Factor Momentum

A river doesn't just have individual eddies of water moving unpredictably — the whole river has a current, and that current itself can speed up or slow down over a season as rainfall changes. Individual stock momentum is like watching one eddy; factor momentum is noticing that the river's overall current — a factor like value, quality, or low-volatility — also tends to keep flowing in the same direction it's been flowing, for a while, before it changes.

Trend-following, one level up

Instead of ranking individual stocks on their own past return, factor momentum ranks entire factors on their own recent performance, then tilts toward the factors that have recently worked:

$w_{f,t} = w_0 + \gamma \cdot r_{f, t-1}$

In words: this month's weight on factor $f$ (say, quality) is a baseline weight plus a sensitivity $\gamma$ times how that factor performed last month — if quality just had a strong month, lean into it a bit more next month; if it had a weak month, lean away. Gupta and Kelly (2019) tested this systematically across dozens of published factors and found it worked about as well as, and largely independently of, ordinary stock-level momentum.

Worked example 1. Suppose the quality factor returned $+1.8\%$ last month and $\gamma = 0.3$ with a baseline weight $w_0 = 10\%$ of the book. This month's tilt: $w = 10\% + 0.3 \times 1.8\% = 10.54\%$ — a small nudge upward. Meanwhile the size factor lost $-1.2\%$ last month, so its weight becomes $w = 10\% + 0.3 \times (-1.2\%) = 9.64\%$. Neither move is large individually, but applied across a dozen factors every month, and compounded over years, Gupta and Kelly found factor momentum's information ratio was comparable to individual-stock momentum's, using far fewer positions and much lower turnover per unit of edge, because you're trading a handful of factor tilts instead of thousands of stock rankings.

The surprising link to stock momentum

Worked example 2. Ehsani and Linnainmaa (2022) asked whether ordinary stock momentum is actually, underneath, mostly factor momentum in disguise. Decompose a stock's past return into its exposure to several factors times each factor's own past performance, plus an idiosyncratic piece — structurally similar to the Residual Momentum decomposition, but applied backward to explain why stock momentum works rather than to strip it out. They found that a large share of stock momentum's profitability could be replicated by a portfolio that only knew each stock's factor loadings (fixed, not time-varying) combined with recent factor returns — no stock-specific past-return information required at all. Concretely: if a stock has a value loading of 0.8 and a quality loading of 0.6, and value returned $+2\%$ while quality returned $+1.5\%$ last month, the factor-momentum-implied prediction for that stock is roughly $0.8 \times 2\% + 0.6 \times 1.5\% = 2.5\%$ — a number derived entirely from factor trends, yet it lines up closely with what standard stock-level momentum would have predicted for the same stock using its own raw past return.

What this means in practice

If factor momentum is doing much of the work, then the classic explanation for stock momentum — investors underreacting to company-specific news — needs revisiting, because the effect would instead be substantially about investors underreacting to macro or style-level trends. Practically, a desk that already runs several factor tilts can add a lightweight factor-momentum overlay — tilting existing factor weights based on their own recent performance — as a cheap complement to (not a replacement for) stock-level momentum, since the two draw on overlapping but not identical information.