Time-Series Versus Cross-Sectional Momentum

Identify the comparison being made

A time-series question is whether one series has risen or fallen over a stated interval. A cross-sectional question is whether its measured change is stronger or weaker than changes elsewhere. The distinction is explicit in Moskowitz, Ooi and Pedersen's academic treatment of time-series momentum. Here we examine the definitions, not reproduce that paper's portfolio or performance results.

Start by asking what could change a label while the instrument's own observations remain fixed. Under a simple own-history sign rule, changes in other instruments do nothing. Under a ranking rule, another instrument's movement can alter the rank immediately. This small thought experiment identifies the comparison more reliably than a strategy name containing the word momentum.

Relative winners can still be falling

Consider three hypothetical positive reference series, all starting at 100. They finish at 108, 103 and 96, giving price changes of 8%, 3% and -4%. An own-history sign rule labels the first two positive and the third negative. A relative rule selecting only the strongest and weakest instead distinguishes the first and third, leaving the middle series unselected.

Now let the endpoints be 98, 94 and 90. Every own-history label is negative, but the first series is still the relative winner. Calling it a winner describes its position within this group, not a positive price change. Conversely, a relative loser in a universally rising group can have a positive own-history label. No future outcome follows from either classification.

The universe is part of the rule

An own-history comparison requires an input series and a time window. A relative comparison additionally needs an eligible universe, ranking measure and selection rule. Adding a fourth series with a 12% change displaces the 8% series from first place without changing any of its prices. A published relative label is therefore incomplete when its comparison group is undisclosed.

The Kenneth French daily momentum factor documentation gives a concrete academic example of portfolios formed using prior returns and size. That equity construction illustrates why eligibility and grouping matter; it is not a futures recipe. Transferring its label to futures without specifying contract histories and measurement conventions would conceal important differences.

Separate direction from weighting

A positive label is not a request to allocate a fixed amount of money. Direction can be represented as +1, 0 or -1, while an independent sizing convention determines any exposure. Two instruments with the same sign may have very different quote units, contract multipliers and variability. Counting matching signs says nothing by itself about balanced economic risk.

Similarly, equal numbers of positive and negative labels do not establish neutrality. In a hypothetical ranking of four instruments, two receive +1 and two receive -1. Their labels sum to zero, but their sensitivities need not cancel. Neutrality requires a specified quantity, such as notional exposure or a measured factor sensitivity, plus the information needed to calculate it.

Make the input measure comparable

Ranking raw point changes can compare incompatible units. A move of five points in one series and two points in another does not establish which changed more proportionally or economically. For strictly positive illustrative levels, percentage changes can answer the proportional question. They still do not automatically represent futures account returns, particularly when observations cross a contract switch.

Negative or zero starting levels make ratio-based interpretations more difficult. A price difference remains an arithmetic difference, whereas division by a near-zero denominator can dominate a ranking. The measurement contract should state admissible inputs and what happens when its assumptions fail. Quietly replacing difficult observations with zero changes the ranking universe and may change unrelated instruments' labels.

Align clocks before comparing labels

A relative ranking needs observations available at a common decision cutoff, not merely rows carrying the same calendar date. If one market's final value becomes available later, using it in an earlier comparison introduces information that was unavailable then. An own-history calculation has the same timing responsibility, although it does not require synchronizing several instruments to construct a rank.

For an illustrative two-observation delay, let a computed label at observation 10 first appear as the delayed label at observation 12. The delay belongs to the ordered data sequence, not automatically to two calendar days. Apply the same convention to every compared rule, and keep unavailable labels distinct from deliberately neutral labels so counts remain interpretable.

Read a comparison without overclaiming

Useful descriptive questions include how often labels agree, how long each state persists and how often states change. If two rules agree on 18 of 24 jointly eligible observations, agreement is 75%. That number measures classification similarity only. It does not establish predictive accuracy, investment suitability or a strategy's ability to withstand transaction costs and funding demands.

A clear comparison therefore names the universe, input measure, lookback, tie convention and timing delay before presenting counts. It also states whether its examples describe generic reference levels or identified futures contracts. Readers can then distinguish a different economic question from an implementation discrepancy, instead of assuming that every use of momentum must produce the same answer.

Source record

1. Moskowitz, Ooi and Pedersen: Time Series Momentum
2. Kenneth French Data Library: Daily Momentum Factor Construction