Timing Factor Regime Shifts

Executive abstract

Style factors are the workhorses of modern equity investing, yet their leadership is unstable. Value can dominate for years, then momentum, then quality, before a regime turns over in months. This report establishes the rotation problem and the predictor landscape, then examines the variable the evidence elevates above all others.

1The question, and why it is hard

The central question is narrow and practical: which observable variables move before factor leadership rotates, and how far ahead? Rotations are rare, so a useful predictor may look unimpressive on raw accuracy while still carrying information. The raw leadership signal is also noisy: the best-performing factor can flip month to month without a durable regime change.

A credible study therefore needs a disciplined definition of “rotation,” leakage-free validation, and metrics judged against appropriate baselines—AUC versus 0.50 and average precision versus the event base rate.

2The factor universe

The study tracks seven equity style factors over 754 monthly observations from July 1963 through April 2026, sourced from the Kenneth French research data library, with a calibrated proxy for low-volatility—the one style with no French equivalent.

3What counts as a rotation

A rotation requires two independent lenses to agree. First, each factor’s trailing six-month return defines relative strength, and a new leader must hold the top slot for at least three months. Second, a change-point detector on the relative-strength panel must independently mark a structural break within a three-month window. Their intersection defines a rotation event.

The process identifies 30 events, with a median regime length of roughly 10 months and a median absolute leadership swing of roughly 14%.

The detector lands on economically coherent months and finds no durable rotation across 2009–2021, when growth, quality, and momentum led without interruption.

4Can rotations be predicted? The driver ranking

A walk-forward gradient-boosting classifier predicting a three-month-ahead rotation flag earns out-of-sample AUC 0.647 versus 0.50 for a coin flip and beats a logistic baseline by 0.09 AUC. The model’s more useful contribution is the ranking of the inputs that carry lead information.

5Definition, data, and provenance

For each month t, let r₁₂^(f)(t) be the trailing 12-month cumulative return of factor f. Cross-factor momentum dispersion is the cross-sectional population standard deviation of those returns:

log_gross = np.log1p(returns)
cum_12m = np.expm1(log_gross.rolling(12, min_periods=12).sum())
factor_mom_disp = cum_12m.std(axis=1, ddof=0)

The measure is large when one or two factors have run far ahead of the pack and small when factor returns are bunched. It is the factor-internal analogue of a turbulence gauge. Two relatives test distinctness later: factor_mom, the cross-factor mean of the same 12-month returns, and factor_val_spread, the standard deviation of trailing 36-month returns.

The factor returns are real and were independently verified. Recomputing factor_mom_disp directly from the factor-return panel reproduces the delivered predictor column with a maximum absolute difference of 0.0 across all 743 months. The series runs from June 1964 through April 2026, with a mean of 0.108, standard deviation of 0.051, minimum of 0.019, and maximum of 0.376 in February 2001.

6Dispersion through history

Across six decades, dispersion peaks line up with major factor-regime stresses: the 2000–2001 dot-com value/momentum unwind, 1974 stagflation, the post-GFC reset, and the 2021 value comeback.

7The leading-indicator result

The lead-lag profile is computed as corr[disp(t), rotation(t+k)] for forward leads from zero to 12 months. Correlation peaks at k = 5 months, at +0.158:

lead 0: +0.129   1: +0.126   2: +0.134   3: +0.126   4: +0.134
lead 5: +0.158 (peak)   6: +0.128   7: +0.089   ...   12: +0.036

The sign is consistent across every lens: high or rising dispersion precedes rotation. The transforms that survive feature selection are dispersion’s six-month momentum and 12-month z-score, indicating the signal lives in its rising and stretched state rather than the raw absolute level.

8Historical event studies

8.1 Average event study across all 30 rotations

Aligning each rotation at t0 and averaging the dispersion z-score across the [−12, 0] window produces a rising-then-cresting profile that peaks five months ahead.

8.2 Individual case studies

The pattern is strongest before large, crowding-fueled rotations and is honest about its misses. Dispersion spikes before the dot-com unwind and the 2022 value/rates turn, is elevated before several others, and is flat or falling before a minority of events.

Counter-examples are real. The October 1979 rotation occurred with dispersion low and falling, near z = −1.3 at t−5, while September 1987 was flat. Dispersion is a tendency, not a law.

9Predictive strength: quintile lift

Because the power is concentrated in the tail, a quintile sort is more revealing than linear correlation. Sorting months by dispersion and measuring whether a rotation follows within six months produces a 25% base rate.

As a binary signal, top-quintile dispersion produces precision of 0.57 and recall of 0.46. In plain terms, when dispersion is extreme, a regime change follows within half a year well over half the time, versus one chance in four unconditionally.

10Why dispersion leads

1. Stretched, crowded trends carry mean-reversion risk. Dispersion is large when one or two factors have run far ahead. Mature, one-sided trends attract crowding and valuation extension; the wider the gap, the larger the eventual snap-back risk.
2. Wide dispersion signals fragile leadership. When returns are bunched, the leader is robust to small shocks. When dispersion is wide, leadership rests on a thin margin of trend rather than breadth, so a modest macro or sentiment shock can reorder the cross-section.
3. Factor-internal turbulence. Dispersion co-moves moderately with VIX and the value spread but dominates both on every importance lens. It captures risk expressed through style rotation rather than index volatility alone.

The five-month lead is best read as roughly one to two quarters ahead, not as a precise point estimate against a sparse binary target.

11Robustness and distinctness

Regressing factor_mom_disp jointly on factor_mom, factor_val_spread, and vix leaves R² = 0.23—approximately 77% of its variance is orthogonal to all three. The strongest overlap, with the 36-month value spread, is expected, but the two separate in stability: dispersion ranks first at 93%, while the value spread reaches only 5% top-five frequency.

The signal is strongest when markets are calm. In low-VIX months, factor-internal dispersion performs the discriminating work and average-precision lift reaches 3.0×. In high-VIX months, where rotation is already visible, the edge collapses toward the base rate.

The result also survives at the per-factor level: a dispersion transform leads the leadership-change event for every modeled factor with a consistent positive sign. Quality, based on only 20 events and sub-random skill, is the lone exception.

12Limitations and conclusion

1. Small event sample. Thirty rotations in 743 months underpin every estimate; the 2000–2001 cluster is influential, and bootstrap ranks move by roughly ±1.8–2.5 positions.
2. Tail-driven, not linear. Peak cross-correlation is approximately 0.16. The power lives in the top quintile, so dispersion is a regime flag rather than a smooth continuous signal.
3. Lead loosely identified. “Roughly one to two quarters ahead” is more defensible than an exact five-month claim.
4. Feature-family overlap. Dispersion shares roughly 23% of its variance with the factor-momentum, value-spread, and VIX block.
5. One proxy factor and shared construction. Low-volatility is a calibrated proxy, and the rotation labels derive from the same factor panel.
6. Indicator, not strategy. Predictive lift is real but modest, while the tradable timing edge in this dataset remains within noise. A naive rotation overlay passes none of three significance tests and sits near the center of a random-rotation null.