Cross-Sectional Momentum

Cross-Sectional Momentum

Cross-sectional momentum (Jegadeesh & Titman, 1993) is a zero-cost long-short equity strategy that ranks all stocks in a universe by their past $J$-month return, buys the top decile of “winners,” shorts the bottom decile of “losers,” and holds the resulting portfolio for $K$ months. For $J,K\in \{3,6,9,12{\}}^2$ all sixteen combinations have historically delivered significant positive average returns.

For each formation date $t$, the strategy ranks the universe on cumulative return over $[t-J,t]$, forms equal-weight decile portfolios $D_1,\dots ,D_{10}$, and trades the long-short Winners-Minus-Losers (WML) portfolio
${\text{WML}}_{t,K}=\frac{1}{K}{\sum }_{k=1}^K(r_{D_{10},t+k}-r_{D_1,t+k}).$

Headline empirical results (1965–1989 NYSE/AMEX):

• WML earns approximately $1\%$ per month ($\approx 12\%$ annualized) with $t$-statistics above $3$ (i.e., high statistical significance) for the standard $(J,K)=(6,6)$ specification.
• The CAPM beta of WML is ${\beta }_{\text{WML}}\approx -0.08$, so the strategy is not compensated for market risk; if anything it tilts slightly against the market.
• Continuation persists over horizons of three to twelve months and reverses over months $13$–$31$ (De Bondt & Thaler, 1985 long-term reversal).
• The result has been replicated internationally (Rouwenhorst, 1998), across asset classes (Asness, Moskowitz & Pedersen, 2013), and is itself one of the four Carhart factors (UMD).

Cross-sectional momentum compares assets against each other — its signal is relative, so a stock that fell less than its peers in a falling market still counts as a “winner.” This is its defining contrast with time-series momentum, which compares each asset against its own past. Together with momentum crashes, these three notes form the empirical core of the momentum section. Cross-sectional momentum is widely regarded as the single strongest piece of evidence against the efficient markets hypothesis.

Intuition

Stocks that have beaten their peers over the past several months tend to keep beating them for several more, then reverse. The standard story (see the over- and under-reaction models) has two stages. First, investors are slow to absorb news — from conservatism, slow information spread, or inattention — so prices drift in the direction of the news for months instead of jumping right away. Second, late buyers pile in and push prices past what the news justified; the overshoot corrects over the following years. Both halves are needed: drift without overshoot would not reverse, and overshoot without drift would not continue.

“Cross-sectional” means the rank is against the rest of the universe, not an absolute benchmark — in a market that fell $15\%$, a stock that fell only $5\%$ counts as a winner. The long and short legs cancel out market exposure by design, which is why the average market beta is near zero. That flat average hides a beta that swings sharply with the market state and drives momentum crashes.

Properties

(Overlapping Portfolios) The practical implementation runs $K$ overlapping portfolios in parallel, one formed each month.

Each month one of the $K$ active portfolios is liquidated and a new one is formed; the reported strategy return is the average across the $K$ currently-held cohorts. This staggering smooths out rebalancing noise and lets the strategy be reported as a monthly time series even though each cohort is held for $K$ months. Deciles are equal-weighted within each leg, which gives small caps a disproportionate influence — value-weighting WML cuts the headline number roughly in half but leaves the alpha statistically significant.

(Returns Decay Across the $(J,K)$ Grid) Longer look-backs and shorter holds give the largest gross returns; the cell-by-cell pattern reveals smooth signal decay.

The $(12,1)$ corner of the grid produces the largest gross monthly return because a long look-back collects more signal and a short hold re-ups the signal frequently. The $(3,12)$ corner produces the smallest because it reverses both choices. The smooth monotonic pattern across the $4\times 4$ grid is itself a robustness check: if any single $(J,K)$ cell were a fluke, neighboring cells would not also be positive. After realistic transaction costs the workhorse $(6,6)$ or $(12,6)$ implementations dominate, which is why those are the cells most cited.

(Conditional-Risk Models Also Fail) Even risk-factor models that allow loadings to vary with the business cycle cannot rationalize the WML alpha.

Static factor regressions — Fama-French three-factor, Carhart four-factor — are the first line of defense and they fail. The deeper rational-pricing defense lets risk loadings vary across states (so that a strategy’s true risk only shows up in bad times). These conditional and ICAPM-style models have also been fitted to WML and also fail to absorb the alpha. Two decades of attempts to find a hidden risk story that justifies momentum have produced none, which is what makes momentum the central piece of evidence against the efficient markets hypothesis rather than just another anomaly.

(Tail Risk Hidden in the Mean) The $\approx 1\%$/month average is generated by many small wins punctuated by rare $-50\%$ to $-75\%$ monthly losses.

The mechanism — covered in momentum crashes — is that the loser leg in a bear market fills with high-beta survivors that explode on a rebound, so the strategy’s conditional beta swings from $-0.08$ on average to $-2$ or more in panic states. A practitioner sizing WML from its unconditional Sharpe ratio of $\approx 0.5$ will be undercapitalized for the next crash month. The right way to read the WML mean is not as a steady cash flow but as a long stretch of small gains plus a small handful of catastrophic months whose timing is partly predictable from market state.

Examples

A $(6,6)$ strategy walk-through

Suppose at the start of January $t$, you rank the 1,000 largest U.S. stocks by their July–December cumulative return (the look-back $J=6$). You go long the top 100 (the winner decile) equal-weighted and short the bottom 100 (the loser decile) equal-weighted. You hold this portfolio for six months — through end of June.

To smooth rebalancing you actually run six overlapping portfolios in parallel: one formed in January (held Jan–Jun), one formed in February (held Feb–Jul), and so on. Each month one of the six portfolios is liquidated and a new one is formed. The reported strategy return for month $t$ is the average return of the six currently-active overlapping portfolios.

Over the 1965–1989 sample, this $(6,6)$ implementation returned approximately $0.95\%$ per month with a $t$-statistic above $3$. The CAPM beta of the strategy was approximately $-0.08$, so the positive return cannot be rationalized as compensation for market exposure under standard asset pricing.

Why $(12,1)$ wins more than $(3,12)$

Two strategies built from the same data behave very differently. The $(12,1)$ strategy uses the longest look-back ($12$ months) and the shortest hold ($1$ month) — it captures a maximally informative signal and turns over every month. The $(3,12)$ strategy uses a short, noisy look-back and holds the resulting (potentially stale) signal for a full year.

In the data, $(12,1)$ delivers higher gross average returns than $(3,12)$, consistent with the picture that the underlying alpha decays smoothly over the holding period: longer look-backs collect more signal, and shorter holds re-up that signal more frequently. The cost of $(12,1)$ is transaction costs, which is why practitioners more often run $(6,6)$ or $(12,6)$ as the workhorse parameterization. The pattern of returns across the $4\times 4$ grid of $(J,K)$ combinations is itself a property of the data — every cell is positive, and the cells with longer look-backs and shorter holds are larger.

The De Bondt-Thaler reversal companion

Take the same winner and loser deciles from the $(12,K)$ strategy, but instead of holding for $K\le 12$ months, hold for months 13 through 60. The sign of the strategy flips. Past winners now underperform past losers by approximately $25\%$ cumulative over the subsequent four years (De Bondt & Thaler, 1985).

This is not a contradiction — it is the same data viewed at a different horizon. The cognitive story is that the initial underreaction-driven drift over months $1$–$12$ eventually overshoots fundamentals as late traders extrapolate, and the overshoot is then corrected over months $13$–$60$. Continuation and reversal are two halves of one trajectory, and any model that explains one must explain the other. This is exactly the empirical target of BSV.

Diagram

The construction below shows one momentum portfolio’s life cycle. In a real implementation, $K$ such portfolios run in parallel, staggered by one month each, so the strategy is rebalanced monthly even though each constituent portfolio is held for $K$ months.

\usepackage{tikz}
\begin{document}
\begin{tikzpicture}[x=0.85cm, y=0.6cm]
  % time axis
  \draw[->,thick] (-0.5,0) -- (14,0) node[right] {time (months)};
  \foreach \x in {0,1,...,13} {
    \draw (\x,0.12) -- (\x,-0.12);
  }
  \node[below] at (0,-0.4) {$t-J$};
  \node[below] at (6,-0.4) {$t$};
  \node[below] at (12,-0.4) {$t+K$};
 
  % look-back window
  \draw[thick, blue] (0,1.6) -- (6,1.6);
  \draw[thick, blue] (0,1.4) -- (0,1.8);
  \draw[thick, blue] (6,1.4) -- (6,1.8);
  \node[blue, above] at (3,1.8) {look-back ($J$ months): rank stocks by cumulative return};
 
  % formation point
  \fill[red] (6,0) circle (2.4pt);
  \node[red, below right] at (6,-0.6) {form WML: long $D_{10}$, short $D_1$};
 
  % hold window
  \draw[thick, green!55!black] (6,2.8) -- (12,2.8);
  \draw[thick, green!55!black] (6,2.6) -- (6,3.0);
  \draw[thick, green!55!black] (12,2.6) -- (12,3.0);
  \node[green!55!black, above] at (9,3.0) {hold ($K$ months): realize WML payoff};
 
  % closeout
  \fill[red] (12,0) circle (2.4pt);
  \node[red, below] at (12,-0.4) {close};
\end{tikzpicture}
\end{document}

See Also

• 8.8. Momentum — parent hub: comparison of cross-sectional, time-series, and momentum-crash treatments.
• 8.8.2. Time-Series Momentum — sibling: signs each asset against its own past, not against the cross-section.
• 8.8.3. Momentum Crashes — sibling: the conditional-beta failure mode that the unconditional $\beta \approx -0.08$ hides.
• 8.9.1. Barberis-Shleifer-Vishny Model — leading behavioral model that jointly generates 3-12 month continuation and 12-31 month reversal.
• 8.9.2. Hong-Stein Model — alternative behavioral model based on slow information diffusion.
• 8.3. Challenges to Rational Pricing — momentum is the canonical “predictability puzzle” listed there.