment to reproduce its result when it is run again. The verdict ran like this. Most of the fallen stars—power posing, social priming, ego depletion—were social psychology, while the core of judgment-and-choice theory—prospect theory, anchoring, framing—passed head-on the very adversarial replication that killed them. The audit did not sentence the whole field. It sorted it by lineage and by trait (What Survived in Behavioral Economics — The Stars the Replication Crisis Toppled Were Mostly Social Psychology). But that piece deliberately left the market empty. It touched neither the disposition effect nor the equity premium puzzle. Now I lay the same ruler against the market.

On the surface, the market side glitters. Kahneman (2002), Thaler (2017), and Shiller (2013) took Nobels; hundreds of papers on anomalies—patterns of excess return that ought not to exist if the market were efficient—cracked the efficient-market hypothesis (EMH); and the smart-beta and factor ETFs that turned those findings into product—exchange-listed funds that build an index by weighting on a chosen characteristic such as value or momentum—held US$1.56 trillion worldwide as of February 2024. The headline is one line. "Behavioral finance broke EMH."

But anomalies have a strange habit. From the moment they are published in a journal, they begin to die. McLean and Pontiff tracked 97 predictors and found their return power fell, on average, 58% after publication. Which prompts the question. If publication kills alpha—excess return over the market—is that a victory for EMH, or is it something else?

Finance Runs Two Audits

There is not one ruler that audits behavioral finance but two. One is post-publication decay: how fast an anomaly fades once the world knows about it. The other is replication: whether the anomaly reappears at all in independent data. The two measure different things. The first ruler measures whether arbitrage—buying and selling a dislocated asset to pocket the gap and pull the distortion back—reaches the places where mispricing (a price adrift from fair value) sits. It measures location and cost. The second measures whether the anomaly ever existed at all; that is, it measures reality. Blur the two and you explain nothing.

The scoring does not split into a true-or-fake binary but into three lineages. (A) what survived; (B) what was arbitraged away after publication; (C) what was never there to begin with. And what sorts the three is not one yardstick but two boundary lines. The arbitrage frontier—the line between where arbitrage reaches and where it cannot—divides A from B. The replication frontier—the line between what reappears in an independent sample and what does not—clears away C.

Let me fix behavioral finance's position here. Behavioral finance is not the thing these rulers measure; it is one of the several things they score. Just as the companion piece scored behavioral economics with the ruler of the replication crisis, here the market anomalies are scored by these two rulers.

Table · The three lineages and two frontiers the finance audit sorted. Within lineage A, the one whose bias is confirmed is the disposition effect; investable survivors like momentum and value are observationally unidentified from a risk premium. The equity premium puzzle is an un-arbitrageable extreme—there is no trade structure to arbitrage at all—so it sits outside this table, on its own. Primary sources: McLean & Pontiff 2016 / Hou, Xue & Zhang 2020 / Odean 1998 / Bernard & Thomas 1989. As of 2026-07-02.

What Was Never There — the Factor Zoo Is a Replication Frontier

First, clear away lineage C, what was never there. Harvey, Liu, and Zhu catalogued 316 factors—a factor being a common characteristic claimed to explain stock returns—that the journals had produced. Test that many factors enough times and some will look significant by pure chance, so they demand not the conventional threshold of t=2.0 but t>3.0. Hou, Xue, and Zhang went one step further. When they re-ran 452 anomalies with a rigorous method that suppressed the distortion from microcaps—the smallest stocks—65% failed to clear even the single-test threshold (absolute t≥1.96), and at the multiple-testing threshold (t=2.78), 82% fell away. This swelling of the factor list into the hundreds is the factor zoo.

This is not a story about the arbitrage frontier. It is a story about the replication frontier. These factors were not arbitraged away; they never existed. Here is the trap the companion piece flagged. To read the collapse of the factor zoo as "the collapse of behavioral finance" is the very misreading that took the collapse of power posing for "the collapse of behavioral economics" (What Survived in Behavioral Economics — The Stars the Replication Crisis Toppled Were Mostly Social Psychology). It is the failure of theory-free mining for t-stats, not the failure of behavioral-finance theory. The misattribution ran both ways. In the boom, behavioral finance shared the credit for the factor zoo's glamorous "discoveries"; in the bust, behavioral finance was made to wear the factor zoo's failure to replicate.

So what is lineage C (mining) and what is lineage A (genuine bias)? The divider is not "did it survive after the fact." Sort it that way and you have a circular argument. The divider is the psychological microfoundation that predates the discovery. The disposition effect and underreaction had prospect theory and laboratory studies first, and were confirmed in market data second. A factor scraped from data with no theory has no such prior grounding. Some factors carried a plausible behavioral-finance story and still failed to replicate. Survival is not by itself "behavioral, confirmed," and decay is not by itself "mining, confirmed."

So what percentage is fake? I will not answer that. The debate is live. If Hou, Xue, and Zhang are the pessimists, Chen and Zimmermann are the optimists: of the 161 predictors that were clearly significant in the original papers, 98% replicated. The difference comes from method. Hou, Xue, and Zhang used a uniform standard that suppresses microcaps; Chen and Zimmermann followed each original paper's own specification. The argument does not hang on "what percentage is fake." It hangs on "the two frontiers as a classifying axis."

The Alpha Publication Killed — That Is No Victory for EMH

Now lineage B, the anomalies publication killed. Look again at McLean and Pontiff's numbers. The return power of the 97 predictors falls 26% already out of sample—the period outside the original paper's own window—and falls 58% after publication. Grant that the 26% out-of-sample drop is the share owed to statistical chance; the 32 percentage points that publication adds are different. Investors who moved on what they read made that.

A sharp reader will want to settle it in one sentence here. "Once published, it is arbitraged away, so EMH is right." A reader one step further along will say, "Some were real, some were data mining." Neither is precise. What is decisive is not how much decay happened but where more of it happened.

McLean and Pontiff measured that the post-publication decay is larger in stocks that are easy to arbitrage. Decay was larger in large caps, in actively traded names, and above all in stocks with low idiosyncratic risk—the volatility specific to a single stock, independent of the market. The higher that risk, the higher the holding cost an arbitrageur bears to carry the position. After publication, the volume and turnover of these stocks rose significantly, and short interest—shares sold short, which is borrowing a stock to sell it and later buying it back to repay, a bet on a decline—jumped roughly ninefold.

Table · Where post-publication decay is larger. The dependent variable is the normalized post-publication long-short return (lower means larger decay). Decay was larger in large-cap, high-liquidity, low-idiosyncratic-risk, and dividend-paying stocks—the ones easy to arbitrage—and when the characteristics are raced together, idiosyncratic risk (holding cost) dominated. The direction and significance are robust, but the coefficient magnitudes are measured from the working-paper version (82 characteristics), so they may differ when recomputed on the published version (97 characteristics). Primary source: McLean & Pontiff (2016), J. Finance 71(1), Table 7. As of 2026-07-02.

Why is this not a "general" victory for EMH? If the market were efficient wholesale, decay would be uniform, independent of how hard a stock is to arbitrage. But decay was large only where arbitrage reaches. This is not the proposition "the market is efficient" but "mispricing is corrected only where arbitrage reaches." It is a map drawn by the limits to arbitrage—the fact that arbitrage takes capital and risk and so cannot be done without bound.

This map must not be read as behavioral finance's private property. The limits to arbitrage are a tool both camps use, and the orthodox rational camp predicts the same picture. Grossman and Stiglitz proved that a perfectly efficient market is impossible. If there is no reward for digging up information, no one will dig, so prices leave exactly that much inefficiency wherever arbitrage is expensive ("efficiently inefficient"). Jensen's cost-adjusted efficiency—that prices are efficient only up to the limit of trading and information costs—says the same thing. So what this cross-section of decay proves is the pattern "correction happens where arbitrage reaches," not that the mispricing was a bias to begin with. The same decay pattern shows up whether it came from a behavioral bias or from information-cost friction. The cross-section of decay points to the location of the correction; it cannot point to the source of the mispricing.

If we call publication's killing of an anomaly "death equals proof," fairness requires hanging a falsification condition right beside the slogan. Had the decay been uniform, independent of liquidity and trading cost, that decay would have been not arbitrage's correction but the uniform product of publication bias—a published effect inflated to begin with, shrinking on its own afterward. In that case it would be "death equals data mining," not "death equals proof," and this argument loses. McLean and Pontiff's non-uniform cross-section is exactly what passes that falsifier. The decay was not uniform; it correlated with how hard a stock is to arbitrage.

So "it died because arbitrage reached it" and "it was a bias to begin with" are not the same claim. Run the elimination honestly. Had the return been a risk premium—a legitimate excess return earned for bearing risk—there would be no reason for it to vanish upon publication, so the risk premium drops out. Had it been pure data mining, the decay would have been uniform, so that drops out too. But stop there and say "what remains is bias," and you have skipped a branch. Information-cost friction remains. On that reading it was never a mispricing but a price gap the size of what a rational investor spends to dig up information, and it narrowed as arbitrage costs did. This reading and the "behavioral bias" reading are not separated by the cross-section of post-publication decay alone—because both predict the same picture. So the proof that lineage B's death is the death of bias is not in this decay pattern. That proof is elsewhere: in the individual's account.

The Bias Arbitrage Left Alive — a Map of the Survivors

The map of surviving anomalies is exactly the map of where arbitrage cannot reach. Small caps, thinly traded names, stocks with short-selling constraints, stocks thick with retail investors. These are the places Shleifer and Vishny flagged, where arbitrage is walled off by risk and cost. That is where bias stays.

The point of highest discriminating power on this map is the disposition effect—the tendency to sell a winner quickly and to hold a loser rather than sell it. Odean took apart 10,000 individual accounts. The rate at which investors realized a gaining stock (0.148) was about 1.5 times the rate at which they realized a losing one (0.098). Why is this decisive? Because it reduces neither to a risk factor nor to information-cost friction. Selling the winner and clinging to the loser is not a matter of the cost of digging up information but the psychology of not wanting to lock in a loss, and it is not a return but the selling behavior itself, measured directly in individual accounts. The bias and the friction that the cross-section of decay could not separate in the previous section separate here. It is not, though, pure psychology. In December the tendency reverses, and investors realize the losers more instead (0.128 versus 0.108)—the calculation of locking in a tax loss to cut the tax bill.

The remaining survivors sit on the same map. PEAD—post-earnings-announcement drift—is the phenomenon in which, after an earnings surprise, a stock keeps drifting slowly in the direction of that surprise for weeks, meaning the market underreacts to new information. The return gap between the highest- and lowest-surprise stocks ran to roughly 4% a quarter, about 18% a year. Momentum—the way stocks that rose over the past 3 to 12 months keep rising—was robustly observed at 0.95% a month for the six-month winners minus losers, at t=3.07; and conversely, over spans of several years, a long-term overreaction appeared in which past losers outran past winners.

Table · Lineage A, the survivors where arbitrage cannot reach. Of these, only the disposition effect is a confirmed bias (individual-account behavior, irreducible to either risk or information cost); for momentum and long-term overreaction, risk premium versus bias is still unresolved. All four persist in the limits-to-arbitrage zone—small, illiquid, retail-dominated. Primary sources: Odean 1998 / Bernard & Thomas 1989 / Jegadeesh & Titman 1993 / De Bondt & Thaler 1985. As of 2026-07-02.

The investable survivors—value, momentum, the ones anyone can hold—carry a trap. Two readings predict survival equally well. One is the bias reading: "arbitrage could not reach, so the bias stayed." The other is the risk-premium reading: "this is not a bias but the reward for bearing risk." The fact of survival alone cannot separate the two. The cross-section of decay from the previous section is no discriminator here either—it shows the limits of arbitrage, but not whether what remains is bias or risk premium. So I put no weight of the argument on this unidentified zone. The one place the full weight of "bias" lands is the disposition effect: not a return but an individual's selling behavior, irreducible to either risk or information cost. For the rest of the survivors I say only "something remains where arbitrage cannot reach," and leave open whether that something is bias or risk premium.

Passing Is Not Confirmation

Having drawn a map of survivors does not mean everything on it is safe. Passing is not confirmation. The passage that most demands honesty is the value premium. When Fama and French re-measured their own factor, the margin by which value beat the market shrank from 0.42% a month in the first half (1963–1991) to 0.11% a month in the second (1991–2019)—about 74%. On the point estimate alone, value too has moved toward lineage B. Yet the same paper cannot reject the hypothesis that the premium is the same across the two periods (F=0.41, p=87%). The month-to-month variation is so large that one cannot statistically declare it "gone." So I do not nail down "value equals a safe survivor." The point estimate says decay; the statistics say unresolved.

This is not the only thing unresolved. Whether momentum and value are risk premia or biases is still fought over in the academy. As we saw with the factor zoo, the share owed to data mining is under debate too (Hou, Xue & Zhang versus Chen & Zimmermann). This is why the argument takes no stand on a particular percentage. The spine is not what percentage of what is fake, but which side of the two frontiers a thing sits on.

The Equity Premium Puzzle — Where There Is No Trade to Arbitrage

The extreme of the arbitrage-frontier logic is the equity premium puzzle. Over a sample of nearly ninety years (1889–1978), U.S. stocks earned about 6.18 percentage points a year more than the risk-free asset. The problem is that the premium a standard consumption-based model can explain at a rational level of risk aversion is only about 0.35 percentage points—about one-eighteenth of what is observed. To close the gap with the model, one must assume an unrealistically high risk aversion.

Why has this puzzle stood unchanged for forty years? Because there is no trade to arbitrage. A single stock's mispricing can be pulled back by buying low and selling high, but a market-wide macro puzzle—"stocks as a whole earned far too much relative to bonds"—has no convergence trade that would bet on two prices coming together to capture a riskless arbitrage. With no frontier to reach, there is neither correction nor resolution. But do not misread this "unresolved." It does not mean an uncorrected bias remains; it means there is not yet an agreed-upon explanatory model. The leading behavioral-finance candidate is myopic loss aversion—the account that a tendency to feel losses more, compounded by frequent checking of one's account, makes investors shun stocks excessively. Put in a loss-aversion coefficient and an evaluation cycle of about a year, and it matches the observed premium. But it is only leading, not settled, and non-behavioral rational solutions—rare disasters, long-run risk, habit formation—remain competing candidates.

If the equity premium puzzle is the extreme arbitrage cannot reach, at the opposite end stand the large-cap, liquid anomalies that die the moment they are published. Set against the whole spectrum, EMH and behavioral finance are not rivals. They are the same equation read from opposite ends. The market learns and corrects mispricing up to the limit arbitrage can reach, and beyond that limit the bias remains. That the 2013 Nobel went at once to Fama of the efficient market and to Shiller of excess volatility—who showed that stock prices swing 5 to 13 times more than the later realized dividends would justify—is no mischief of coincidence. But that shared award was a prize for "the empirical analysis of asset prices," not a proof that "both ends are right." It is only the institutional circumstance of this complementarity.

What Comes Next

So when publication makes alpha disappear, who bears the cost? The academy took the reputation from the discovery; the asset managers turned the discovery into smart-beta product and took the fees—US$1.56 trillion worth as of February 2024. Not all of it is empty. Factors with firm prior grounding that still replicate are mixed in there too. The problem is the products built on factors that decayed, or were never there. The individuals and institutions who bought them late bear the excess return that has already vanished. The sell side has settled up; the decay is handled in a footnote; the fees keep being collected. This structure, in which a delegated judgment defers the reckoning, overlaps with the money flows of Bitcoin ETFs: The Outflows Stopped, the Price Won't Follow and the delegation of Passive Crossed Half the Market. Who Took Over the Judgment?.

So what is predictable? The two frontiers move like this going forward. First, the gap in decay rates holds. In easy-to-arbitrage large-cap, liquid stocks, post-publication alpha vanishes fast; in the arbitrage-walled zone of small, illiquid, retail-dominated names, it vanishes significantly slower. Second, the more recently a large-cap, liquid anomaly is published, the faster its post-publication decay, in step with the growth of arbitrage capital. Third, the performance of smart-beta and factor ETFs splits. The returns of products on factors with firm prior grounding (value, momentum, quality) diverge from those on data-mined factors.

This outlook must be able to be wrong in a checkable way. If anomalies in the arbitrage-walled marginal zone vanish as fast as large caps do—that is, if the cross-section of decay rates goes uniform—the signature of the decay cross-section collapses and the limits-to-arbitrage argument misses. Then "death equals data mining," not "death equals proof," was right. Or if the factor zoo keeps expanding even after replication infrastructure—open-source tools for validating factors—has spread widely, the replication-frontier forecast is wrong. Let me note one limit as well. The data that drew this map are mostly the U.S. market. Whether the same frontiers replicate in other countries' markets is itself an open test. I set the horizon at 2030.

Here is one decoder you can use right away. When you meet an anomaly or a factor product, look at whether it sits in large-cap, liquid stocks with little retail participation, or in the zone of small, illiquid, short-constrained, retail-dominated names with a psychological basis that predates the discovery. The former does not last long. The latter stays. And if it does not replicate in independent data, it is lineage C—never there to begin with.

Whether the market is efficient is the wrong question. Ask instead whether this bias is somewhere arbitrage reaches or somewhere it cannot, and whether it replicates at all. And the bill for the alpha that decayed—or was never there—is always left, quietly, with whoever bought late.