A History of Thought in Asset Pricing

Asset pricing has one question, and it has changed shape roughly once a generation. The question is always some version of “what determines the price of a claim to uncertain future cash flows?” But what the field takes that question to be about, what would count as an answer, and what object an answer must produce have all moved. This essay traces those movements. It is a history of a question, not a catalog of results, and it treats individual papers as evidence for claims about how the question shifted rather than as entries in a bibliography.

The through-line is a single object that took the field eighty years to name and thirty more to measure. By 1980 the theory had converged on the stochastic discount factor: a positive random variable whose covariance with an asset’s payoff determines that asset’s price. Everything before 1980 can be read as the slow assembly of that object; everything after can be read as the struggle to say what it is made of and where to find it. In 1992 the field split over that struggle into distinct research programs, and the most interesting recent development is that the programs are converging again, around a shared methodological move: replace the inferred latent object with a measured one.

Prologue: Prices Before Theory (1900–1951)

Before 1952 there was no field of asset pricing, and so there was no shared question. What existed instead were two practical puzzles that different communities attacked with whatever tools they had. Practitioners and practitioner-adjacent economists asked what a security is worth and how its price relates to that worth. Mathematicians and statisticians asked whether price movements can be forecast at all. The two groups developed largely in isolation, and neither asked the question that would organize everything after 1952: what return should an asset earn in equilibrium, given its risk? There is no portfolio, no risk-return trade-off, and no equilibrium anywhere in the founding pre-theory texts. That absence is what makes this period a prologue rather than a first chapter.

The honest chronology of the period is that it barely has one. Three parallel strands developed in near-total mutual ignorance and were braided together only in retrospect. Bachelier stands outside the conversation entirely. His 1900 thesis, defended under Poincaré, derived the Gaussian law of price changes with dispersion growing as the square root of time from the single postulate that the speculator’s mathematical expectation is zero, named volatility (the “coefficient d’instabilité ou de nervosité”), priced options, and solved first-passage problems by a reflection argument (Bachelier, 1900). He framed all of this as a contribution to probability theory, with the Paris Bourse as a convenient domain of application, and closed the thesis with the claim that the market “without knowing it, obeys a law that dominates it, the law of probability” (Bachelier, 1900). Economists then ignored the work for half a century. It earned only “mention honorable,” and re-entered the field when Savage circulated postcards about it around 1954. Bachelier is chronologically first and dialogically absent: nothing else in the prologue responds to him because nothing else knew he existed.

The statistical strand did form a genuine, if thin, conversation. Cowles, presenting to a joint Econometric Society and ASA meeting three years after the 1929 crash, audited forty-five professional forecasting agencies against explicit chance benchmarks, including twenty-four artificial forecasters generated by drawing shuffled cards, and found that no group beat chance. The sixteen financial services he tracked averaged 1.43 percent per year below the market (Cowles, 1933). Working, a year later and citing Slutsky and King, supplied the complementary instrument: a published artificial series cumulating random draws, against which the trends and apparent cycles that norm-and-cycle analysts prized were shown to be entirely without forecasting significance, and against which the chart formations technicians prized could be checked, if they occurred there just as often they would carry no forecasting value either (Working, 1934). Together they produced a negative fact and a method. Neither offered a theory of why prices should be unforecastable. Cowles was explicit that his null of “pure chance” was a statistical benchmark, not an equilibrium prediction, and Working’s real interest lay in the places where wheat prices departed from randomness (Cowles, 1933; Working, 1934).

The valuation strand is the period’s one clear case of direct succession. Graham and Dodd, writing in 1934 in the immediate shadow of a collapse that had discredited both speculation and time-honored investment rules, codified intrinsic value as a disciplined craft: value justified by the facts, held as a range rather than a point, protected by a margin of safety, and deliberately without a valuation formula, since they treated any single capitalization of earning power as an unsafe premise (Graham and Dodd, 1934). Williams, four years later, supplied exactly the formalism they had refused. His investment value is the discounted infinite stream of dividends, with closed-form solutions including the growth formula eighteen years before Gordon and Shapiro and the Law of the Conservation of Investment Value twenty years before Modigliani and Miller (Williams, 1938). Williams also picked the period’s one explicit doctrinal fight, arguing against the earnings-multiple orthodoxy that dividends, not earnings, are the object of value, since reinvested earnings either fund later dividends already counted or are money lost (Williams, 1938). The movement within the strand is from craft to algebra. The field later kept Williams’s identity and converted Graham and Dodd’s case-by-case characteristics into mechanical sorting variables, which is precisely the formula-driven selection the book distrusted.

Keynes cuts across both strands without engaging either. Chapter 12 of the General Theory was a load-bearing wall of a macroeconomic argument: Keynes needed volatile investment demand, and the stock market’s conventional valuations supplied the mechanism (Keynes, 1936). Read against the other five works, though, it functions as the period’s internal critique. Against the valuation strand it argues that the knowledge required to compute intrinsic value does not exist, so valuation rests on a convention that the present will persist. Against the statistical strand it explains, before the fact, why professional activity might add nothing: professionals are paid to anticipate what average opinion expects average opinion to be, not to forecast prospective yields (Keynes, 1936). Cowles’s finding that forecasters underperform chance is, I would argue, exactly what one would expect of an industry playing Keynes’s beauty contest, though the two works never cite each other and I know of no record that either author drew the connection.

The period thus posed a question it could not answer, and its deepest fault line was inherited rather than resolved. Graham, Dodd, and Williams all presupposed that a value distinct from price exists and can be approximated from facts; Keynes denied the premise, holding that the basis for such calculation “amounts to little and sometimes to nothing” (Keynes, 1936; Graham and Dodd, 1934). The later field split this inheritance cleanly: the efficiency tradition claimed Cowles’s and Working’s negative results, while behavioral finance claimed Keynes’s Chapter 12 and Graham and Dodd’s voting-machine passage as its pre-history. What none of these authors could say is what return an asset should earn. Bachelier assumed zero drift, Williams gestured at a personal discount rate plus an unspecified premium, Graham and Dodd substituted the margin of safety for a price of risk. The prologue supplied the numerator of asset pricing and left the denominator, the risk premium, as its structural absence and the next era’s founding problem.

Era I: The Rational Revolution (1952–1972)

The era did not begin with a question; it converged on one. At its outset there were three disconnected puzzles in three literatures. Markowitz, in 1952, asked a decision-theoretic question: what rule should an investor follow in choosing a portfolio, given beliefs (Markowitz, 1952)? Tobin, six years later, asked a monetary question: what individual behavior microfounds the Keynesian demand for cash (Tobin, 1958)? Modigliani and Miller asked a corporate question: what is the cost of capital to a firm whose returns are uncertain (Modigliani and Miller, 1958)? None of these was posed as a question about the prices of risky assets in equilibrium. By the mid-1960s the field had learned to state the question it had been circling. Sharpe posed the first half explicitly: the field had normative models of individual choice but “no positive microeconomic theory” of how the price of risk emerges (Sharpe, 1964, pp. 425-426). Samuelson posed the second half in its sharpest form: is the observed unpredictability of speculative prices an empirical accident, a refutable hypothesis, or an immutable deduction (Samuelson, 1965)? Fama, the same year, ran the empirical version of the contest, chartists against the random walk (Fama, 1965). By 1970 Fama had fused the two into one agenda. Prices “fully reflect” available information, but that claim has empirical content only inside a model of equilibrium expected returns, and he named the “development and testing of models of market equilibrium under uncertainty” the most pressing task ahead (Fama, 1970, p. 416). The rational revolution was the two-decade process by which “how should I invest?” and “why can’t I predict prices?” became one question: what is the equilibrium price of risk?

The era runs on two strands that develop separately and are spliced at the end. The first runs from portfolio choice to equilibrium. Markowitz is the origin point, but a deliberately modest one: a single investor, no market, no prices, aimed at demolishing the “maximize discounted expected returns” rule inherited from Williams, which can never rationalize observed diversification (Markowitz, 1952, pp. 77-79). His replacement, the mean-variance efficient frontier with covariance rather than own-variance as the operative notion of risk, supplied the raw material for everything that followed. Tobin took the same apparatus for a different purpose, rescuing Keynesian liquidity preference, but delivered the load-bearing technical bridge: the separation theorem, under which the composition of the risky bundle is independent of risk tolerance, so every investor behaves as if there were one composite risky asset (Tobin, 1958). Sharpe then asked the question neither predecessor had. If all investors are Markowitz-Tobin optimizers with homogeneous expectations and a common riskless rate, what must prices be? His answer, the capital market line and the result that only systematic responsiveness, not total risk, is priced, converted a normative toolkit into a positive theory (Sharpe, 1964). Lintner, independently and nearly simultaneously, derived the same equilibrium in dollar-covariance form and pushed it into capital budgeting (Lintner, 1965). Mossin closed the strand by disciplining it: he complained that Sharpe’s diagrammatic equilibrium conditions were “somewhat indefinite” and re-derived the result as a genuine exchange equilibrium, proving rather than assuming that every investor holds the same fraction of every risky asset (Mossin, 1966, pp. 769, 775).

That the central theory of the era was a multiple discovery is worth pausing on, because the literature has never fully settled the attribution and there is no reason to pretend it has. Lintner’s opening footnote records that Sharpe’s paper appeared while his own was at the printer, “parallels the first half,” and neglects capital budgeting (Lintner, 1965). Lintner believed his equilibrium genuinely differed from Sharpe’s; the dispute was dissolved only when Fama showed in 1968 that the two models are identical once notation is aligned. Mossin’s stated purpose was to make precise what Sharpe had left indefinite (Mossin, 1966). Treynor’s unpublished note, conceded in Sharpe’s own footnote 7, is a fourth claimant. The CAPM was thus born as a three- or four-way multiple discovery accompanied by live disagreement about whether it was one theory at all. Modigliani and Miller sit beside this strand rather than inside it. Written before any equilibrium theory of risky returns existed, their paper had to invent the “risk class” as a substitute, and its arbitrage proof that two packages of the same cash-flow stream must sell for the same price introduced a second pricing logic, replication rather than optimization, whose full consequences lay outside the era (Modigliani and Miller, 1958).

The second strand runs from random walks to efficiency, and it ran backwards. Fama himself described the pre-1965 serial-correlation results as “a large body of empirical results in search of a rigorous theory” (Fama, 1970, p. 389). Samuelson supplied the theory: under the axiom that competitive bidding sets prices at conditional expectations, properly anticipated prices are a martingale, a deduction by the law of iterated expectations rather than an empirical claim, and he warned explicitly that the axioms yield no statement about whether actual markets work well (Samuelson, 1965). Fama supplied the evidence at scale, with thirty Dow stocks, serial correlations near zero, and filter rules dead after commissions (Fama, 1965). Fama, Fisher, Jensen, and Roll answered the objection both had left open, that independence tests only infer efficiency: their stock-split study invented event time, market-model residuals, and the cumulative-average-residual plot in one stroke, and showed that prices impound the information in a split by the end of the split month (Fama et al., 1969). Fama then codified the strand, organizing it under the weak, semi-strong, and strong taxonomy he credited to Harry Roberts, and delivering the verdict that contradictory evidence was sparse (Fama, 1970). The efficient markets hypothesis was not proposed and then tested; it was reverse-engineered from tests already run.

The splice is Fama-MacBeth. Fama’s 1970 survey demanded equilibrium models of expected returns, and the CAPM strand had built exactly that; the 1973 paper joins them into a single test of the two-parameter model and market efficiency at once. Its two-pass procedure asks whether the cross-section of NYSE returns is linear in beta, whether beta is a sufficient measure of risk, and whether the monthly deviations are fair games (Fama and MacBeth, 1973). Its verdict reads as the era’s consolidated self-assessment: linearity and beta-sufficiency are not rejected, the coefficients on beta-squared and residual risk carry t-statistics near zero even in the full sample, the monthly slopes are serially uncorrelated, and only the Sharpe-Lintner restriction that the zero-beta return equal the riskless rate is rejected, in favor of Black’s variant. The rational edifice, tested against itself, passes.

One number in that paper deserves the attention the era did not give it. The zero-beta return exceeds the T-bill rate by generally more than 0.4 percent per month, with a full-sample t-statistic of 2.55 (Fama and MacBeth, 1973). At the time this was absorbed as support for Black’s zero-beta model. Read later, it is the first sample-wide documentation of a security market line too flat to match Sharpe and Lintner, the ancestor of the betting-against-beta literature and the first quiet crack in the edifice the era thought it had certified.

Two things the era left unresolved would define everything after it. The first is the joint-hypothesis problem, which Fama flagged as a methodological aside on a single page: stating efficiency in terms of expected returns means every test is conditional on an equilibrium model, so a rejection can never be assigned to the pricing model or to efficiency alone (Fama, 1970). The second is that the era’s durable legacy turned out to be a research technology, not a validated theory. Fama, Fisher, Jensen, and Roll’s residual-averaging machinery became the event study; Fama-MacBeth’s two-pass regression with time-series standard errors became the standard estimator of empirical asset pricing. It was this very machinery that Fama and French would use in 1992 to declare beta dead. The era built tools sturdier than the theories they were built to confirm, and handed the field a testing apparatus it did not yet know how to interpret.

Era II: Arbitrage and State Prices (1973–1980)

The CAPM era had answered “what is the price of risk?” with a single equilibrium trade-off between mean and variance. The next era opened with a narrower-sounding question that turned out to be deeper: how much of an asset’s price is pinned down without knowing anyone’s preferences, expectations, or the market equilibrium at all? Black and Scholes posed it concretely for options. Can the market’s own no-arbitrage condition determine the price completely, where every earlier warrant formula had smuggled in an arbitrary discount rate or expected return (Black and Scholes, 1973)? By the end of the decade the question had been generalized and inverted: what is the minimal structure any coherent price system must have, and what economic primitive stands behind it? The answer, a positive linear pricing rule, equivalently state prices, equivalently a probability measure under which prices are martingales, equivalently the marginal utility of consumption, is the closest thing asset pricing has to a unification. It was assembled piece by piece between 1973 and 1979 by two research lines that did not at first know they were converging.

No contemporary described it this way. Black and Scholes thought they were completing the warrant-pricing literature and building a theory of corporate liabilities; Merton thought he was repairing the CAPM’s fit; Ross thought he was simplifying capital budgeting; Lucas thought he was adjudicating the martingale definition of efficiency; Breeden thought he was making Merton testable (Black and Scholes, 1973; Merton, 1973; Ross, 1978; Lucas, 1978; Breeden, 1979). The unification is visible only in retrospect, which is exactly what makes this era the pivot of the whole history.

The founding pair arrived in 1973. Black and Scholes derived the option pricing equation twice, once from a continuously rebalanced riskless hedge and once from the CAPM, and presented the two routes as co-equal (Black and Scholes, 1973). Merton, publishing the same spring, immediately audited which assumptions did the work. He re-derived the equation from a zero-net-investment hedge with no CAPM at all, arguing that the CAPM assumption “may be a red herring” since the formula depends only on the interest rate and the total variance of the stock (Merton, 1973). He also built the floor under the whole enterprise, a set of distribution-free dominance bounds (put-call parity, no early exercise, convexity) that any rational pricing theory must satisfy. In a separate paper the same year Merton ran the equilibrium line forward: in continuous time with a shifting investment opportunity set, portfolio demands acquire hedging components and the CAPM’s security market line becomes a multi-beta plane, in which expected returns compensate covariance with state variables rather than with the market alone (Merton, 1973). This was not only a pricing result but a decision-theoretic one: long-horizon investors hedge changes in the opportunity set rather than merely hold the static tangency portfolio, a problem later turned into a full strategic-asset-allocation program by Campbell and Viceira (Merton, 1973; Campbell and Viceira, 2002). This was the first genuine dispute of the era over foundations, and the field sided with Merton so completely that the CAPM route to Black-Scholes is now a curiosity. Replication beats equilibrium wherever replication is available, and that division of labor still holds.

The generalizations came in 1976, from two directions at once. Ross took the arbitrage logic out of the option box and into the cross-section: with a factor structure on returns, approximate linear factor pricing follows with no normality, no quadratic utility, and, pointedly, no special role for the market portfolio (Ross, 1976). Ross was careful about what does the work. His Section I counterexample shows that the pure diversification-and-no-arbitrage argument can fail on its own, undone by relative risk aversion that grows without bound as the number of assets rises; the pricing theorem holds only once he adds an agent with uniformly bounded relative risk aversion, together with homogeneous expectations, nonnegative aggregate demands, and bounded expected returns. Even then the result bounds only the sum of squared pricing deviations across the cross-section, leaving individual assets free to misprice. Rubinstein, coming from capital-budgeting theory, proved from the single-price law of markets together with nonsatiation that a single positive random variable prices every security, identified it with the marginal utility of aggregate consumption under aggregation, and then showed that joint lognormality of the stock and that variable reproduces the Black-Scholes formula with discrete trading and no hedge whatsoever (Rubinstein, 1976). Rubinstein is a hinge document of the era, an early explicit statement that the arbitrage strand and the preference strand price the same claims through the same object, anticipating Ross’s 1978 operator and Harrison-Kreps. His answer to whether continuous hedging is the essence of Black-Scholes was no: constant proportional risk aversion, he showed, “substituted for continuous trading opportunities to achieve the same end.”

In 1978 the object was named twice. Ross extracted the theorem from his factor work: no arbitrage implies a positive linear valuation operator, that is, state prices, unique exactly when markets are complete, with Modigliani-Miller and a preference-free version of martingale efficiency falling out as corollaries (Ross, 1978). Lucas, coming from macroeconomics rather than finance, closed the general-equilibrium loop: in an exchange economy the price function is the fixed point of the stochastic Euler equation, tying the pricing object directly to consumption and rational expectations (Lucas, 1978). Then in 1979 the convergence was made exact. Breeden collapsed Merton’s intractable multi-beta plane into a single beta on aggregate consumption growth, since consumption, not wealth, is the sufficient statistic for marginal utility, fusing the two lines at no loss of generality (Breeden, 1979). Harrison and Kreps supplied the definitive mathematics for the arbitrage line: a price system is viable if and only if an equivalent martingale measure exists, a claim is priced by arbitrage if and only if its expectation is common across all such measures, and the measure is unique if and only if markets are complete (Harrison and Kreps, 1979). The Radon-Nikodym derivative of that measure is Rubinstein’s pricing variable, Ross’s operator, and Lucas’s marginal utility. By 1980 the field possessed, without yet using the phrase, the stochastic discount factor.

That three unconnected motivations, capital budgeting, macroeconomics, and portfolio theory, produced the same object is the strongest contemporaneous evidence that the object is fundamental. The arbitrage line is best read as a progressive stripping-away of equilibrium: Black and Scholes needed an equilibrium model badly enough to offer the CAPM as a co-equal derivation; Merton showed the hedge alone suffices; Ross showed even the hedge is a special case of a positive linear valuation operator implied by no arbitrage; Harrison and Kreps proved the exact equivalence, after which the equilibrium scaffolding they themselves retained was stripped by successors (Black and Scholes, 1973; Merton, 1973; Ross, 1978; Harrison and Kreps, 1979). Each paper subtracts an assumption from its predecessor and shows the price survives.

The era also quietly dismantled the martingale definition of efficiency it had inherited. Lucas showed that with curved utility, equilibrium prices are generally not martingales; only the marginal-utility-weighted return series is a martingale-difference, so a failure of the martingale property cannot by itself indict rationality, a point Lucas himself said precisely replicates LeRoy’s 1973 result (Lucas, 1978). Ross’s efficient-market theorem restated efficiency without risk neutrality (Ross, 1978). The joint-hypothesis problem that the next era would confront empirically was thus established theoretically here. So was the era’s greatest unfinished business, which its own endpoints were engineered to foreground. Ross left the factors of the APT entirely unidentified, offering only a passing “GNP factor” example, and was candid that his theorem bounds only the sum of squared pricing deviations, so individual assets can misprice (Ross, 1976). Breeden’s closing pitch for consumption data reads as an implicit answer to Roll, whom he cites: consumption is measured badly, but the true market portfolio, including human capital, real estate, and durables, is not measured at all (Breeden, 1979). The theory looked complete. Whether it was complete is a question worth taking a position on, and I will take one: the theory was formally complete and empirically empty. It had a pricing object and no way to observe it. The next era is the discovery of exactly how empty the box was.

A parallel fixed-income tradition was moving toward the same object in its own language. Vasicek and Cox, Ingersoll, and Ross made the term structure a dynamic equilibrium problem, while the affine tradition that followed treated bond prices as restrictions on the time-series behavior of the pricing kernel (Vasicek, 1977; Cox et al., 1985). The bond literature developed partly apart from equity pricing, but its question was the same: what stochastic object prices payoffs at every horizon, and how does that object move with real rates, inflation, and investment opportunities?

Era III: The Empirical Crisis (1977–1991)

The previous era had left a working consensus: the CAPM described the cross-section of expected returns, the efficient markets hypothesis described their time series, and the standard tests had certified both. The question the crisis era inherited was therefore not “is the theory true?” but “how do we test it properly?”, and its defining discovery was that the field did not know how. Between 1977 and 1991 the question mutated three times. First: what is the theory’s rejectable empirical content, and did the celebrated tests have any? Second: given data the consensus cannot absorb, is the failure in the pricing model, in efficiency, or in the test itself? Third: once the failures are quantified, what must a successor theory explain, a set of cross-sectional characteristic premia or a large, persistent, time-varying discount rate? By 1991 the field had no agreed answer, only a much sharper statement of the question. That unresolved state is what made the coming split possible.

The critique and the first crack arrived together in 1977. Roll’s result was that the two-parameter model has exactly one testable implication, the mean-variance efficiency of the market portfolio, which cannot be tested without observing the true market portfolio (Roll, 1977, pp. 130-131). This reframed every prior confirmation as a tautology given an efficient proxy and every rejection as a joint-hypothesis ambiguity (Roll, 1977, pp. 144-145). In the same year Basu delivered the first well-executed anomaly using precisely the orthodox toolkit Roll was undermining: Jensen alphas under the Sharpe-Lintner and zero-beta CAPM show low price-earnings stocks earning two to four and a half percent per year more than their betas warrant (Basu, 1977). Basu himself stated the finding as a violation of a joint hypothesis and flagged that the ratio might proxy for an omitted risk variable, the escape route the entire subsequent debate would run through.

The crack widened on two fronts in 1981. Banz generalized Basu’s move from a valuation ratio to a firm characteristic, showing that market value predicts returns beyond beta, concentrated in the smallest firms, and, in direct response to Roll, tried three market proxies while refusing to say whether size itself matters or merely proxies for unknown true factors, in which case, in his words, size would be “complete nonsense” as a primitive (Banz, 1981). Simultaneously Shiller attacked the consensus from the aggregate time series rather than the cross-section. Under constant discount rates, price volatility should be bounded by dividend volatility; he found the bound violated five- to thirteen-fold, with the price series roughly five times more volatile than the ex post rational price it was supposed to forecast (Shiller, 1981). Where Basu and Banz questioned the CAPM leg of the joint hypothesis, Shiller questioned the efficiency leg, and he identified, then dismissed as “academic,” the very escape route that would define the rest of the era: hugely volatile expected real rates, swinging from roughly negative four to positive fourteen percent (Shiller, 1981).

The consumption-based theory was then put on trial and lost twice. It was the natural successor to the CAPM that Roll had cornered, since it replaced the unobservable market portfolio with observable consumption. Hansen and Singleton made it testable without solving for equilibrium, estimating the Euler equation directly by GMM on orthogonality conditions, and their illustration rejected constant-relative-risk-aversion preferences on multiple-return systems even while producing plausible point estimates (Hansen and Singleton, 1982). Mehra and Prescott then made the same failure quantitative and vivid where the test statistic had been abstract. Calibrated to a century of consumption data, their model’s maximum equity premium is 0.35 percent against an observed 6.18 percent (Mehra and Prescott, 1985). Mehra and Prescott rejected a model they had expected to pass and say so explicitly, and their own diagnosis pointed less at the premium than at the puzzlingly low risk-free rate (Mehra and Prescott, 1985). Hansen and Singleton, by contrast, framed their rejection as a byproduct of an econometrics methods illustration and reported it flatly, without alarm about what it implied for asset pricing. The pairing set the template for what followed: models would henceforth be judged by whether they escaped that figure, not by whether they passed a test statistic.

De Bondt and Thaler, meanwhile, opened a genuinely new front by inverting the era’s epistemology. Rather than documenting a pattern and asking which leg of the joint hypothesis failed, they derived a prediction from psychology, the Kahneman-Tversky overreaction hypothesis, and confirmed it: prior losers beat prior winners by 24.6 percent over three years despite lower betas (De Bondt and Thaler, 1985). They were explicit that rationalizing a known anomaly after the fact would be unfalsifiable curve-fitting, and that predicting a new pattern was the point (De Bondt and Thaler, 1985). This is the moment psychology switched from residual explanation to hypothesis source, and it chartered the behavioral program that formalized the move after the split.

A parallel literature deserves separate treatment here, because it shares the era’s preoccupation with information but asks a question the rest of the crisis never posed. Everyone from Fama onward had asked whether prices reflect information; almost no one asked how it gets in. The efficient-markets tradition treated price formation as instantaneous and costless, a black box converting information into price the moment it arrived. Market microstructure opened the box. Glosten and Milgrom built a sequential-trade model in which a risk-neutral, competitive market maker earning zero expected profit posts quotes to a stream of anonymous traders, some of them better informed than the maker, and showed that a bid-ask spread arises from adverse selection alone, with no inventory cost, no processing cost, and no monopoly rent: the maker must lose to the informed and recoup from the uninformed, and the spread is exactly that recoupment (Glosten and Milgrom, 1985). Their transaction prices are a martingale with respect to public information and the maker’s information at once, a mechanism-level derivation of semi-strong efficiency rather than an assumption of it, and one that holds even against a better-informed insider, because each quote already embeds the inference the maker will draw from the very order it attracts (Glosten and Milgrom, 1985). Kyle, published the same year, reached the same economics by a different route: a single strategic informed trader hides inside noise-trader flow, a competitive maker prices only the aggregate order imbalance, and information enters price at a rate set by one depth parameter, the price impact per unit of net order flow (Kyle, 1985). The two are the twin foundations of the adverse-selection view, and together they say what the reduced-form efficiency statement cannot: information does not simply appear in the price, it is paid for, and the payment is a permanent price move in the direction of the trade.

Glosten and Harris made the mechanism estimable, decomposing the NYSE spread into a permanent adverse-selection component that scales with trade size and a transitory component covering inventory and order processing, and finding the information component negligible for small trades but dominant for large ones (Glosten and Harris, 1988). That permanent, size-scaled price impact is an object the rest of this history will rediscover under other names: it is Kyle’s depth parameter read off transaction data, the ancestor of the illiquidity measures of the following decade, and, read forward thirty years, the first structural evidence that the demand curve for an individual security slopes down. I place this literature as a counterpoint rather than a chapter deliberately, because it prices something the rest of the essay does not. Its spread is compensation for adverse selection, not for bearing risk; in its base model every agent is risk-neutral and there is no discount factor at all. It sits orthogonal to the price-of-risk question that organizes this history, which is exactly why the traditions fighting over that question could ignore it for two decades. Its debt comes due only in Era V, when the field, having failed to locate the pricing kernel in preferences, begins measuring price impact directly on balance sheets and in holdings, and finds that microstructure had built a structural theory of the thing all along.

The era’s last movement reorganized the crisis around time-varying expected returns and, in doing so, produced one of the most striking convergences in the field’s history. Fama and French showed that dividend yields forecast returns with R-squared rising from under five percent monthly to over twenty-five percent at horizons of two to four years, and interpreted this inside an efficient-markets framework as slowly mean-reverting expected returns (Fama and French, 1988). Campbell and Shiller supplied the accounting that made the two readings one: their log-linear present-value identity forces variation in the dividend-price ratio to reflect forecasts of dividend growth or of returns, they measured that dividend forecasts have only half to two-thirds the volatility of the actual ratio, and they showed every observable discount-rate proxy fails, with the consumption-based one carrying the wrong sign (Campbell and Shiller, 1988). Fama, defending efficiency, and Shiller, attacking it, had arrived at the same empirical object, predictable long-horizon returns, from opposite directions. Excess volatility and return predictability were revealed to be the same phenomenon, and the fight moved from whether expected returns vary to why.

Three unresolved matters passed to the next era, and they map almost exactly onto the programs that would carry them. The joint-hypothesis stalemate was never broken: every anomaly admits a risk reading and a mispricing reading, and Roll’s critique guaranteed no test could adjudicate. Basu and Banz stated both readings and declined to choose; De Bondt and Thaler chose mispricing and were immediately met with risk-based reinterpretations. The stalemate was productive precisely because it was irresolvable, and it forced the field to divide into camps defined by which leg of the joint hypothesis each held fixed. Second, the search for observable discount rates had failed completely. Campbell and Shiller tried real interest rates, which were too smooth, consumption growth, which had the wrong sign, and volatility, which showed no forecasting relation, and all failed (Campbell and Shiller, 1988). That failure defined the agenda of the macro-finance program: habit, long-run risk, and disaster models are all attempts to build a discount-rate process volatile and countercyclical enough to fill the residual observables could not. Third, and most consequentially, the era transformed the question “is the market efficient?” into “what is the stochastic discount factor?” without anyone deciding to do so. Roll had shown the CAPM question untestable as posed; Hansen and Singleton relocated testing from portfolios to Euler equations; Mehra and Prescott showed the leading Euler-equation model failed by an order of magnitude; and Hansen and Jagannathan generalized that failure into nonparametric bounds on the volatility of the discount factor. By the era’s end every camp, the behavioral one included, stated its claims as claims about discount rates. The crisis began with one theory certified and ended with three questions certified in its place.

Era IV: The Divergence (1992–2007)

The split of 1992 is the hinge of this history, and it is worth saying why I date it there rather than at 1985, when De Bondt and Thaler opened the behavioral front, or at 1977, when Roll and Basu opened the crisis. The earlier dates produced anomalies; they did not produce programs. What Fama and French did in 1992 was force every subsequent researcher to choose an interpretation of a common body of facts, and the choices hardened into research programs with their own methods, benchmarks, journals, and students. Before 1992 the field argued within one framework about whether the framework held. After 1992 it argued across frameworks that no longer shared a null. The divergence did not create the three questions the crisis had certified; it made official that the joint hypothesis had been severed and that each program would hold a different leg fixed. That is a sociological event as much as an intellectual one, and it is the reason the field spent the next fifteen years talking past itself.

IV.A: The Empirical Factor Program

By 1992 the inherited question, does the CAPM describe average returns, had collapsed into a negative answer, and the factor program’s real question took its place: if beta does not price the cross-section, what does, and is whatever does it a risk? The question had a surface layer, descriptive, about which of the proliferating predictors carry independent information, and a deeper layer, interpretive, about whether the winning predictors earn premia because they load on pervasive undiversifiable risk or because the characteristics themselves predict returns. The program’s distinctive move was to let the descriptive layer run ahead of the interpretive one. Factors were constructed, benchmarks installed, and alphas computed years before anyone agreed on what a factor was compensation for. That tolerance for interpretation-free machinery is what made this a program rather than a single debate.

Fama and French (1992) is the opening move and the site of the split. Its explicit target was the anomalies literature of the prior decade: run the candidate variables jointly in Fama-MacBeth regressions and announce that size and book-to-market suffice, while beta, even after a double-sort trick designed to generate beta variation unrelated to size, explains nothing; in the univariate regression on beta alone the average slope is 0.15 percent per month with a t-statistic of 0.46. The paper contains no factors. It offers characteristics in cross-sectional regressions plus a promissory note, and it thought it was interring the CAPM and closing the anomalies literature. Instead it opened the largest empirical literature in asset pricing and forced every program to take a position on the same facts.

Fama and French (1993) is the program’s true founding document, and it responds directly to the interpretive gap. If size and book-to-market predict returns and pricing is rational, then they must proxy for sensitivity to common risk factors; so the paper changes the method, replacing Fama-MacBeth cross-sections with time-series regressions in the Black-Jensen-Scholes tradition, and invents the technology: the size and value factors built from double sorts, the twenty-five test portfolios, and the intercept test as the criterion of success. The move from characteristics on the right-hand side to mimicking-portfolio returns on the right-hand side is the era’s central methodological event. It converted anomaly documentation into model building, and it changed the field’s operative question from “what predicts returns?” to “what is the benchmark for alpha?” Within five years an anomaly had been redefined from a violation of efficiency into a nonzero intercept against the reigning factor model.

Momentum shows how completely the new question won, and it exposes what the program was really willing to do. Jegadeesh and Titman, publishing the same year, came from outside the program entirely. Their paper was a market-efficiency test adjudicating between the De Bondt-Thaler contrarian result and practitioner relative-strength rules, and it proposed no factor, no model, and not even the word momentum (Jegadeesh and Titman, 1993). Its finding, twelve months of return continuation earning roughly one percent per month followed by partial reversal, was framed as evidence of systematically biased expectations. Fama and French conceded in 1996 that continuation was the anomaly their three factors could not absorb, and Carhart completed the domestication a year later. He appended a one-year momentum portfolio to the three factors purely as a performance-attribution device for the mutual-fund persistence debate, explicitly declining to take a stand on whether momentum is a risk (Carhart, 1997). The four-factor benchmark became standard. A variable with no accepted economic interpretation became a permanent fixture of the toolkit, which is the program’s clearest revealed preference for description over explanation, and the template for the factor-zoo logic that would follow.

The program’s internal reckoning came from Daniel and Titman, and it was made possible by the program’s own founding design choice. Because characteristics and factor loadings are nearly collinear in standard sorts, the 1993 time-series tests could not distinguish “value stocks earn more because they load on the value factor” from “value stocks earn more because they are value stocks.” Their triple-sort design broke the collinearity, and the characteristic-balanced portfolios delivered the verdict for characteristics over covariances: a mean return statistically indistinguishable from zero coexisting with a three-factor alpha of 0.354 percent per month (Daniel and Titman, 1997). The exchange that followed, with Davis, Fama, and French answering that the result was a short-sample artifact and Daniel, Titman, and Wei answering back with Japanese data where characteristics won again, ran on as a decade of replication warfare that settled nothing rather than resolving into a verdict (Daniel and Titman, 1997). What the field took from it was not a winner but a methodological lesson: the program’s central tests have low power against characteristics alternatives, so its central evidence was weaker than its adoption suggested.

The program left two things unresolved. The first is the interpretive question it had deferred by construction. Fama and French floated the value factor as a proxy for a priced distress state variable, but conceded that without a theory of the state variables any particular version of the factors is somewhat arbitrary, and the promised link from earnings fundamentals to a priced state variable was never established (Fama and French, 1993). The distress story survived as citation ritual, not as a live research result. The second is the question of whether the factor models were asset-pricing theory or bookkeeping at all, which the program never settled. Carhart’s agnosticism and Fama and French’s own concession that factor choice was motivated by empirical experience meant the program’s flagship objects had officially undetermined interpretations for its entire duration. The program grew, in the end, by systematically re-reading its inputs as contributions to benchmark construction, whatever their authors had intended.

IV.B: The Macro-Finance Program

The factor program took the anomaly catalog as its problem statement. The macro-finance program took a different reading of the 1980s wreckage. On this reading, Mehra-Prescott, Shiller, Hansen-Singleton, and the predictability evidence had not refuted the consumption-based paradigm; they had refuted one narrow parameterization of it. The program’s central question was accordingly unitary and ambitious: can the entire constellation of aggregate puzzles, the equity premium, the low and stable risk-free rate, excess volatility, and long-horizon predictability, be explained jointly, from a single representative-agent equilibrium, with macroeconomic fundamentals as the source of risk? The founding conviction was that these were not separate puzzles but reflections of one object, a large, volatile, countercyclical risk premium, and that a model matching one moment while missing the others had explained nothing.

Cochrane’s textbook consolidated the program’s constitution: every model is a specification of the discount factor in the pricing equation, and the empirical target is a single calibration. A Sharpe ratio near 0.5 against one percent consumption volatility requires a risk aversion coefficient of 50, or 250 once the low consumption-return correlation is taken into account (Cochrane, 2001). The book inverted the field’s history, teaching the consumption-based model first and the CAPM as a special case, and made “what is your discount factor?” the default seminar question (Cochrane, 2001). That calibration functioned as the program’s entrance exam, which is why three mutually incompatible models could all count as members.

The preference technology preceded the program. Epstein and Zin, writing before the split and not as macro-finance at all, saw themselves fixing a specification problem: the expected-utility straitjacket that forces the elasticity of intertemporal substitution to be the reciprocal of risk aversion (Epstein and Zin, 1989). Their recursive utility broke that link and nested the CAPM and the consumption CAPM as special cases. It also carried a feature they flagged uneasily, a preference for early resolution of uncertainty, on which the program’s second generation would run.

The program then produced three canonical mechanisms, each a minimal deviation from Lucas along a different margin. Campbell and Cochrane kept i.i.d. lognormal consumption growth and time-additive power utility but subtracted a slow-moving external habit, making the surplus consumption ratio the single state variable, so that investors fear stocks because stocks do badly in recessions (Campbell and Cochrane, 1999). Bansal and Yaron took the opposite margin: keep moderate risk aversion of 10, adopt the Epstein-Zin recursion with an elasticity of 1.5, and relocate the risk into a small, highly persistent expected-growth component plus stochastic volatility, which raised the maximal Sharpe ratio from 0.27 to 0.73 (Bansal and Yaron, 2004). The mechanism is powered directly by the early-resolution preference Epstein and Zin had derived and worried about. Barro chose the most conservative deviation of all: retain the Lucas tree, time-additive power utility, complete markets, and i.i.d. shocks, and rehabilitate rare disasters by measurement rather than assumption, counting sixty contractions of fifteen percent or more across thirty-five countries in the twentieth century to give a disaster probability of 1.7 percent per year, with the levered premium coming out at 5.4 percent against seven percent observed (Barro, 2006).

These three are one research design executed on three margins, and each purchased its success by making its key state variable hard to observe. In the habit model the sensitivity function governing how the surplus consumption ratio responds to consumption shocks is reverse-engineered, chosen to deliver a constant risk-free rate and a predetermined habit (Campbell and Cochrane, 1999). Long-run risk’s persistent growth component is by construction nearly indistinguishable from i.i.d. growth in any finite sample, which was Bansal and Yaron’s own defense against the charge that no one could see it in the data (Bansal and Yaron, 2004). Disaster risk lives in the probability of events mostly absent from any single country’s history (Barro, 2006). The shared move, inferring macro risk dynamics from the very asset prices they are meant to explain, was later named “dark matter,” but it was already the program’s method rather than an accident of any one paper. The data of the era could not adjudicate between habit and long-run risk, which matched largely the same moments with opposite economics, and the dispute was deferred rather than resolved. Each mechanism, in the end, relocated rather than eliminated the implausibility the calibration demanded: Campbell and Cochrane conceded value-function risk aversion near 80, Bansal and Yaron hid the cost in enormous premia for early resolution of uncertainty, and Barro’s disasters were recast by critics as a peso problem.

A sidebar to this program deserves separate treatment, because it shares the era’s dates and its equilibrium discipline but not its representative agent. Scheinkman and Xiong, reaching back to Harrison and Kreps, answered a question the three consumption-based mechanisms could not pose: why the internet episode combined high prices with frenzied volume. Harrison and Kreps had shown that heterogeneous beliefs plus a short-sale prohibition make the resale option worth paying for, so the stock trades above what even the most optimistic investor would pay to hold it forever; their numerical example has the price exceed every investor’s own valuation (Harrison and Kreps, 1978). Scheinkman and Xiong built the fully dynamic version, microfounding the disagreement in overconfidence, deriving the trading-frenzy prediction in closed form, and adding the signature policy result that a transaction tax of one percent of prices throttles volume but cuts the bubble and its excess-volatility component together by less than twenty percent (Scheinkman and Xiong, 2003). I place this as a sidebar rather than as a fourth consumption-based mechanism deliberately, because it is not one. The paper is disagreement-based theory, closer in spirit to the behavioral program than to the representative-agent machinery around it. In Harrison and Kreps no agent is irrational beyond holding heterogeneous priors and agreeing to disagree, yet there is no representative valuation to aggregate to, since equilibrium prices strictly exceed every investor’s fundamental value (Harrison and Kreps, 1978). Scheinkman and Xiong go one step further and microfound the disagreement in overconfidence imported explicitly from the behavioral literature, so their belief differences are a belief-formation bias rather than mere priors (Scheinkman and Xiong, 2003). Its inclusion in the macro-finance orbit is defensible only if that program is defined by its question, what equilibrium object generates observed price behavior, rather than by its representative-agent technology. Read that way, the disagreement strand is the era’s internal admission that some price behavior, bubbles with turnover, resists any single discount factor, an admission the intermediary and demand-based literatures of the next era would generalize.

The program ended by converting its own central question, performed by the textbook rather than by any single result, from “is the premium explicable?” into “which discount factor is it?” (Cochrane, 2001). That reframing legitimized habit, long-run risk, and disasters as coequal competitors, and simultaneously pushed offstage the demand-based and heterogeneous-agent questions that the disagreement sidebar had kept alive. What it left unresolved was whether the pricing object could be tied to anything observable at all, or whether the program had simply replaced one unobservable, the market portfolio, with another, the latent state variable. The next era’s institutional literature was, in large part, the attempt to answer that by measuring the marginal investor directly.

IV.C: The Behavioral Program

The third branch took the mispricing reading of the anomalies, and its central question crystallized quickly: can systematic mispricing be turned into a disciplined scientific program rather than a list of embarrassments for efficiency? Answering yes required two things at once. It required a supply-side account of why mispricing survives, since the standard rejoinder was that smart money should trade it away, and a demand-side account of which psychological distortions produce the observed pattern. The constraint that Fama made canonical in 1998 gave the demand side its shape: the evidence showed both short-horizon underreaction, in drift and momentum, and long-horizon overreaction, in reversal and value, and any behavioral theory worth taking seriously had to generate both from one mechanism, plus predictions it was not built to fit. The program’s self-understanding was explicitly architectural. It would stand on two legs, limits of arbitrage and investor psychology, and no single paper would contain both.

Shleifer and Vishny built the load-bearing wall first. Before any psychology could matter for prices, the program had to disarm the argument that arbitrage eliminates mispricing, and their move was to attack the arbitrage mechanism itself. Real arbitrage is delegated, specialized, capital-constrained, and disciplined by uninformed outside investors who allocate on past returns, so performance-based arbitrage fails precisely when mispricing is deepest, forcing arbitrageurs to liquidate into widening dislocations (Shleifer and Vishny, 1997). The collapse of Long-Term Capital Management the following year gave the paper an almost experimental confirmation and settled, for the era, that “why doesn’t smart money fix it?” had an answer.

With the survival of mispricing secured, 1998 and 1999 produced three rival accounts of its origin, and the fact that they were rivals rather than a single theory mattered. Barberis, Shleifer, and Vishny came first and defined the template: a representative investor who is a competent Bayesian inside a permanently wrong model, oscillating between a mean-reverting regime and a trending regime, generating underreaction to single news shocks and overreaction to strings of them. They did not model limits to arbitrage; they imported them by citation, making the two-leg architecture explicit (Barberis et al., 1998). Daniel, Hirshleifer, and Subrahmanyam answered with a different psychology and a different fault line: overconfidence about private signals plus biased self-attribution, so that prices overreact to private information and underreact to public information, with momentum as continuing overreaction rather than slow correction (Daniel et al., 1998). Hong and Stein closed the trio by relocating the action from biases inside one head to interaction between boundedly rational types, newswatchers who never condition on prices and momentum traders who condition only on past price changes, so that gradual information diffusion generates underreaction and momentum trading necessarily converts it into overreaction (Hong and Stein, 1999). Of the three, only Hong and Stein delivered a confirmed out-of-sample prediction quickly, that momentum concentrates in small, low-coverage stocks (Hong and Stein, 1999).

That the three models fit the same facts with contradictory mechanisms was exactly what Fama weaponized. Three theories fitting one pattern was, he argued, evidence of reverse-engineering, and the degrees-of-freedom critique stuck to the program permanently. The models made promises the psychology could not keep: Barberis, Shleifer, and Vishny staked their out-of-sample credibility on finding an a priori strength-and-weight classification of news for out-of-sample tests, and admitted their model predicts overreaction to earnings surprises for exactly the extreme growth stocks where the data show underreaction (Barberis et al., 1998). The mechanism dispute, whether momentum is underreaction slowly corrected or continuing overreaction, was never settled by the profession.

Barber and Odean supplied the missing empirical actor. The theory had assumed noise traders; nobody had measured them. Using account-level data on seventy-eight thousand households at one discount broker, of which the sixty-six thousand holding common stocks form the performance sample, they showed that households turn over seventy-five percent of their portfolios per year, that gross returns are flat across turnover quintiles while the highest-turnover quintile’s net returns fall far below the lowest’s, and that rational explanations, liquidity, rebalancing, and taxes, cannot account for the volume (Barber and Odean, 2000). The paper framed itself as investor welfare, not asset pricing, but the program later read it as the microfoundation, the correlated, attention-driven demand the models required. The direction of intellectual traffic ran, unusually for this literature, from theory to measurement.

By 2001 the program had bifurcated, and the second branch fought the macro-finance program rather than efficiency. Barberis, Huang, and Santos embedded prospect theory, loss aversion over financial-wealth fluctuations with the loss-aversion coefficient varying with prior gains and losses, into a Lucas exchange economy, generating a five percent equity premium, excess volatility, and low return-consumption correlation with modest loss aversion (Barberis et al., 2001). Its investor is fully rational and dynamically consistent, with nonstandard utility rather than distorted beliefs, and its declared rival is Campbell-Cochrane habit formation, with a stated testable difference: in habit models risk aversion moves with consumption, in this one with past market outcomes (Barberis et al., 2001). The early behavioral papers argued against efficient markets; the late ones competed with the macro-finance program on its own quantitative terrain.

What the program left unresolved is unusually easy to state, because its most durable exports turned out to be its least psychological parts. The mechanism dispute among the three momentum models was never resolved. The prospect-theory account survived as proof of concept rather than as a workhorse aggregate model, since its quantitative success depended on prior-outcome dynamics that later evidence rendered fragile (Barberis et al., 2001). And the program’s deepest legacy ran sideways: Shleifer and Vishny were retrospectively read as a paper about the industrial organization of asset management, constrained intermediaries as price-setters, seeding intermediary asset pricing and fire sales; Hong and Stein’s surviving contribution was gradual information diffusion, a friction rather than a bias; and even Barber and Odean’s households were later recast as one demand class among many (Shleifer and Vishny, 1997; Hong and Stein, 1999; Barber and Odean, 2000). The behavioral program’s historical function was partly to smuggle frictions and institutions into asset pricing under a psychological flag, a reading no participant would have endorsed in 1999, and the direct bridge to the era that followed.

Era V: The Programs Mature (2008–present)

Each of the three programs entered its mature phase after the financial crisis, and each did so by confronting the same limitation from a different angle: the objects it reasoned about, factors, discount factors, sentiment, were inferred from returns rather than observed. The maturation of all three consists in replacing an inferred latent object with a measured one. That common move is what the epilogue will argue is closing the divergence, but it is easiest to see program by program first.

V.A: The Factor Zoo and Machine Learning

By the late 2000s the factor program had succeeded so thoroughly that its success became the problem. Hundreds of published characteristics claimed to predict the cross-section, each certified by the same convention, a t-statistic above two, that had certified the first, and Cochrane’s 2011 address named the predicament the factor zoo. The era’s central question was reflexive in a way no prior era’s had been. It was not “what explains expected returns?” but “which of our own past findings should we believe, and what is the right statistical technology for a cross-section whose candidate predictors now outnumber any small model’s capacity to test them one at a time?” The question ran in three movements: consolidation, audit, and re-representation.

Consolidation came first. The initial response to the zoo was to shrink it back into a small model, now with theory attached. The q-factor model revived a production-based route that had run alongside consumption-based asset pricing: infer discount rates from firms’ investment first-order conditions rather than from household marginal utility (Cochrane, 1991). Hou, Xue, and Zhang built a four-factor model from that production-side logic, market, size, investment, and profitability, and confronted it with nearly eighty anomalies, finding that half died under proper breakpoints and value weights and that most survivors’ alphas vanished, momentum included (Hou et al., 2015). Fama and French answered within months with a five-factor model adding profitability and investment, motivated by the valuation identity rather than by q-theory, and discovered along the way that their own value factor had become redundant, an ironic epitaph delivered by its creators (Fama and French, 2015). The two papers together launched the “factor wars,” which consumed enormous energy and settled little: both models share nearly identical investment and profitability factors, though they differ elsewhere, the q-theory model carrying four factors and a monthly-sorted profitability factor with no value factor while the valuation-identity model keeps five and retains HML; both are rejected by their own intercept tests, and the model-comparison literature that followed turned benchmark choice into a permanent horse race.

The audit came next, and it turned the program’s statistical machinery on the program itself. Harvey, Liu, and Zhu hand-collected the full census of factor discoveries, three hundred sixteen of them from three hundred thirteen articles, published and unpublished, treated four decades of factor production as one giant multiple-testing problem, and concluded that the defensible hurdle was a t-statistic above three, not two, under which between eighty and one hundred fifty-eight of the two hundred ninety-six published significant factors are false discoveries (Harvey et al., 2016). Their data are not stock returns but the t-statistics of three hundred thirteen articles, which is the moment the program turned reflexive. Hou, Xue, and Zhang supplied the empirical companion, replicating four hundred fifty-two anomalies under a common design and finding that sixty-five percent fail even single tests once microcaps are controlled and eighty-two percent fail the multiple-testing hurdle, with the trading-frictions category collapsing almost entirely (Hou et al., 2020). Discovery had given way to meta-science. The audit was itself contested: Chen and Zimmermann and Jensen, Kelly, and Pedersen reinterpreted the low replication rate as re-weighting rather than replication failure, arguing that published t-statistics reproduce almost exactly under original procedures and that Bayesian pooling of the correlated factor universe implies most factors replicate (Hou et al., 2020; Harvey et al., 2016). The settled residue is narrower than either side’s headline: anomalies concentrate in microcaps, investable magnitudes are far smaller than advertised, and the robust core is small.

Re-representation abandoned the premise that consolidators and auditors had shared, that the goal is a short list of true named factors. Kozak, Nagel, and Santosh showed that the mere absence of near-arbitrage forces the discount factor to load on the first few principal components of returns, regardless of whether expected returns come from risk or from sentiment (Kozak et al., 2018). This result is the era’s quiet bombshell, because it deflates both the inference that “factor models work, so pricing is rational” and the Daniel-Titman characteristics-versus-covariances test at once: a sentiment economy with constrained arbitrageurs produces exactly the same low-dimensional factor pricing as a risk economy, so zero alphas on latent factors demonstrate comovement, not risk compensation. Their sequel drew the constructive conclusion, estimating the discount factor on all candidate factors at once with an economically motivated shrinkage prior and finding that sparsity in named characteristics fails, since dozens of anomalies carry marginal pricing information, while sparsity holds only in unnamed principal-component space (Kozak et al., 2020). The multi-decade quest for a sparse characteristics-based model was, on their verdict, futile because the anomalies are simply not redundant enough.

The machine-learning phase generalized the move. Kelly, Pruitt, and Su fused the latent-factor and characteristics traditions, using characteristics as instruments for the dynamic loadings on latent factors, with the factors themselves estimated jointly by alternating least squares, and showing that with two or more factors anomaly alphas vanish (Kelly et al., 2019). Gu, Kelly, and Xiu recast the cross-section as a high-dimensional nonlinear prediction problem and ran a horse race of thirteen methods on nearly a thousand predictors, crowning shallow neural networks, while insisting that the forecasts are “only measurements” and that mechanism comes later; the field then received the exercise as a paradigm statement that recast decades of one-characteristic-at-a-time testing as a dominated special case (Gu et al., 2020). Kelly, Malamud, and Zhou carried the complexity argument over to the time series of aggregate market returns, in the Goyal-Welch tradition rather than the cross-section, and inverted the field’s oldest methodological commitment: under shrinkage, out-of-sample Sharpe ratios increase in model complexity even when parameters exceed observations, so that “Occam’s razor may instead be Occam’s blunder” (Kelly et al., 2024). The era that opened by asking which of three hundred sixteen factors to keep closed by arguing the analyst should use the largest model she can compute. Whether that virtue of complexity is real or an artifact of the setup remains actively contested, and it is the era’s live frontier rather than its settled conclusion (Kelly et al., 2024).

The episode says something uncomfortable about the sociology of the field, and it is worth saying plainly. The factor zoo was not produced by fraud or incompetence. It was produced by a publication technology, the single t-statistic above two, applied honestly and independently thousands of times against a finite dataset that cannot generate fresh observations the way particle physics can. The program’s own success at certifying findings became indistinguishable from data mining, and the field needed a full decade of reflexive audit to notice. That is a lesson about incentives and about the difference between a discipline that can run new experiments and one that must re-analyze a fixed history. What the era left unresolved is the deepest question of all: it showed that the rational-versus-behavioral question inherited from the split cannot be answered by cross-sectional factor evidence, because risk and sentiment are observationally equivalent at the level of the low-dimensional factor structure (Kozak et al., 2018). Kozak, Nagel, and Santosh’s own prescription, that only structural models of beliefs and preferences can discriminate, went largely unheeded within the program; the field took the statistical half and left the structural half. That road not taken is exactly the road the other two programs of this era were building.

V.B: Intermediaries, Institutions, and Demand Systems

For three decades after Lucas, the pricing kernel belonged to a representative household; intermediaries were a veil, and the stock market and mortgage-backed securities shared a single consumption-based discount factor. The post-crisis intermediary literature did not invent constrained asset pricing; it made an older incomplete-markets problem measurable. Since Hart and Geanakoplos and Polemarchakis, the field had known that when risks cannot be fully shared and borrowing is constrained, equilibrium allocations need not be efficient and prices can depend on wealth distributions and collateral values (Hart, 1975; Geanakoplos and Polemarchakis, 1986; Geanakoplos, 2010). The crisis made that assumption look not merely inaccurate but disabling. Risk premia in intermediated markets spiked nonlinearly, arbitrage relations that no household ever traded, such as the CDS-bond basis, went deeply negative and stayed dislocated, and the recovery tracked the recapitalization of dealer banks rather than any change in household consumption (Duffie, 2010; He and Krishnamurthy, 2013). This wing of the macro-finance program asked what happens to asset pricing when the marginal investor is not a household but an institution, a levered dealer or a mandate-constrained fund, whose capital, balance sheet, and demand curve are observable state variables. Having tried habits, long-run risks, and disasters as amendments to the household kernel, the program now replaced the household itself.

Duffie’s presidential address was the charter. It presented no structural theory, offering its own general-equilibrium inattentive-investor model as deliberately illustrative rather than structural; it assembled the empirical signatures of price impact and slow reversal, index deletions, catastrophe premium cycles, Treasury auction dips, the crisis-era arbitrage breakdowns, and declared that asset-pricing fundamentals must include state variables tracking the availability of capital, above all intermediary capitalization (Duffie, 2010). This was also the moment the limits-to-arbitrage literature was absorbed from its behavioral-finance annex into the mainstream: slow-moving capital was now everyone’s problem, not the anomalies literature’s. Its other ancestry was informational and microstructural. Grossman and Stiglitz had shown that perfectly informative prices are impossible when information is costly, and Kyle and Glosten had turned private information into price impact and a bid-ask spread (Grossman and Stiglitz, 1980; Kyle, 1985; Glosten and Milgrom, 1985). Duffie’s slow-moving-capital program absorbed that lesson into asset pricing proper: liquidity was no longer a trading friction around prices but a state variable for prices. He and Krishnamurthy and Brunnermeier and Sannikov answered the call with general-equilibrium theory, breaking with the log-linearized financial-friction macro that could not speak to an event happening far from any steady state. He and Krishnamurthy built a crisis model around an equity capital constraint on intermediaries, generating the signature asymmetry, premia flat when the constraint is slack and exploding as specialist capital falls below a threshold, and matching crisis half-lives of eight to ten months (He and Krishnamurthy, 2013). Brunnermeier and Sannikov solved the global nonlinear dynamics and found the volatility paradox, that low fundamental risk breeds high leverage and hence systemic fragility (Brunnermeier and Sannikov, 2014). Neither paper aimed at everyday cross-sectional pricing: Brunnermeier and Sannikov presented itself as a crisis-motivated contribution to macroeconomics rather than asset pricing, and He and Krishnamurthy, though titled intermediary asset pricing, framed itself as a crisis model and a policy laboratory with the pricing-kernel message treated as a byproduct. The field retroactively made them the founding documents of intermediary asset pricing because both delivered the same reduced-form message, that the pricing kernel lives on intermediary balance sheets.

The empirical wing converted that message into cross-sectional tests, deliberately mimicking the two-pass methodology of the factor literature so the new kernel could compete on the incumbent’s home turf. Adrian, Etula, and Muir showed that a single broker-dealer leverage factor prices forty-one stock and bond portfolios with a seventy-seven percent cross-sectional fit, matching a five-factor benchmark whose own fit was slightly higher at eighty-one percent while winning on pricing errors, a twelve-basis-point intercept against sixty-six, and on parsimony, one factor against five (Adrian et al., 2014). He, Kelly, and Manela built a primary-dealer capital-ratio factor and took it to the markets where the theory should bite hardest, options, credit, currencies, sovereigns, and commodities, finding a pooled price of capital risk of about nine percent per quarter that could not be rejected as common across the classes, even though the point estimates ranged widely, from around seven to eleven percent in most classes up to nineteen percent in currencies and twenty-two percent in options (He et al., 2017). The two papers staged the era’s defining internal dispute knowingly: Adrian, Etula, and Muir’s leverage factor is procyclical with a positive price of risk, which they read as evidence against the wealth-based theories, while He, Kelly, and Manela found the opposite cyclicality, and in head-to-head tests the leverage factor’s risk price turned insignificant and flipped sign in several markets (Adrian et al., 2014; He et al., 2017). The dispute was productive precisely because it turned “whose constraint binds, and when” from a modeling choice into an empirical question. It also inherited the vulnerabilities of the borrowed methodology: Adrian, Etula, and Muir’s headline fit proved sensitive to data vintage and sample, making the paper a cautionary exhibit in the same power debates that dogged the factor program (Adrian et al., 2014).

Koijen and Yogo then widened the question from intermediaries to institutions in general, and from the discount factor to the demand curve. Their complaint was methodological: every equilibrium model is demand plus market clearing, yet the field estimated models from returns alone and discarded holdings, which are direct observations of demand. Using a toolkit the field would later recognize as empirical industrial organization, characteristics-based demand, instruments built from investment mandates, and GMM estimation, though the authors themselves positioned the paper within asset pricing theory rather than the differentiated-products tradition, they estimated demand investor by investor from institutional filings and found it strikingly inelastic: latent demand explains eighty-one percent of the cross-sectional variance of stock returns and supply-side fundamentals only twelve percent, so that a ten percent demand shock moves the median stock’s price by roughly twenty-six percent (Koijen and Yogo, 2019). Gabaix and Koijen completed the arc at the aggregate level with the inelastic markets hypothesis: moving one dollar from bonds into aggregate equities raises total market value by roughly five dollars, implying a macro demand elasticity two orders of magnitude below standard models, which predict a move of essentially zero (Gabaix and Koijen, 2021). Their own survey of economists had found a median predicted multiplier of exactly zero against their estimate of five, and they framed the result as Shiller-style excess-volatility revisionism made structural, replacing the dark matter of latent discount-rate shocks with measurable flows from concrete investors (Gabaix and Koijen, 2021).

Two points about the demand-system turn need care, because they are easy to overstate. First, it is a synthesis, not a pure descendant of the macro-finance program. Koijen and Yogo derive characteristics-based demand from Markowitz-style optimal portfolio choice with heterogeneous beliefs and short-sale constraints, and are explicit that they estimate demand plus market clearing, which is what every equilibrium model contains (Koijen and Yogo, 2019). The estimation itself, I would add, does not pin down whether the inelasticity it measures originates in preferences, mandates, or slow-moving capital, a point later critics of the demand-system literature pressed (Gabaix and Koijen, 2021). It is best read as equilibrium machinery, the estimation of a structural demand system with instruments, put in the service of a question the limits-to-arbitrage literature had posed. Which leads to the second point: the low estimated elasticities are, in substance, a vindication of limits to arbitrage. Shleifer and Vishny’s argument was that arbitrage is delegated, specialized, and capital-constrained, disciplined by outside investors who chase past returns, so that performance-based arbitrage fails and mispricing can persist; read retrospectively, that is a world in which demand curves for individual securities slope down (Shleifer and Vishny, 1997). An aggregate multiplier of five and a latent-demand share of eighty-one percent are what a world of scarce, slow-moving arbitrage capital looks like when you measure it directly rather than infer it. The demand-system program did not overturn the frictionless benchmark by discovering a new anomaly; it estimated, with equilibrium machinery and holdings data, the price impact that the behavioral program’s limits-to-arbitrage wall had predicted must exist. The two literatures, born on opposite sides of the divergence, describe the same friction.

What this wing left unresolved is whether measured demand is structural, which is to say whether the estimated elasticities would survive a policy change. Koijen and Yogo flag the Lucas critique themselves and proceed under the null that demand is structural; Gabaix and Koijen concede that their hypothesis “remains just that: a hypothesis” (Koijen and Yogo, 2019; Gabaix and Koijen, 2021). Whether measured inelasticity reflects preferences, mandates, or simply Duffie’s slow-moving capital in new clothing is the live question, and it is the same question the beliefs program of this era was asking from the other side.

V.C: Subjective Beliefs and Survey Expectations

By the late 2000s the field had lived for two decades with the settlement that followed the divergence: price ratios predict returns, and the macro-finance program read that predictability as rationally time-varying required returns while the behavioral program read it as sentiment. Both, however, argued from the same evidence, prices and realized returns, and inferred expectations rather than observing them. The central question of this era was almost embarrassingly simple: what do investors actually expect, and do those measured expectations behave like the expected returns that theory attributes to them? The polemical form was sharper. Rational representative-agent models make a joint prediction, that the expected return inferred from the dividend-price ratio and the expected return in investors’ heads are the same object, and testing that equivalence became an overlooked but decisive falsification test (Greenwood and Shleifer, 2014). Answering it required first winning a prior fight, establishing that survey data were admissible evidence at all, against a profession-wide dismissal that Cochrane had voiced by recommending the survey evidence be discarded as uncomfortable (Greenwood and Shleifer, 2014; Barberis et al., 2015).

The survey turn was part of a broader household-finance turn that Campbell had named explicitly: ordinary households were no longer background suppliers of aggregate consumption, but heterogeneous decision-makers whose participation, diversification, borrowing, and belief formation could be measured directly (Campbell, 2006). Limited participation and investment mistakes mattered not only because they might move prices, but because they changed the allocation of risk across households (Vissing-Jørgensen, 2002; Calvet et al., 2007). The microfoundational opening move for the beliefs branch came from Malmendier and Nagel, before the aggregate survey program existed as such. Using household survey data matched to lifetime return histories, they showed that personally experienced returns, weighted with a roughly linearly declining function of how long ago they occurred, predict risk-taking, and that stock experiences move stock holdings while bond experiences move bond holdings, evidence that kills the wealth-effects alternative and, the paper suggested tentatively, favors the beliefs channel over endogenous preferences, though it remained deliberately agnostic between the two until the 2016 follow-up (Malmendier and Nagel, 2011). Two features proved generative: beliefs are formed from a subset of history, so they never converge across cohorts, and the weighting function estimated purely from microdata aggregated into something tracking the Shiller price-earnings ratio with a correlation near 0.6. That aggregation exercise, modest at the time, was the bridge the rest of the era walked across.

Greenwood and Shleifer delivered the founding fact. They assembled six survey series, showed they were mutually consistent and consistent with mutual-fund flows, thereby answering the admissibility objection, and then showed that survey expectations are extrapolative, explained by past returns and the price level, and negatively correlated with model-based expected-return proxies, minus 0.33 against the log dividend-price ratio and minus 0.48 against the Campbell-Cochrane surplus consumption ratio, forecasting realized returns with the wrong sign (Greenwood and Shleifer, 2014). Their calibrated economy showed that the dividend-price ratio predicts returns even with constant risk aversion, so the predictability evidence the macro-finance program had claimed as its own could no longer be read automatically as evidence for time-varying rational risk premia (Greenwood and Shleifer, 2014). Theory then caught up. The extrapolative capital asset pricing model of Barberis, Greenwood, Jin, and Shleifer estimated its key extrapolation parameter from the very survey series it sought to explain and compared itself head-to-head with habit and long-run risk on a common scorecard whose expectations column was empty for every model but its own (Barberis et al., 2015). Its sequel extended extrapolation to bubbles, adding a “wavering” term to generate the volume that pure common extrapolation cannot, and confirmed the model’s distinctive volume-and-past-return prediction in four historical episodes (Barberis et al., 2018). Bordalo, Gennaioli, La Porta, and Shleifer then moved the program from “beliefs extrapolate” to “beliefs overreact for a specifiable psychological reason,” estimating a diagnostic-expectations filter from expectations and fundamentals alone and asking it to deliver a return spread it was never fit to, which it roughly did (Bordalo et al., 2019). Their explicit ambition was to answer the degrees-of-freedom critique that had dogged the behavioral program: relaxing rational expectations, they argued, adds discipline, because the model must now fit measured beliefs too.

Martin’s contribution sits inside this chronology as the counterparty, and it is important to frame it correctly, because it is not a behavioral paper. Martin derived a lower bound on the equity premium read off option prices in real time, a no-arbitrage result in the direct lineage of the Hansen-Jagannathan bounds, and the observable mirror image of them: where Hansen and Jagannathan bounded the volatility of the unobservable discount factor, Martin bounds the conditional expected return using traded option prices, with no estimation and no behavioral assumption (Martin, 2017). It is rational no-arbitrage theory. Its role in this era’s story is that it gave the beliefs literature exactly what it had lacked, a theory-disciplined, estimation-free statement of what conditional expected returns should look like, and Martin himself documented that all four of the main survey measures correlate negatively with his bound (Martin, 2017). A paper from the rational tradition thereby sharpened the behavioral anomaly. That a no-arbitrage bound and a survey both say expected returns rise when subjective expectations fall is the cleanest evidence in the whole literature that measured beliefs are not the expected returns of rational theory.

Three papers then complicated the program from within, and each is a stress test rather than a refutation. Giglio, Maggiori, Stroebel, and Utkus linked a quantitative survey to anonymized administrative records of the Vanguard portfolio holdings of wealthy US retail investors, averaging around half a million dollars, and found that beliefs matter but pass through weakly, a one-percentage-point rise in expected returns moving equity shares only about seven-tenths of a point, an order of magnitude below the frictionless benchmark, and that the dominant feature of belief data is persistent cross-sectional disagreement, with individual fixed effects absorbing nearly half the variation and time effects only a twentieth (Giglio et al., 2021). This is a direct corrective to representative-agent belief calibrations that assume strong common time variation. De la O and Myers re-ran the Campbell-Shiller decomposition under the subjective measure and inverted the post-1988 consensus: subjective cash-flow expectations account for essentially all price-ratio variation and subjective return expectations for almost none, shifting the program’s center of gravity from return extrapolation to distorted fundamentals expectations (De la O and Myers, 2021). Nagel and Xu closed the arc Malmendier and Nagel had opened: a representative agent with fading memory, using the same gain parameter estimated from the microdata and not fit to prices, generates a high, countercyclical objective premium alongside acyclical subjective expectations, doing with one psychologically grounded friction what habit and long-run risk had done with preference machinery, while matching the survey moments those models contradict (Nagel and Xu, 2022).

What this era achieved, and what it left open, both concern measurement. Its decisive weapon against the macro-finance program was not a new anomaly in returns but a sign flip in measured beliefs: across otherwise different papers, where rational models require expectations to rise, measured expectations fall or stay flat (Greenwood and Shleifer, 2014; Martin, 2017; Giglio et al., 2021; Nagel and Xu, 2022). And it answered the degrees-of-freedom critique by inverting it, making measured beliefs a source of discipline rather than an escape, so that a behavioral model must now match belief moments as well as return moments (Barberis et al., 2015; Bordalo et al., 2019; De la O and Myers, 2021). What it left unresolved is whose beliefs. Retail, analyst, and executive expectations may not be those of the wealth-weighted marginal investor, and this is now where the action is (Greenwood and Shleifer, 2014; De la O and Myers, 2021). It also left open whether the distortion lives in beliefs about cash flows or about returns, with the cash-flow reading currently ascendant (De la O and Myers, 2021); the two need not be exclusive, though that is my reading rather than a claim the paper makes.

Epilogue: Convergence

The divergence of 1992 severed a joint hypothesis into three programs, and for fifteen years those programs talked past one another because they held different legs of it fixed. The factor program held efficiency fixed and asked which risks price the cross-section. The macro-finance program held rationality fixed and asked which discount factor a representative agent must carry. The behavioral program held neither fixed and asked which psychological distortion, disciplined by limits to arbitrage, produces the anomalies. The claim of this epilogue is that the divergence is now closing, and that it is closing because all three programs, in their mature phase, made the same move: they replaced an inferred latent object with a measured one.

State the three descendants side by side and the shared move is unmistakable. The factor program had inferred the pricing kernel from returns; its machine-learning phase estimates it as a high-dimensional statistical object on all candidate signals at once, and its own theory shows that the object it recovers is silent on the risk-versus-sentiment question it was built to answer (Kozak et al., 2018; Gu et al., 2020). The macro-finance program had inferred the discount factor from consumption and preferences, the dark matter of latent state variables; its post-crisis phase measures it directly, on the balance sheets of intermediaries and in the holdings of institutions (Duffie, 2010; Koijen and Yogo, 2019). The behavioral program had inferred sentiment from prices; its mature phase measures it, in surveys and administrative records of what investors actually expect and hold (Greenwood and Shleifer, 2014; Giglio et al., 2021). Holdings, balance sheets, stated beliefs: three programs, one methodological turn, from inference to measurement. That is why the divergence is closing. The programs no longer disagree about how to infer an unobservable, because they have stopped inferring it.

The convergence is visible in the cross-links, not merely in the parallel. Giglio, Maggiori, Stroebel, and Utkus is not cleanly a beliefs paper or a demand paper; it is both, and it is the bridge between them (Giglio et al., 2021). Its finding that even wealthy, engaged investors move their equity share only about seven-tenths of a point per point of expected return is simultaneously the beliefs program’s reality check and founding micro-evidence for inelastic-markets thinking, cited alongside Gabaix and Koijen (Giglio et al., 2021; Gabaix and Koijen, 2021). The two literatures are two halves of one post-representative-agent account: beliefs supply the shifts in demand, and inelasticity supplies the price impact that turns those shifts into returns. Weak accommodation of belief-driven demand by rational traders is exactly what lets measured beliefs move prices as much as the aggregate evidence says they do. Read together, the subjective-beliefs and demand-system programs are not rivals but complements, and the complementarity is quantitative, not rhetorical.

A second cross-link runs the other way, from the statistical to the structural. Kozak, Nagel, and Santosh is the bridge between the factor program and the beliefs program, because it proves that the cross-sectional evidence the factor program produces cannot by itself distinguish risk from sentiment, and then names the only thing that can: structural models of beliefs and preferences (Kozak et al., 2018). The factor program’s own frontier result thus hands the discriminating question to the beliefs program, which is measuring exactly the beliefs that would resolve it. The machine-learning phase declared, in effect, that its methods are only measurements and that mechanism comes later (Gu et al., 2020); the beliefs and demand programs are where “later” is happening.

The third cross-link is the one I have argued for at length, and it deserves restating as the epilogue’s structural claim: inelastic demand is a limits-to-arbitrage result estimated with equilibrium machinery. Shleifer and Vishny argued, on the behavioral side of the divergence, that performance-based arbitrage is scarce and specialized so that mispricing persists, which I read as a world in which demand curves slope down (Shleifer and Vishny, 1997). That downward slope had an older structural home still: in the microstructure models of the crisis era, adverse selection made every trade move the price permanently against the trader, so an individual security’s demand curve sloped down one transaction at a time, and the price impact Kyle and Glosten formalized in 1985 is the same object the demand-system program would later estimate from holdings, reached from information economics rather than from institutional mandates (Glosten and Milgrom, 1985; Kyle, 1985). Koijen and Yogo and Gabaix and Koijen, using the equilibrium apparatus of structural demand estimation, measured how far down those curves slope, and the answer, a multiplier of five and a latent-demand share of eighty-one percent, is the friction the behavioral program had posited, now quantified (Koijen and Yogo, 2019; Gabaix and Koijen, 2021). A result born in behavioral finance is confirmed by the estimation technology of equilibrium economics. The two sides of the 1992 divergence meet in a single number.

It would be too tidy to declare the field reunified, and I will not. Real disputes remain live, and they are the honest residue of the convergence rather than its refutation. Whether the virtue of statistical complexity is real or an artifact is unsettled (Kelly et al., 2024). Whether measured demand is structural or would dissolve under a policy change is unsettled (Koijen and Yogo, 2019). Whose beliefs are the marginal investor’s, and whether the distortion lives in cash-flow or return expectations, is unsettled (De la O and Myers, 2021). What has changed is not that these questions are answered but that the three programs now argue about them in a common currency: observable quantities, measured beliefs, and measured balance sheets, tied together by equilibrium market clearing. The stochastic discount factor that the field spent from 1900 to 1980 assembling, and from 1980 to 2007 arguing about how to infer, is now something the field increasingly tries to read off the world rather than back out of prices. That is the sense in which the divergence of 1992 is closing. The question that organized the whole history, what determines the price of a claim to uncertain cash flows, has not changed. What has changed, after more than a century, is that the field has decided to answer it by looking at who holds the claim, what they believe about it, and what it would cost them to change their minds.