Residual Momentum

Imagine ranking swimmers by how fast they finished a race, when some swam with a strong current behind them and others swam against it. The fastest raw times are dominated by who got the current, not who's the best swimmer. To rank actual swimming ability you'd subtract out each lane's current first, then rank what's left. Ordinary price momentum has the same flaw: a stock that rallied 40% might just be a high-beta name riding a market that rallied 25%, not a stock with genuine idiosyncratic strength. Residual momentum subtracts the "current" — the return explained by the market and other common factors — and ranks stocks on what's left over.

Stripping the current out

For each stock $i$, run a rolling regression of its daily or weekly return $r_{i,t}$ on the market and a small set of factors:

$r_{i,t} = \alpha_i + \beta_i^{MKT} r_{t}^{MKT} + \beta_i^{SMB} r_t^{SMB} + \beta_i^{HML} r_t^{HML} + \varepsilon_{i,t}$

In words: today's return is some baseline drift $\alpha_i$, plus how much the stock moves per unit the market moves times how much the market actually moved, plus the same for size and value factors, plus $\varepsilon_{i,t}$ — everything left over that the factors don't explain. Sum or compound the residuals $\varepsilon_{i,t}$ over the standard momentum window (typically 12 months, skipping the most recent month), and you get a cumulative residual return: the stock's factor-adjusted performance, stripped of the current.

Rank stocks each month on cumulative residual return, go long the top decile and short the bottom decile — the same construction as ordinary cross-sectional momentum, just applied to $\varepsilon$ instead of raw return.

Worked example 1. A stock returns 3% in a month while the market returns 2% and the stock's estimated market beta is 1.4. The regression predicts the stock "should" have returned $1.4 \times 2\% = 2.8\%$ from market exposure alone. The residual is $3\% - 2.8\% = 0.2\%$ — almost nothing. A second stock returns 3% in the same month with a beta of only 0.5, so its factor-predicted return is $0.5 \times 2\% = 1.0\%$, leaving a residual of $3\% - 1.0\% = 2.0\%$. Both stocks show identical raw 3% returns, but the second one did ten times more of its own work — ordinary momentum treats them as tied; residual momentum ranks the second one far above the first.

Drag the correlation toward zero on this scatter — that is the target shape of residual returns against the market after the factor regression is done correctly: whatever comovement existed should be scrubbed out, leaving a cloud with no visible slope, so that residual momentum is genuinely ranking idiosyncratic behavior rather than a leveraged bet on market direction.

Why bother: the crash-risk motivation

The main reason to build the residual version isn't extra return, it's a smoother ride. Ordinary momentum's short leg tends to fill up with high-beta losers — stocks that crashed because they're leveraged and cyclical, not because of anything company-specific. When the market snaps back sharply (as it did in 2009), that short leg is exactly the wrong thing to be short, and raw momentum can lose double digits in a matter of weeks (see Momentum Crashes). A residual-momentum short leg holds fewer high-beta names by construction, because part of what made them crash — the beta itself — has already been subtracted out before ranking.

Worked example 2. Take a hypothetical momentum crash month where the market rallies 12% off a bottom. A raw-momentum short book, loaded with high-beta losers averaging $\beta = 1.8$, would be expected to rally about $1.8 \times 12\% = 21.6\%$ against the short position from market exposure alone — a brutal squeeze. A residual-momentum short book built from stocks selected on idiosyncratic weakness tends to run closer to market-neutral, say average $\beta = 1.0$, predicting only a $1.0 \times 12\% = 12\%$ move against it — still painful, but roughly half the damage, purely from having stripped systematic exposure out of the ranking before selecting names.

What this means in practice

Residual momentum needs an extra moving part — a rolling factor regression, re-estimated on some window, feeding into every stock's ranking every month — which is more machinery, more parameters to get wrong, and more sensitivity to which factors you choose to strip out. It's a reasonable default when a book already runs factor-neutral and momentum is one signal among several; it's overkill for a standalone long-only tilt where a simple 12-1 sort is easier to explain and audit.