Testing Regime-Based Factor Rotation
Factor returns vary across market conditions. The investable question is narrower: can a rule identify those conditions before the relevant returns arrive, and improve a portfolio after trading costs? A chart that colours past periods by their eventual regime does not answer that question.
Start with one decision rule
Consider a deliberately simple hypothesis: when trailing market volatility is high, hold less of a value portfolio and more of a low-volatility portfolio. The exact threshold and weights are research choices, so record them before examining the evaluation period. For an illustration, use the previous 63 trading days of market returns to estimate volatility at each month end. At the next trading session, allocate 25% to value and 75% to low volatility if that estimate exceeds a threshold fixed on training data; otherwise reverse the weights.
Those percentages are an illustrative rule, not a recommended allocation or a measured result. The point is that every ingredient is observable when the trade is placed. A regime label assigned after a market crash is not.
| At the decision date | Allowed input | Unavailable input |
|---|---|---|
| Month end | Returns published through that close | Next month’s factor returns |
| Next session | Portfolio weights from the frozen rule | A revised macro series released later |
| Later evaluation | Realised return and turnover | A new threshold chosen after seeing the outcome |
A yield-curve or inflation signal can be tested the same way, but use the value and release date available at the time. Revised economic data can make a historical rule look informed when it was not.
Give rotation a fair opponent
Compare the rule with a fixed 50/50 value and low-volatility allocation on the same dates, assets, and return series. Rebalance both on the same schedule. Charge each for its actual turnover, spreads, fees, and market impact. If the rotation rule trades more often, its additional cost belongs in the comparison.
Use a rolling evaluation: set the volatility threshold from an initial training window, freeze it, score the next window, and move forward. Do not tune the threshold on the final evaluation months. Keep a log of every threshold, factor definition, and rebalance frequency tried. If dozens of variants were examined, a single winning Sharpe ratio is selected evidence, not an independent test.
Report at least net return, volatility, drawdown, turnover, and the difference from the fixed portfolio. Show the time series, not only an aggregate. A gain concentrated in one short period is a different claim from a gain that persists across market conditions. Also report where the rule did worse. A model that mostly identifies regimes after they have changed may be descriptive but not useful for allocation.
Add complexity only when it earns its place
A hidden Markov model can estimate changing states, but it also adds state count, initialization, fitting window, and state-label choices. Its inferred states can shift when new data arrives. Test its real-time assignments, then ask whether it improves on the fixed rule and the simple volatility threshold after costs. If it does not, the simpler baseline is the more honest result.
This is a testing protocol, not evidence that regime rotation works. It gives the hypothesis a way to fail before anyone treats a backtest as a strategy. For more on trial selection and trading assumptions, see Five Backtesting Pitfalls That Fake Your Sharpe and the reading list on multiple testing.