TA Review - Multifactor Models

TA Review - Multifactor Models#

TA: Dipesh Ghimire

Autumn 2026. Review session: November 5, 2026.

Linear Factor Decomposition vs Pricing#

Let \(\tilde r_{i,t}\) denote an excess return and \(f_t\) the vector of factor realizations. A linear factor decomposition (LFD) describes realised returns:

\[ \boxed{\tilde r_{i,t} = \alpha_i + \beta_i^\top f_t + \varepsilon_{i,t}, \qquad \mathrm{Cov}(\varepsilon_{i,t}, f_t) = 0} \]

It is a statistical projection. It is used for attribution, hedging, tracking, and covariance estimation. A high time-series \(R^2\) means the factors account for much of the realised variance. It says nothing, by itself, about average returns.

A linear factor pricing model (LFP) restricts expected returns:

\[ \boxed{E[\tilde r_i] = \beta_i^\top \lambda} \]

There is no population pricing error in this expected-return restriction. Realized returns still have residuals. The vector \(\lambda\) contains the factor risk premia.

Tangency Pricing in Sample#

Using the same sample means and covariance matrix, an unconstrained tangency portfolio \(p\) prices every asset in its investment universe. Assume the covariance matrix is nonsingular and the tangency portfolio is well-defined:

\[ \bar r_i^e = \widehat\beta_{i,p}\,\bar r_p^e, \qquad \widehat\beta_{i,p} = \frac{\widehat{\mathrm{Cov}}(r_i^e,r_p^e)}{\widehat{\mathrm{Var}}(r_p^e)}. \]

A cross-sectional regression of sample mean returns on these betas has zero intercept and \(R^2=1\), provided the mean returns vary across assets. Individual time-series \(R^2\) can still be low.

This in-sample identity does not establish out-of-sample pricing. Estimation error in means and covariances can make the estimated tangency portfolio unstable. Factor models use a smaller set of exposures; their out-of-sample pricing remains an empirical question.

Question

Object

Do factors explain realised comovement?

Time-series \(R^2\), residual covariance

Are average pricing errors zero?

Individual and joint alphas

Do exposures explain return differences?

Risk-price estimates, cross-sectional fit

Does one factor add to existing factors?

Spanning alpha, mean–variance improvement

CAPM as the one-factor benchmark#

\[ \boxed{E[\tilde r_i] = \beta_{i,M}\, E[\tilde r_M], \qquad \beta_{i,M} = \frac{\operatorname{Cov}(\tilde r_i, \tilde r_M)}{\operatorname{Var}(\tilde r_M)}} \]

The CAPM posits one priced risk factor: the market. Diversifiable idiosyncratic risk earns no premium. The market factor has clear economic motivation, correlating with aggregate investor wealth (labor, housing, consumption).

Multifactor pricing generalizes this to \(K\) exposures: \(E[\tilde r_i] = \sum_{k=1}^K \beta_{i,k}\lambda_k\).

Empirically, market beta alone fails the cross-section; portfolios sorted on size, book-to-market, profitability, investment, and past returns earn anomalous returns. This motivates additional factors but does not prove every predictor compensates for risk. Testing the CAPM is also subject to Roll’s critique: the true market portfolio is unobservable. Any test is a joint test of the model and the proxy; rejection may reflect a flawed proxy, a flawed model, or both.

Fama–French Three-Factor Model (FF3)#

\[ \boxed{E[\tilde r_i] = \beta_{i,M}\, E[\tilde r_M] + \beta_{i,S}\, E[\tilde r_S] + \beta_{i,H}\, E[\tilde r_H]} \]

  • MKT (\(\tilde r_M\)): Broad equity market excess return.

  • SMB (\(\tilde r_S\)): Small minus big (diversified small-stock minus large-stock return).

  • HML (\(\tilde r_H\)): High minus low (high book-to-market value minus low book-to-market growth return).

Fama and French (1992, 1993) showed that size and value explain average returns where the CAPM fails. Pricing relies on multivariate betas (the marginal effect of each factor holding others fixed), not univariate correlations.

Book-to-market is book equity divided by market equity (\(B/M\)); high \(B/M\) denotes value. Alternative metrics include earnings/price or EBITDA/price (accounting flows) and dividend/price (cash payout), but \(B/M\) is the standard balance-sheet measure.

The \(2\times3\) Construction#

Independent size and book-to-market (\(B/M\)) assignments are intersected:

Low \(B/M\) (Growth)

Middle \(B/M\)

High \(B/M\) (Value)

Small

\(SL\)

\(SM\)

\(SH\)

Big

\(BL\)

\(BM\)

\(BH\)

Standard U.S. FF3 construction uses June sorts with NYSE size-median and \(B/M\) 30th/70th percentile breakpoints. Book equity is lagged, and portfolio returns are value-weighted. The six portfolios generate:

\[\begin{split} \begin{aligned} SMB_t &= \frac{SL_t+SM_t+SH_t}{3} - \frac{BL_t+BM_t+BH_t}{3} \\ HML_t &= \frac{SH_t+BH_t}{2} - \frac{SL_t+BL_t}{2} \end{aligned} \end{split}\]

This design balances size across value categories and value across size groups. The long-short legs offset much of the broad-market exposure, though neither operation guarantees zero market beta or exact factor orthogonality.

Other Models and Economic Interpretation#

Model

Factor Set

Additional Dimension

CAPM

MKT

Market exposure

FF3

MKT, SMB, HML

Size and value

Carhart 4-factor

FF3 + MOM

Momentum

FF5

MKT, SMB, HML, RMW, CMA

Profitability and investment

  • RMW: Robust minus weak operating profitability.

  • CMA: Conservative minus aggressive investment.

Economic interpretation: A factor premium may reflect compensation for adverse state-dependent payoffs, persistent mispricing, or both.

Characteristics versus betas#

Factor models price beta, not characteristics. An asset’s premium derives from \(\beta_i\), not its \(B/M\) ratio, size, or past return. A high-\(B/M\) stock with \(\beta_H \approx 0\) earns no HML premium, whereas a low-\(B/M\) stock that covaries strongly with HML does.

This distinction separates the Fama–French and Buffett–Graham approaches to value. Buffett and Graham buy individual stocks with strong fundamentals relative to price (a characteristic rule). The Fama–French factor is constructed by sorting on characteristics, while its pricing model relates expected returns to covariance exposure to that factor. They may often select the same stocks, but the underlying logic is fundamentally different.

Two Schools of Thought: Risk Premia vs. Behavioral Biases#

Factor premia are explained by two paradigms. The risk-based view argues that premia compensate investors for exposure to systematic, non-diversifiable macroeconomic risks (bad states). The behavioral view (e.g., Richard Thaler) attributes premia to systematic cognitive biases among investors and limits to arbitrage, allowing mispricing to persist.

Value (High Book-to-Market)#

  • Risk-based: High \(B/M\) firms are often financially distressed or cyclical. Higher expected returns compensate investors for bearing elevated discount rates and vulnerability to economic downturns.

  • Behavioral: Investors overreact to past poor performance, overselling value stocks and overhyping “glamour” (low \(B/M\)) stocks. The premium reflects mean reversion as these valuation errors correct.

Momentum#

  • Risk-based: Momentum strategies bear asymmetric tail risk, notably rare, severe “momentum crashes” (e.g., sharp reversals during market recoveries).

  • Behavioral: The underreaction hypothesis suggests investors conservatively update beliefs. Prices drift gradually as new information is slowly incorporated, creating predictable trends until full adjustment occurs.

Note: A factor’s positive mean return does not uniquely identify its source. Persistent premia may reflect rational compensation for hidden tail risks, behavioral mispricing sustained by limits to arbitrage, or both.

Testing Multifactor Pricing Models#

A multifactor pricing model makes two related predictions:

  1. Time-series restriction: An asset’s expected excess return should be explained by its exposure to the traded factors. Its regression intercept (alpha) should therefore be zero.

  2. Cross-sectional restriction: Differences in expected excess returns across assets should be explained by differences in their factor exposures (betas).

These predictions motivate two complementary approaches: time-series tests of alpha and cross-sectional tests of expected returns. A third approach, factor spanning, evaluates whether a new factor provides incremental value beyond existing factors.

Time-Series Tests: Are the Alphas Zero?#

For \(N\) test assets and \(K\) traded factors over \(T\) periods, estimate the time-series regression for each asset \(i\):

\[ r_{i,t}^{e}=\alpha_i+\boldsymbol{\beta}_i^\top\mathbf f_t+\varepsilon_{i,t} \]
where \(r_{i,t}^{e}\) is the excess return. A correctly specified model implies no unexplained average returns:
\[ \boxed{H_0:\alpha_1=\alpha_2=\cdots=\alpha_N=0.} \]

OLS Estimation Let \(\boldsymbol{\theta}_i=(\alpha_i, \boldsymbol{\beta}_i^\top)^\top\), \(X\) be the \(T\times(K+1)\) design matrix (including an intercept), and \(\mathbf y_i\) the vector of excess returns. Minimizing the sum of squared residuals yields the OLS estimator:

\[ \boxed{\widehat{\boldsymbol{\theta}}_i=(X^\top X)^{-1}X^\top\mathbf y_i \quad\Longrightarrow\quad \widehat\alpha_i=\bar r_i^{e}-\widehat{\boldsymbol{\beta}}_i^\top\bar{\mathbf f}.} \]
The estimated alpha is the asset’s average excess return minus the return explained by its factor exposures.

Inference and Interpretation Test individual alphas using \(t_{\alpha_i}=\widehat\alpha_i/\widehat{\operatorname{SE}}(\widehat\alpha_i)\). Standard errors must reflect the residual dependence structure; use HAC estimators to account for heteroskedasticity, serial correlation, or overlapping holding periods.

Crucial distinction: A high time-series \(R^2\) measures the fraction of realized return variation explained by the factors. It does not establish correct pricing. A model can explain 95% of return variance while leaving an economically significant alpha.

Joint Alpha Tests: Wald and GRS#

Individual alpha tests ignore correlations among pricing errors. To test whether the model jointly prices all \(N\) assets, evaluate:

\[ \boxed{H_0:\boldsymbol{\alpha}=\mathbf{0}} \]

Wald Test#

Let \(\widehat{\boldsymbol{\alpha}}\) be the estimated alphas and \(\widehat V_\alpha\) their sampling covariance matrix. The Wald statistic is:

\[ \boxed{W=\widehat{\boldsymbol{\alpha}}^\top \widehat V_\alpha^{-1} \widehat{\boldsymbol{\alpha}} \xrightarrow{d}\chi_N^2} \]

Under standard regularity conditions, \(W\) is asymptotically chi-squared with \(N\) degrees of freedom, measuring the joint magnitude of alphas relative to estimation uncertainty. Note: If \(\widehat V_\alpha\) estimates the covariance of \(\sqrt{T}\widehat{\boldsymbol{\alpha}}\) rather than \(\widehat{\boldsymbol{\alpha}}\), multiply the quadratic form by \(T\).

Gibbons–Ross–Shanken (GRS) Test#

The GRS test is exact when regression errors are independent over time, multivariate normal with a constant nonsingular covariance matrix, and independent of the factors. The factor design must have full rank and \(T>N+K\). Conditional on observed factors, the covariance of estimated alphas is:

\[ \operatorname{Var}(\widehat{\boldsymbol{\alpha}}\mid X) = [(X^\top X)^{-1}]_{11}\Sigma_\varepsilon \]

where \(\Sigma_\varepsilon\) is the residual covariance matrix. Defining the sample factor covariance as \(S_f=\frac{1}{T}\sum_{t=1}^T (\mathbf{f}_t-\bar{\mathbf{f}})(\mathbf{f}_t-\bar{\mathbf{f}})^\top\), block matrix inversion yields:

\[ [(X^\top X)^{-1}]_{11} = \frac{1}{T}\left(1+\bar{\mathbf{f}}^\top S_f^{-1}\bar{\mathbf{f}}\right) \]

Using the residual covariance estimator \(S_\varepsilon=\frac{1}{T}\sum_{t=1}^T \widehat{\boldsymbol{\varepsilon}}_t \widehat{\boldsymbol{\varepsilon}}_t^\top\), the GRS statistic is:

\[ \boxed{GRS = \frac{T-N-K}{N} \frac{\widehat{\boldsymbol{\alpha}}^\top S_\varepsilon^{-1} \widehat{\boldsymbol{\alpha}}}{1+\bar{\mathbf{f}}^\top S_f^{-1}\bar{\mathbf{f}}} \sim F_{N,T-N-K}} \]

Interpretation and Assumptions: The numerator measures joint pricing errors relative to residual risk. The denominator adjusts for the squared sample Sharpe ratio of the maximum-Sharpe portfolio of the factors.

Inference Caveats: Wald is a general large-sample test; GRS is an exact finite-sample test under stronger assumptions. With nonnormality or dependence, use robust inference or resampling methods. Failure to reject \(H_0\) does not prove correct pricing, as the test may lack power to detect economically meaningful alphas.

Cross-Sectional Tests: Do Betas Explain Average Returns?#

Time-series tests examine unexplained average returns. Cross-sectional tests ask whether differences in factor exposures explain differences in expected returns.

The standard approach uses two passes.

First pass: Estimate betas: For each asset, regress excess returns on the factor returns:

\[ r_{i,t}^{e} =\alpha_i+\boldsymbol{\beta}_i^\top\mathbf f_t+\varepsilon_{i,t}. \]

Second pass: Estimate factor premia: Regress average excess returns across assets on their estimated betas:

\[ \boxed{ \bar r_i^{e} =\gamma_0+\widehat{\boldsymbol{\beta}}_i^\top \boldsymbol{\lambda}+u_i. } \]

Here \(\gamma_0\) is the cross-sectional intercept, \(\boldsymbol{\lambda}\) contains factor risk premia, and \(u_i\) is the pricing residual.

With \(Z=[\mathbf1,\widehat B]\) and \(\boldsymbol{\theta}_{CS}=(\gamma_0,\boldsymbol{\lambda}^\top)^\top\), OLS gives

\[ \boxed{ \widehat{\boldsymbol{\theta}}_{CS} =(Z^\top Z)^{-1}Z^\top\bar{\mathbf r}^{e}. } \]

The regression requires sufficient independent variation in factor betas across assets.

Under an exact traded-factor pricing model, the population restrictions are

\[ \boxed{ \gamma_0=0,\qquad \boldsymbol{\lambda}=E[\mathbf f],\qquad E[\mathbf r^{e}]=B\boldsymbol{\lambda}. } \]

Expected excess returns must equal the sum of factor exposures multiplied by their corresponding premia. Estimating unrestricted second-pass slopes tests a weaker relation unless the traded factors’ own-return restrictions are also imposed.

Time-Series vs. Cross-Sectional \(R^2\)#

Time-series

Cross-sectional

Observations

Time periods

Assets

Dependent variable

Realized excess returns

Average excess returns

Explanatory variables

Factor returns

Estimated betas

Pricing restriction

\(\boldsymbol{\alpha}=\mathbf0\)

\(E[\mathbf r^e]=B\boldsymbol{\lambda}\)

\(R^2\) interpretation

Return variation explained

Mean-return dispersion explained

Exact population pricing implies cross-sectional \(R^2=1\) when expected-return dispersion is nonzero. In samples, estimation noise makes this an unrealistic decision rule.

Test-asset selection also matters: a model may fit portfolios sorted on the same characteristics used to construct its factors but fail to price other assets.

Fama–MacBeth Regressions (Advanced Extension)#

Instead of a single regression on average returns, Fama–MacBeth runs a cross-sectional regression each period and averages the factor premia. For each date \(t\):

\[ r_{i,t}^{e} = \gamma_{0,t} + \widehat{\boldsymbol{\beta}}_{i,t-1}^\top \boldsymbol{\lambda}_t + u_{i,t} \]

The risk price estimates are the time-series average:

\[ \boxed{\widehat{\boldsymbol{\lambda}} = \frac{1}{T}\sum_{t=1}^{T}\widehat{\boldsymbol{\lambda}}_t} \]

If period-by-period slopes are independent, their sampling covariance is:

\[ \boxed{\widehat{\operatorname{Var}}(\widehat{\boldsymbol{\lambda}}) = \frac{1}{T(T-1)}\sum_{t=1}^{T}(\widehat{\boldsymbol{\lambda}}_t-\widehat{\boldsymbol{\lambda}})(\widehat{\boldsymbol{\lambda}}_t-\widehat{\boldsymbol{\lambda}})^\top} \]

For serially correlated slopes, apply a long-run covariance adjustment (e.g., Newey–West). This approach bases inference on the time series of cross-sectional slopes, accommodating contemporaneous cross-asset dependence without treating every asset-period observation as independent.

Crucial Caveats:

  • Errors-in-variables: Estimated betas introduce measurement error. Fama–MacBeth standard errors do not automatically correct for beta estimation uncertainty.

  • Predictive validity: For forecasting, estimate betas using strictly pre-period information. Model selection and validation must remain outside the final evaluation window.

Factor Spanning: Does a New Factor Add Anything?#

A factor may earn a significant average return without improving an existing model if its premium is already spanned by other factors. Regress a candidate traded factor \(g_t\) on incumbent factors \(\mathbf{f}_t\):

\[ \boxed{g_t = a_g + \mathbf{b}_g^\top \mathbf{f}_t + u_t} \]

Test \(H_0: a_g = 0\). With unrestricted portfolio weights, access to the risk-free asset, and a nonsingular covariance matrix, the population implications are:

  • \(a_g = 0\): The factor adds no incremental Sharpe ratio.

  • \(a_g \neq 0\): The factor improves the attainable mean-variance frontier.

This factor spanning test determines whether a candidate provides investment opportunities not already spanned by the existing set.

Mean-Variance Interpretation For factor expected excess returns \(\boldsymbol{\lambda}_z\) and covariance matrix \(\Sigma_z\), the unconstrained optimal portfolio is:

\[ \boxed{\mathbf{w}^* = \frac{1}{\gamma} \Sigma_z^{-1} \boldsymbol{\lambda}_z} \]

If a factor receives zero optimal weight, removing it does not reduce the maximum attainable Sharpe ratio (assuming unrestricted weights and nonsingular covariance).

Example (SMB): Construct the tangency portfolio from MKT, SMB, and HML. If SMB receives zero weight, it adds no incremental mean-variance benefit beyond MKT and HML, regardless of its stand-alone average return or \(t\)-statistic.

A factor’s stand-alone profitability is not evidence of incremental value. Spanning requires appropriate statistical inference, including multiple-testing adjustments when evaluating many candidates.

Pricing Diagnostics#

Question

Method

Criterion

Do factors explain realized return variation?

Time-series regression

Time-series \(R^2\)

Does the model jointly price assets?

Wald / GRS

\(\boldsymbol{\alpha}=\mathbf0\)

Do betas explain average returns?

Cross-sectional regression

Pricing errors and factor premia

Are factor premia statistically significant?

Fama–MacBeth

Average slopes and adjusted standard errors

Does a factor add incremental value?

Factor spanning

\(a_g=0\)

Does it improve portfolio efficiency?

Mean–variance optimization

Incremental maximum Sharpe ratio

Momentum Anomaly#

Momentum is the tendency for assets with high past returns to outperform those with low past returns over intermediate horizons. Jegadeesh and Titman (1993) documented this in U.S. equities, demonstrating that while individual returns are difficult to predict, small, persistent predictive relationships can support profitable, diversified strategies.

Small Predictive Power Can Produce Meaningful Returns#

Individual stock returns exhibit little exploitable short-horizon autocorrelation. For a stylized AR(1) process with \(\rho=0.05\), the predictive \(R^2\) is merely \(0.0025\) (0.25%). While negligible for a single stock, weak predictability becomes economically valuable when applied systematically across many securities.

Information Coefficient (IC)
The IC measures the correlation between forecasted and realized active returns:

\[ \boxed{IC = \operatorname{Corr}(\widehat r_i, r_i)} \]
An IC of 0.05 is a small correlation. Its economic value depends on diversification, portfolio constraints, and trading costs.

Interpreting IC as Directional Accuracy
Under the assumption of centered, jointly Gaussian forecasts and outcomes, the probability of correctly forecasting the sign of a residual return (Hit Rate) is:

\[ \boxed{\text{Hit Rate} = \frac{1}{2} + \frac{\arcsin(IC)}{\pi} \approx \frac{1}{2} + \frac{IC}{\pi} \quad \text{(for small } IC)} \]
An IC of 0.05 implies only ~51.6% directional accuracy. The economic value of this edge depends on exploitation frequency and the diversification of independent risk. This mapping does not apply to arbitrary classification scores or rank ICs.

The Fundamental Law of Active Management
Grinold’s Fundamental Law relates forecasting skill to the attainable Information Ratio (IR):

\[ \boxed{IR \approx IC\sqrt{BR}} \]
where \(BR\) (breadth) is the number of effectively independent investment opportunities per year. With portfolio constraints, the Transfer Coefficient (\(TC\)) accounts for implementation efficiency:
\[ \boxed{IR \approx TC \cdot IC\sqrt{BR}} \]
With \(IC=0.05\) and \(TC=1\), 400 independent opportunities theoretically support an IR of 1.0. The mechanism is diversification: weak expected returns accumulate across independent bets, offsetting uncorrelated risks.

Caveat: Effective breadth depends on cross-sectional correlations, common factor exposures, and overlapping holding periods. Trading 1,000 correlated stocks does not yield 1,000 independent bets.

Constructing the Momentum Factor#

Conventional cross-sectional momentum measures cumulative returns over the prior 12 months, skipping the most recent month to avoid short-term reversal.

\[ \boxed{MOM_{i,t}^{12-2} = \prod_{j=2}^{12}(1+R_{i,t-j}) - 1} \]
Stocks are ranked by this metric to form a Winner-Minus-Loser (WML) or Up-Minus-Down (UMD) portfolio: \(R_{WML,t} = R_{W,t} - R_{L,t}\). Novy-Marx (2012) notes that predictive power largely stems from intermediate performance (months 12 through 7), not the recent lookback.

Distinction: \(MOM_{i,t}^{12-2}\) is a stock-level sorting characteristic. WML/UMD is the realized factor return. In factor models (e.g., Carhart), stocks are priced by their covariance exposure to UMD, not their raw past-return ranking.

Implementation: Turnover and Transaction Costs#

Momentum requires periodic rebalancing. While a 12-month lookback yields more persistent rankings than a 1-month signal due to overlapping windows, turnover can remain substantial. Its magnitude depends on the universe, portfolio cutoffs, weighting, and rebalancing rule. Rankings are relative; a stock with positive trailing returns can exit the winner portfolio if outperformed by peers.

One-way portfolio turnover is:

\[ \boxed{TO_t = \frac{1}{2}\sum_i |w_{i,t}^{target} - w_{i,t}^{pre-trade}|} \]
Approximate net returns are:
\[ \boxed{R_t^{net} \approx R_t^{gross} - \sum_i c_{i,t}|\Delta w_{i,t}| - \text{Borrow}_t - \text{Financing}_t} \]
where \(c_{i,t}\) captures execution costs (commissions, spread, market impact). Implementation mitigations include longer signal windows, rebalancing bands, less frequent trading, and liquidity-aware sizing. These reduce costs but may dilute signal exposure. The objective is net risk-adjusted performance, not gross returns.

Momentum Across Asset Classes#

Asness, Moskowitz, and Pedersen (2013) study value and momentum across markets. Table I reports the following results for their rank-weighted, long-short momentum factors, before trading costs. Returns and alphas are annualized; Sharpe ratios are annualized.

Asset class

Excess return

Benchmark alpha

Sharpe ratio

U.S. stocks

7.7%

8.7%

0.45

Global stocks

7.1%

6.4%

0.59

Currencies

3.0%

3.1%

0.32

Commodities

11.5%

10.5%

0.51

The sample runs from January 1972 to July 2011 for the stock aggregates and commodities, and from January 1979 to July 2011 for currencies. Alpha uses the paper’s asset-class benchmark; it is not uniformly an equity-CAPM alpha. Global stock strategies are combined using equal-volatility weights.

Source: Asness, Moskowitz, and Pedersen (2013), Value and Momentum Everywhere, Table I, momentum “Factor” column.

Value and momentum are negatively correlated in these samples. These historical results do not identify whether the premia compensate for risk or reflect mispricing.

Economic Explanations for Momentum#

Three primary explanations dominate the literature:

Explanation

Economic mechanism

Empirical implication

Underreaction

Gradual information incorporation into prices

Returns continue in the direction of earlier news

Overreaction

Extrapolation of past performance beyond fundamentals

Short/intermediate continuation followed by long-term reversal

Risk compensation

Undesirable payoffs in specific economic states

Positive average returns compensate for exposure to these states

Underreaction and overreaction yield similar short-run predictions but diverge over longer horizons (gradual convergence vs. eventual correction). Risk-based explanations attribute profits to exposures like liquidity risk or adverse tail events. Momentum’s persistence remains an open question; trading costs, limits to arbitrage, and crash risk may prevent its full elimination.

Momentum Crashes and Idiosyncratic Momentum#

Momentum Crashes
Conventional momentum suffers severe losses during sharp market rebounds following prolonged declines. Past losers often include distressed, high-beta stocks, while past winners are defensive. A sharp rebound causes former losers to outperform winners, generating massive drawdowns for the WML portfolio. Daniel and Moskowitz (2016) show these crashes concentrate in panic states (prior market declines, high volatility, subsequent rebounds), creating an asymmetry of modest steady gains punctuated by severe drawdowns.

Idiosyncratic (Residual) Momentum
Conventional momentum uses total past returns, conflating systematic factor performance with firm-specific returns. Residual momentum isolates firm-specific continuation by removing returns explained by common risk factors.

First, estimate historical factor exposures to calculate residual returns:

\[ \widehat\varepsilon_{i,t} = r_{i,t}^{e} - \widehat\alpha_i - \widehat{\boldsymbol\beta}_i^\top \mathbf{f}_t \]
Then, construct a standardized residual momentum signal:
\[ \boxed{RMOM_{i,t} = \frac{1}{\widehat\sigma_i} \sum_{j=2}^{12} \widehat\varepsilon_{i,t-j}} \]
where \(\widehat\sigma_i\) is the estimated idiosyncratic volatility, using only information available at the signal formation date. Stocks are ranked on \(RMOM\) rather than total returns. This discounts performance driven merely by favorable market or industry exposure.

Blitz, Huij, and Martens (2011) find residual momentum exhibits more consistent performance, and stronger risk-adjusted returns than conventional momentum.

Blitz, Hanauer, and Vidojevic (2020) confirm idiosyncratic momentum is distinct, and conventional crash-risk explanations do not fully account for its returns. However, residualization reduces but does not eliminate crash risk; factor exposures shift, and residual returns can correlate during market stress.

The Factor Zoo, Data Mining, and Machine Learning#

The Factor Zoo: Are There Hundreds of Factors?#

The “factor zoo” refers to the hundreds of characteristics proposed to explain cross-sectional returns. The core issue is distinguishing distinct priced risks from overlapping signals, statistical artifacts, and data-mined discoveries.

Claims are often conflated. Evidence requirements differ:

Claim

Evidence Required

Characteristic predicts returns

Reliable out-of-sample return predictability

Factor earns a premium

Statistically significant portfolio returns

Factor adds new information

Incremental pricing or spanning evidence

Factor represents priced risk

Credible economic mechanism and risk exposure

Strategy is investable

Net performance, stability, liquidity, and capacity

Replication Evidence (Hou, Xue, and Zhang, 2020):
Systematic replication of 452 published anomalies using consistent methods (NYSE breakpoints, value-weighting to mitigate microcap effects) reveals:

  • 65% fail the conventional \(|t|=1.96\) threshold.

  • 82% fail the multiple-testing-adjusted threshold of \(|t|=2.78\). Many effects rely heavily on microcaps, equal-weighting, or specific implementation choices, raising the evidentiary standard for new factors.

Multiple Testing and P-Hacking#

Conventional significance thresholds are inadequate when testing hundreds of candidate factors (Harvey, Liu, and Zhu, 2016).

The Multiple-Testing Problem
Testing \(m\) independent null hypotheses at significance level \(\alpha\) yields:

\[ \boxed{E[V]=m\alpha} \quad \text{and} \quad \boxed{P(V\geq1)=1-(1-\alpha)^m} \]
For 100 tests at \(\alpha=0.05\), we expect 5 false positives, with a 99.4% probability of at least one.

Harvey’s Critique
Due to p-hacking (testing variations until one works), publication bias, and selection bias, Harvey et al. recommend \(|t|>3\) as a more appropriate benchmark for newly proposed factors.

Post-Publication Decay (McLean and Pontiff, 2016)
Average predictor-portfolio returns decline:

  • 26% out-of-sample relative to the original study.

  • 58% after publication (a difference of 32 percentage points relative to the original-study return, which the authors attribute to publication-informed trading).

Machine Learning in Asset Pricing#

Machine learning manages high-dimensional, correlated predictors and nonlinearities, but three distinct objectives must be separated:

Objective

Model

Evaluation

Predict expected returns

\(\widehat{\mu}_{i,t+1}=g_\theta(\mathbf{z}_{i,t},\mathbf{s}_t)\)

OOS \(R^2\), IC, portfolio performance

Model return comovement

\(\mathbf{r}_{t+1}=B_t\mathbf{f}_{t+1}+\boldsymbol{\varepsilon}_{t+1}\)

Covariance fit and residual risk

Price expected returns

\(E_t[\mathbf{r}_{t+1}]=B_t\boldsymbol{\lambda}_t\)

Pricing errors and factor spanning

Better return prediction does not imply better factor pricing. A model can forecast returns accurately without identifying systematic risk, or explain return covariance without correctly pricing average returns.

Return prediction (Gu, Kelly, and Xiu, 2020). Trees and neural networks outperform linear models by capturing nonlinear characteristic interactions. Individual-stock out-of-sample \(R^2\) is small, but a diversified long–short portfolio on the forecasts earns a value-weighted Sharpe ratio of about 1.35, gross of costs.

\[ \boxed{R^2_{\mathrm{OOS}} = 1 - \frac{\sum_{(i,t)}(r_{i,t+1}-\widehat{\mu}_{i,t+1})^2}{\sum_{(i,t)}(r_{i,t+1}-\widehat{\mu}^{\mathrm{bench}}_{i,t+1})^2}} \]

Research Standards and Practical Implications#

Machine learning exacerbates overfitting; searching numerous features, architectures, and hyperparameters can fit historical noise despite regularization. A credible empirical design requires:

  1. Economic motivation and point-in-time data: Avoid survivorship, look-ahead bias, and unrealistic microcap assumptions.

  2. Independent evaluation: Separate model development from testing using temporal validation, penalizing repeated specification searches.

  3. Incremental evidence: Test new factors against established models, not just stand-alone returns.

  4. Robustness: Evaluate across periods, markets, weighting schemes, and post-publication samples.

  5. Economic implementation: Report turnover, transaction costs, borrow constraints, and capacity.

ML improves prediction and manages high-dimensional universes, but it does not eliminate multiple-testing bias or provide economic explanations merely through better fit.

Selected References#

Contents