1.4.2 Cross-Sectional Momentum

Basic Prerequisites: Econometrics & Time Series (serial correlation) and Linear Algebra (inner products, rank).

---

1. Intuition & Practical Objective

Cross-sectional momentum (XSMOM) is the Jegadeesh–Titman strategy: every month, rank your universe by trailing 12-month return, buy the winners and short the losers, dollar-neutral, hold ~1 month, repeat. It is a relative bet - an asset “wins” because it beat its peers over the ranking window, regardless of the absolute direction of its return.

Two implementation facts shape everything:

1. Skip the most recent month (the $12\text{-}1$ convention). Rank on months $t-12\dots t-2$, not $t-1$. The last month reverses (Jegadeesh 1990; Lehmann 1990) and, if included, drags the signal toward a trade that loses. This one convention is the single most important detail of the strategy.
2. The portfolio is dollar-neutral by construction. Rank weights $w_i\propto ({\text{rank}}_i-\frac{N+1}{2})$ sum to zero, so the long and short legs offset exactly - no net market exposure. The loser short leg is where both the profit and the crash risk live.

The verified US facts: winners decile (a decile = one tenth of the cross-section, i.e. a 10% rank bucket by past return) earns $\sim 15\%$/yr, losers $\sim -2.5\%$/yr, WML (winners-minus-losers) has a Sharpe $\approx 0.60$–$0.71$ and a negative CAPM beta $(-0.58)$ with annual alpha $\approx 22\%$ (Daniel & Moskowitz 2016, 1927–2013). Momentum profits are essentially a non-January phenomenon (JT 2001).

---

2. Mathematical Ground Truth & Derivations

Signal. For asset $i$ with monthly returns $r_{i,t}$, the $12\text{-}1$ cumulative signal is

$$R_i^{(12\text{-}1)}=\prod _{k=2}^{12}(1+r_{i,t-k})-1.$$

Rank-to-weight map. Rank assets by $R_i$ (1-indexed, best = $N$). Set

$$w_i=\frac{{\text{rank}}_i-\frac{N+1}{2}}{{\sum }_{j=1}^N|{\text{rank}}_j-\frac{N+1}{2}|}.$$

Because ranks run $1\dots N$, the raw weights $({\text{rank}}_i-\frac{N+1}{2})$ are symmetric about zero: the portfolio is dollar-neutral (${\sum }_i{w}_i=0$) and 100% gross (${\sum }_i|w_i|=1$). The next-period WML return is $r_{t+1}^{\text{WML}}={\sum }_i{w}_i{r}_{i,t+1}$.

Why it earns (decomposition). Moskowitz–Ooi–Pedersen (2012) decompose the expected cross-sectional momentum profit as

$$\mathbb{E}\left[r_{t,t+1}^{\text{XS}}\right]=\frac{\mathrm{tr}(\Omega )}{N}-\frac{1^{\prime }\Omega 1}{N^2}+12{\sigma }_m^2,\Omega =\mathbb{E}[(R_{t-12,t}-12\mu )(R_{t,t+1}-\mu {)}^{\prime }],$$

where ${\sigma }_m^2$ is the cross-sectional variance of mean returns. The three channels:

1. Own autocovariance $\mathrm{tr}(\Omega )$ - time-series predictability (this is the TSMOM channel; not required for XSMOM).
2. Cross-serial covariances $-1^{\prime }\Omega 1$ - temporal lead-lag across assets (a past move in one asset predicts another’s next return). Negative cross-serial covariances alone can generate XSMOM profits with no own autocorrelation (Lewellen 2002).
3. Cross-sectional dispersion of unconditional means $12{\sigma }_m^2$ - high-mean assets are also high-past-return assets on average.

Why skip the last month. If the one-month autocorrelation ${\rho }_1<0$ (reversal), then including $r_{i,t-1}$ in the signal loads the portfolio onto an asset whose next-month return is negatively predicted by its most recent move. Dropping month 1 removes this contamination, leaving the positive 2–12 month continuation.

---

3. Computational Implementation - the ranking engine

Standard library only. Builds the skip-month signal, verifies dollar-neutrality and 100% gross leverage exactly, and shows that the skip-month convention matters: skipping the last month preserves the premium, while including it destroys it (short-term reversal).

---

4. Failure Modes & First-Principles Breakdowns

1. The loser short leg is where crashes live. WML’s profit and its crash risk both concentrate in the short losers. In a panic rebound the losers “crash up” (they are levered, high-beta, option-like), so the short leg loses violently - this is the entire momentum-crash mechanism (see 05 · Failure Modes).
2. Skip-month violations destroy the edge. Using month $t-1$ in the signal loads one-month reversal; the verified example above shows SR dropping from $+1.75$ to $-0.07$. This is a first-principles failure: wrong autocorrelation band.
3. Beta exposure contamination. Because winners are stocks that did well in the past, WML loads positively on whatever factor did well over the formation window - after market declines it is long low-beta, short high-beta (loser beta can exceed 3), giving time-varying, crash-prone exposure (Kothari–Shanken 1992; Grundy–Martin 2001).
4. Multiple testing & crowding. With 200+ papers and endless parameter sweeps over lookbacks/holdings, the published momentum Sharpe is inflated by selection; capacity and crowding compress realized returns (see Backtesting Hygiene).

---

5. References

• Jegadeesh & Titman (1993), Returns to Buying Winners and Selling Losers, J. Finance 48(1)
• Jegadeesh (1990) and Lehmann (1990)
• Moskowitz & Grinblatt (1999)
• Kothari & Shanken (1992) and Grundy & Martin (2001)
• Moskowitz, Ooi & Pedersen (2012), Time Series Momentum

---

6. Connected Graph Bridges

• Base: Econometrics & Time Series · Linear Algebra
• Continue: 03 · Time-Series Momentum · 04 · Value–Momentum
• Failure analysis: 05 · Failure Modes · 06 · Advanced Extensions
• Sibling factor view: Fundamental Multi-Factor Models (UMD / momentum factor)