Cross-Sectional Momentum and Equity Factor Portfolios

Cross-sectional momentum ranks a universe of assets by past return and goes long the winners while shorting the losers — it bets on relative performance persisting, not on the market going up. This is the single most studied equity anomaly and the workhorse of quant equity long-short, but the version that survives live is almost unrecognizable from the textbook "buy the top decile" sketch. The money is in the neutralization, the turnover control, and the honest accounting of crowding and crash risk. It is a close cousin of, but mechanically distinct from, time-series trend following, and it sits inside the broader factor investing framework.

Cross-sectional vs time-series momentum

The distinction is not pedantic — it determines your market exposure:

Time-series (trend) momentum asks "is this asset above its own past?" and

takes a long or short on each asset independently. The aggregate book can be net long or net short the market, so it carries directional beta and tends to be long volatility (it profits from sustained trends and large moves).

Cross-sectional momentum asks "which assets beat their peers?" and is

dollar-balanced long-short by construction. It is (approximately) market-neutral and bets purely on the ranking persisting. Winners keep beating losers even if the whole market is flat.

You can be long-trend and short-cross-sectional simultaneously, and they have different factor loadings, drawdown timing, and capacity. The classic formation is 12-1 momentum: rank on the cumulative return from 12 months ago to 1 month ago, skipping the most recent month to avoid the short-term reversal (1-month) effect, which is microstructure-driven mean reversion that contaminates the signal if you include it.

Ranking, deciles, and the signal

The raw signal is a cross-sectional rank or z-score of the formation-period return, recomputed each rebalance. The two standard ways to turn ranks into weights:

Decile (or quantile) sort: long the top 10%, short the bottom 10%, equal- or

cap-weighted within. Simple, robust, throws away information in the middle.

Rank- or z-score-weighted: weight monotonically in the signal across the whole

universe. Uses all the cross-section, smoother, but loads more on the extreme, often illiquid names unless capped.

import numpy as np
import pandas as pd

def momentum_signal(prices, lookback=252, skip=21):
    # 12-1 momentum: return from t-lookback to t-skip, skipping last month
    past = prices.shift(skip)
    formation = past / prices.shift(lookback) - 1.0
    return formation

def cross_sectional_zscore(signal_row):
    s = signal_row.dropna()
    if len(s) < 20:
        return pd.Series(np.nan, index=signal_row.index)
    z = (s - s.mean()) / s.std()
    return z.clip(-3, 3).reindex(signal_row.index)   # winsorize tails

def long_short_weights(z_row, gross=1.0):
    z = z_row.dropna()
    w = z - z.mean()                 # demean -> dollar neutral
    w = w / np.abs(w).sum() * gross  # scale to target gross
    return w.reindex(z_row.index).fillna(0.0)

The unglamorous truth: most of the cross-sectional spread in raw momentum returns is not pure alpha — it is exposure to other factors that happen to correlate with the ranking at that moment. Which is why neutralization is not optional polish; it is the strategy.

Neutralization is where the alpha actually is

A raw winners-minus-losers portfolio is a tangle of unintended bets. Momentum portfolios systematically drift into sector tilts (whatever sector ran loads the long book), into beta (in a recovery, high-beta names dominate the winners), and into size and value. When that tilt reverses you get a "momentum crash" that has nothing to do with the ranking thesis failing. Neutralize explicitly:

Sector/industry neutral: demean the signal within each sector, so you are

long the best names in each sector against the worst in the same sector. This removes the accidental sector bet.

Beta neutral: regress the book's returns on the market and size the legs so net

beta is ~0; otherwise your "market-neutral" book is a leveraged market bet in disguise.

Dollar and (optionally) factor neutral: demean to dollar-neutral, then

optionally residualize the momentum signal against value, size, and quality so you isolate the momentum premium rather than re-buying correlated factors. This is best done inside a portfolio optimizer with factor constraints rather than by naive sorting.

The empirical regularity: residualizing momentum against the other factors raises its risk-adjusted return and dramatically reduces the crash, because the crashes are largely a beta/illiquidity artifact, not a failure of relative-strength itself. Run a performance attribution on the raw book and you will see how much of the "momentum P&L" was really uncompensated beta.

Turnover, costs, and the capacity wall

Momentum is a high-turnover factor — names enter and exit the winner/loser sets constantly — and turnover is the enemy that quietly converts a great paper Sharpe into a mediocre live one. The half-life of the 12-1 signal is months, but a naive monthly rebalance to fresh deciles can churn 30-60% of the book per rebalance, and every unit of turnover pays the spread plus market impact.

Rebalance / band	Annual turnover (1-way)	Gross Sharpe (paper)	Net after ~15bps
Monthly, hard decile	very high	high	sharply reduced
Monthly, no-trade band	moderate	slightly lower	materially better
Quarterly + banding	low	lower	often best net

The non-obvious lever is no-trade banding / hysteresis: do not rebalance a name out of the book until its rank degrades past a buffer, and do not chase tiny weight changes. This sacrifices a little gross return for a large cut in turnover, and net of costs it usually wins. Combine with trading the change in target weights through a cost-aware scheduler (TWAP/VWAP or impact-optimal) rather than dumping the full delta at the close. Capacity is set by the least liquid names in your extreme buckets: scaling AUM pushes you into smaller names or larger participation rates, and impact grows roughly with the square root of size — past some AUM the strategy mathematically cannot clear its own costs.

def apply_no_trade_band(target_w, current_w, band=0.003):
    # Only move weights that have drifted beyond the band; reduces turnover
    delta = target_w - current_w
    delta[np.abs(delta) < band] = 0.0
    return current_w + delta

Factor crowding and the momentum crash

Momentum's defining risk is not gradual decay — it is the momentum crash: sharp, violent losses concentrated in market rebounds after a selloff. The mechanism is structural. After a crash, the short (loser) book is full of beaten-down, high-beta, distressed names; when the market snaps back, those losers rip upward fastest and the short leg detonates. The factor's own return distribution is left-skewed and its volatility is time-varying — high exactly when the crash is coming. The single most effective mitigation is volatility scaling: lever the factor inversely to its recent realized vol, which targets constant risk and sidesteps much of the crash, since the crash is preceded by a vol spike. See volatility targeting for the mechanics shared with trend.

Crowding compounds this. Momentum is the most widely replicated quant factor, so a large pool of capital holds similar winner/loser baskets. When that capital deleverages together — a risk-off shock, a factor crowding unwind — the crowded names move against the consensus position regardless of fundamentals. The August 2007 "quant quake" was exactly this: a forced deleveraging of crowded long-short books, invisible in any single-name fundamental, that mean-reverted within days but ruined anyone levered into it. Monitor crowding through short interest, factor-return correlation across your peers' proxies, and the liquidity of your extreme buckets — and keep gross leverage well below what the calm- period Sharpe would tempt you into.

Signal variants that actually differ

"Momentum" is a family, and the variants are not cosmetic — they load on different risks and decay at different rates:

Residual (idiosyncratic) momentum. Rank on the residual of a Fama-French (or

statistical-factor) regression rather than raw return. You momentum-trade the part of past performance that is not explained by factor exposure. Empirically this has a higher and more stable risk-adjusted return than raw momentum and a far smaller crash, for the same reason neutralization helps: it strips the beta/size tilt that causes the blowups. The cost is estimation noise in the residuals.

52-week-high momentum. Rank on proximity to the 52-week high price rather than

cumulative return. It captures an anchoring effect and is mechanically lower turnover (a name near its high stays near its high), but it overlaps heavily with raw momentum.

Frog-in-the-pan / path quality. Two names with the same 12-month return are not

equivalent: one that ground up steadily is "higher quality" momentum than one driven by a single jump, and the smooth path predicts continuation better. Encoding path consistency (e.g. the fraction of up-days, or return divided by path volatility) adds information orthogonal to the level of the return.

The practical lesson: the raw 12-1 sort is the worst member of the family that still counts as momentum. Residualizing and conditioning on path quality are where the incremental, less-crowded edge lives, and they degrade more gracefully because they are less mechanically replicated than the textbook decile sort. Combine them as inputs to a single composite signal rather than running three overlapping books.

Implementation horizon and rebalance timing

The signal has a months-long half-life, so the rebalance frequency and the entry timing are free parameters that change the realized return without changing the thesis. Rebalancing all positions on the same calendar day (month-end) concentrates your trades into the most crowded liquidity window and exposes you to month-end factor rotation; overlapping portfolios — splitting capital into, say, four sleeves each rebalanced on a staggered schedule — smooths turnover, diversifies entry timing, and reduces the sensitivity to any single rebalance date. It is the same averaging logic that makes a laddered bond portfolio less path-dependent than a single bullet maturity, applied to factor entry points. None of this adds alpha in a frictionless world; all of it adds net alpha once costs and timing luck are real.

Honest limits

Cross-sectional momentum is real and persistent — it shows up across decades, countries, and asset classes — but the headline academic Sharpe is gross of costs, gross of crowding, and assumes shorting the illiquid loser names is free, which it is not (borrow can be expensive or unavailable exactly when you need the short). The live strategy is a neutralized, vol-targeted, turnover-banded, cost-aware shadow of the paper factor, and its honest net Sharpe is moderate with a nasty left tail. Treat it as one residualized sleeve in a multi-factor book, size it to its tail risk, and never forget that its worst days arrive together with everyone else's worst days.