Cross-Sectional vs. Time-Series Strategies

Almost every systematic strategy is built in one of two ways, and the choice shapes everything downstream, the P&L profile, the market exposure, the breadth, the crash behavior. Cross-sectional strategies rank assets against each other and bet on the winners beating the losers. Time-series strategies compare each asset against its own history and bet on the sign of its own signal. Understanding the difference, and the precise sense in which they are related, is foundational to systematic trading.

The two designs

Take momentum as the running example. Let $s_{i,t}$ be a signal (say, past 12-month return) for asset $i$.

Cross-sectional (relative). Demean the signal across assets at each date, then weight in proportion to the demeaned signal:

$w^{\text{CS}}_{i,t} \propto s_{i,t} - \bar s_t, \qquad \bar s_t = \frac{1}{N}\sum_j s_{j,t}.$

You are long the above-average assets and short the below-average ones. Because the weights sum to zero by construction ($\sum_i w^{\text{CS}}_{i,t}=0$), the portfolio is dollar-neutral and bets purely on relative performance, it profits if winners beat losers even if the whole market falls.

Time-series (absolute). Compare each asset to a fixed benchmark (often zero excess return) and weight by its own signal's sign/magnitude, typically scaled by its own volatility:

$w^{\text{TS}}_{i,t} \propto \frac{s_{i,t}}{\sigma_{i,t}}.$

Here the weights do not sum to zero: if most assets have positive signals you are net long, if most are negative you are net short. A time-series strategy therefore takes directional / net market exposure that varies with the aggregate signal, which is exactly why Trend Following is long the market in bull runs and short in crashes.

Demeaning is the whole story

The single operational difference is what you subtract. Cross-sectional demeans across the cross-section (assets at one time); time-series demeans across time (one asset's own history). That choice determines:

• Market neutrality. Cross-sectional zero-sum weights ⇒ (dollar) market-neutral; time-series ⇒ net directional. To make a time-series book market-neutral you must separately hedge the aggregate exposure.
• What you're paid for. Cross-sectional harvests dispersion (the spread between best and worst); time-series harvests direction (the average asset going the way its signal points).
• Breadth. Cross-sectional breadth grows with the number of assets $N$ (many simultaneous relative bets); time-series breadth is closer to the number of assets times independent trends, often the strategies are combined precisely to stack their breadth (Signal Construction).

The algebraic link

Cross-sectional and time-series momentum are not independent phenomena, they are related by an identity. Lo and MacKinlay decomposed the expected profit of a cross-sectional (relative-strength) momentum strategy into three pieces:

$\mathbb{E}[\pi^{\text{CS}}] = \underbrace{\sigma^2(\mu)}_{\text{cross-sectional dispersion in means}} + \underbrace{\text{own-autocovariance}}_{\text{time-series predictability}} - \underbrace{\text{cross-serial covariance (lead-lag)}}_{\text{one asset predicts another}}.$

Reading it: cross-sectional momentum profits come partly from time-series auto-predictability (each asset trending, the time-series effect), partly from stable differences in mean returns across assets, and partly (with a minus sign) from lead-lag relationships. The practical upshot, made precise by Moskowitz-Ooi-Pedersen, is that time-series momentum is the more primitive effect, a single asset's own trend, and cross-sectional momentum is largely time-series momentum plus the dispersion in average returns. They are two views of overlapping return predictability, which is why they are correlated but not identical, and why running both adds diversification rather than redundancy.

Worked example

Three stocks have 12-month returns of +30%, +10%, and −20%; the cross-sectional mean is +6.7%.

• Cross-sectional: demeaned signals are +23.3%, +3.3%, −26.7%. You go long the first two (above average) and short the third, dollar-neutral, even though two of three rose, you short the laggard. If the market rallies and all three rise 10%, your relative bet is unaffected by the common move.
• Time-series: signs are +, +, − (using zero as benchmark). You go long the first two and short the third, here it coincides, but note you are net long (two longs, one short), so a market-wide selloff hurts you. If instead all three signals were positive, the time-series book would be fully long the market while the cross-sectional book would be roughly flat.

Same raw signals, different portfolios, different exposures, that is the entire distinction.

Failure modes

• Hidden market beta in "neutral" books. Dollar-neutral is not beta-neutral; if your longs are higher-beta than your shorts, a cross-sectional book still has net market exposure. Neutralize on beta, not dollars.
• Time-series books are directional bets. A time-series strategy's returns are dominated by the aggregate signal; in a whipsaw market it gets chopped up flipping net-long/net-short.
• Correlated cross-section = low effective breadth. Cross-sectional strategies rely on many independent relative bets; in a correlated selloff the cross-section moves together and the diversification the design assumes disappears.
• Demeaning against the wrong universe. Cross-sectional demeaning implicitly defines "average" by your universe; a biased or too-narrow universe injects unintended factor bets.

In interviews

Explain the two constructions in one line each: cross-sectional ranks assets against each other and demeans across the cross-section (dollar-neutral, harvests dispersion); time-series compares each asset to its own history and does not demean across assets (net directional, harvests trend). The subtle question is "are cross-sectional and time-series momentum the same thing?", no, but they are algebraically linked (Lo-MacKinlay decomposition), time-series momentum is the more primitive own-asset effect and cross-sectional adds the dispersion-in-means term, which is why the two are correlated yet diversifying. Flag that dollar-neutral ≠ market-neutral. See Trend Following (the canonical time-series strategy) and Statistical Arbitrage (the canonical cross-sectional one).