Cross-Sectional Momentum as a Fifth Factor: Improving R² Beyond the Standard Four

Abstract

The four-factor regression — market, sector, size, value — explains 78% of daily variance on the illustrative $2M industrials micro-cap. Adding 12-month cross-sectional momentum lifts cumulative R² to 0.85 on certain micro-cap cohorts — 7 points of additional explained variance, the largest single-factor gain available after the base four. The catch: the momentum loading fails the stability screen the other four pass. It flips sign across rolling windows, decays within months, and reverses at turning points. This note documents the construction, the gain, and why the fifth factor is usually excluded from the proxy hedge despite the higher R².

1. Baseline: The Four-Factor R² of 0.78

OLS on 252 trading days of daily total returns. Independent variables: Russell 2000 excess returns, small-cap industrials sector index, SMB, HML. Retention threshold t > 2.0, re-estimated monthly.

On the illustrative position, SMB drops out (t = 1.1). The retained loadings:

• Market (Russell 2000): β = 1.35, t = 6.1, marginal R² = 0.62
• Sector (Industrials SC): β = 0.48, t = 3.4, marginal R² = 0.11
• Value (HML): β = 0.22, t = 2.3, marginal R² = 0.05
• Size (SMB): β = 0.09, t = 1.1 — dropped
• Cumulative R²: 0.78

This is the operating baseline. Stable across rolling windows, all retained t-stats above 2.0, loadings drifting less than 0.10 quarter to quarter.

2. Momentum Construction: What the Fifth Factor Measures

Cross-sectional momentum ranks stocks by trailing 12-month total return, skipping the most recent month, and goes long the top decile and short the bottom decile. The UMD series is the daily return of that long-short portfolio.

As a regressor, the momentum loading answers a narrow question: after controlling for market, sector, size, and value, does the micro-cap co-move with recent winners or recent losers?

Data requirements are strict:

• Formation window: 12-month return ending one month before the regression date. Shorter windows add noise without explanatory power on daily micro-cap returns.
• Skip month: exclude the most recent 30 calendar days to avoid bid-ask bounce contaminating the signal.
• Universe: the sort must include small- and micro-caps. A large-cap-only UMD understates sub-$150M momentum exposure by roughly half.

Same OLS, same 252-day window, same t > 2.0 retention rule.

3. The Gain: R² From 0.78 to 0.85

Adding UMD produces the following on the momentum-tilted cohort — micro-caps in the top or bottom performance tercile of their sector over the estimation window:

The 7-point gain is real in-sample. Momentum’s marginal R² of 0.07 exceeds value’s 0.05 and is the largest increment of any fifth-factor candidate tested, including profitability (+0.02) and investment (+0.01).

Two qualifications. First, the gain concentrates in trending cohorts. On the full micro-cap sample pooled, the average lift is 0.03, not 0.07. Second, market and sector loadings compress slightly when momentum enters (1.35 → 1.28, 0.48 → 0.44), shifting dollar-beta 5–8%. That difference is the tracking decision.

4. Instability: Why the Loading Does Not Hold

The momentum loading of 0.31 clears t > 2.0 in the estimation window. It does not survive the stability screen the other factors pass: drift under 0.15 across three consecutive rolling windows. Six monthly re-estimations on the same name:

• Window 1: β_mom = 0.31, t = 2.8
• Window 2: β_mom = 0.24, t = 2.2
• Window 3: β_mom = 0.11, t = 1.2 (fails retention)
• Window 4: β_mom = −0.08, t = −0.7
• Window 5: β_mom = −0.19, t = −1.6
• Window 6: β_mom = −0.27, t = −2.4 (significant, opposite sign)

Over six months the loading travels from +0.31 to −0.27, ending significant in the reverse direction. The base factors over the same span: market beta 1.28–1.41, sector 0.42–0.50, value 0.17–0.24. None breach the 0.15 drift trigger. Momentum breaches it twice.

The cause is structural. Cross-sectional momentum is a ranking, not a characteristic. A micro-cap that outperformed for eleven months and then stalls drops out of the winner decile within one rebalance. Its return stream detaches from UMD abruptly. Market beta reflects economic linkage; momentum loading reflects temporary leaderboard position.

5. Reversal: The Cost of Holding the Exposure

Instability understates the problem. Momentum does not merely decay toward zero — it reverses. At market turning points, prior winners underperform and prior losers outperform over 1–6 months. The 2009 and 2020 reversals each erased more than 20% from UMD in under eight weeks.

For a proxy hedge built on a positive momentum loading, a reversal produces correlated failure on both legs. The micro-cap, as a recent winner, sells off harder than its market beta predicts. The momentum proxy leg sells off simultaneously. The hedge adds loss instead of offsetting it. In a simulated reversal month on the illustrative position, the four-factor proxy tracks within 1.8% of the factor-driven move; the five-factor proxy misses by 4.3% in the wrong direction.

Shortening the rebalance cycle does not fix this. Monthly re-estimation still leaves a 30-day window of stale loading against a reversed regime. Weekly windows produce t-stats that oscillate around 2.0 on sampling noise alone, triggering false hedge-ratio changes.

6. The Tracking Tradeoff: When 0.85 Is Worse Than 0.78

The decision rule: higher in-sample R² justifies inclusion only if it improves realized out-of-sample tracking. Walk-forward — estimate on trailing 252 days, track over the next 21, roll monthly:

• Four-factor proxy: mean absolute tracking error 2.1% per month, worst month 4.6%.
• Five-factor proxy with momentum: mean absolute tracking error 1.9% per month, worst month 7.8%.

Average tracking improves by 0.2 points. Tail tracking deteriorates by 3.2 points. The extra 7% of explained variance buys a small improvement in normal conditions and a large degradation in the months when the hedge carries the load.

The operating policy follows directly:

1. Estimate five factors for diagnostics. If momentum clears t > 2.0 and adds over 0.03 to R², flag the position as momentum-exposed.
2. Build the proxy off four factors. Exclude momentum unless the loading holds sign and t > 2.0 across three consecutive windows — fewer than one in five pass.
3. Treat a significant loading as a warning. A large positive β_mom signals reversal exposure. Tighten stops and reduce short-call overwrite; do not tilt the basket.

7. Bottom Line

Cross-sectional momentum raises the factor regression R² from 0.78 to 0.85 on trending micro-cap cohorts. The marginal power exceeds every other fifth-factor candidate. But the loading is unstable, reverses sign at turning points, and converts a 0.2-point average tracking gain into a 3.2-point tail-tracking loss. For proxy hedging — where the hedge must perform in stress — the four-factor model at 0.78 remains the specification the basket is built on.