Cross-Sectional Momentum

Cross-sectional momentum is a relative game: take a universe of stocks, rank them by how they have done over the past several months, buy the ones near the top, and short the ones near the bottom. Unlike Time-Series Momentum, the absolute direction of any single stock does not matter — what matters is whether it beat its peers. A stock that fell $5\%$ can still land in the "long" bucket if everything else fell $15\%$. This is the cross-sectional approach, and it is the version economists usually mean when they say "the momentum factor."

Because you are long some names and short others in equal measure, the portfolio is roughly market-neutral: it strips out the overall market move and bets purely on the spread between winners and losers. That long-short factor has a name — WML (winners-minus-losers) or UMD (up-minus-down) — and it is the extra factor Carhart bolted onto the The Fama-French Factor Models model because nothing else in that model explained it.

Building the signal

Rank every stock by its trailing return, then split into buckets (deciles are standard). One important convention: use the return from twelve months ago up to one month ago, skipping the most recent month.

$$\text{signal}_i = r_{i,\,t-12 \to t-1}, \qquad \text{WML} = \bar r_{\text{top decile}} - \bar r_{\text{bottom decile}} .$$

Here $r_{i,\,t-12\to t-1}$ is stock $i$'s return over that window, and WML is the average return of the top-ranked names minus the average of the bottom-ranked names. The skipped last month matters: over very short horizons stocks tend to reverse, not continue, so including the latest month pollutes a momentum signal with short-term mean-reversion noise.

Worked example: rank, then trade

Rank eight stocks by their trailing 12-month return and go long the top two, short the bottom two, equal-weighted.

Stock	Trailing return	Rank	Position
A	+48%	1	Long
B	+35%	2	Long
C	+22%	3	—
D	+14%	4	—
E	+9%	5	—
F	+1%	6	—
G	−12%	7	Short
H	−27%	8	Short

You put $50\%$ of your long capital in A and $50\%$ in B, and short H and G equally. Suppose next month A returns $+3\%$, B $+1\%$, G $-2\%$, H $-4\%$. The long book earns $\tfrac{1}{2}(3\%+1\%)=2\%$; the short book gains $\tfrac{1}{2}(2\%+4\%)=3\%$ because those names kept falling. The strategy return is $2\% + 3\% = 5\%$ — earned with no net market exposure, since the long and short legs offset the market move.

Momentum crashes

The catch is buried in the short leg. After a market crash, the losers you are short are the most beaten-down, highest-beta names. When the market violently rebounds, those names snap back hardest — and your short leg gets run over exactly when you can least afford it. This is why cross-sectional momentum has a sharply negative skew: years of steady gains punctuated by rare, brutal drawdowns, almost always at market turning points (spring 2009 is the textbook case). Daniel and Moskowitz showed these crashes are somewhat predictable — momentum is most dangerous when the market has fallen a lot and volatility is high — which motivates scaling the strategy down in those regimes.

Where it stumbles

• Turnover. Rankings shift monthly, so momentum trades a lot; a gross edge can be eaten alive by costs, especially in small or illiquid names.
• Crowding. Momentum is one of the most-run factors on the planet. Heavy crowding means everyone owns the same winners and shorts the same losers, amplifying the unwind when it turns.
• Negative correlation with value. Momentum's winners are usually expensive and its losers cheap, so it tilts against value — which is precisely why the two are run together, since their drawdowns rarely coincide.
• Tuning. The 12-month formation and skip-a-month rules are conventions; fitting them to your sample is overfitting.