SignalTrace

Style Factor Rotation

John Montgomery4 min read

Style Factor Rotation

Style factor rotation is a strategy that switches a portfolio’s exposure among style-factor models — value, momentum, quality, low-volatility, and multi-factor blends — depending on the currently detected market regime, on the premise that factor risk premia are cyclical and that the factor which leads in a bull market is not the factor that leads in a bear or sideways market. It is a concrete sub-type of Factor Timing: the rotation rule is a timing rule, with a regime label standing in for the more usual valuation or trend signal. In the Markov-model context the regime label comes from a fitted HMM, which is why this strategy sits in the vault’s regime-switching cluster.

Wang Lin Mikhelson 2020 (“Regime-Switching Factor Investing with Hidden Markov Models,” Journal of Risk and Financial Management) is the vault’s worked example. A three-state Gaussian HMM, trained on S&P 500 ETF daily return and volatility and retrained daily on a rolling 2707-day window, infers a bull / bear / sideways state; a Kolmogorov-Smirnov confidence rule then selects whichever of six style-factor models (Fama-French 3-factor, a modified Fama-French, Carhart 4-factor, a value model, an AQR-style model, and the plain S&P 500 ETF) historically performed best in the inferred regime. The factor models are built and backtested on QuantConnect. The reported result is a roughly 2.0 out-of-sample Sharpe ratio over a Sep 2017-Apr 2020 test window, beating the individual factor models and the index.

That headline number must be read with care, and the vault grades it weak. Four problems undercut it. First, no transaction costs and no slippage are modelled, yet a strategy that retrains an HMM daily and switches among six factor portfolios on a KS-confidence trigger is turnover-heavy — the after-cost Sharpe is unknown and could be materially lower. Second, the out-of-sample window is short (~2.5 years) and crash-dominated: it ends in the COVID-19 collapse of early 2020, so the reported Sharpe is heavily influenced by the regime model’s behaviour through one extreme event rather than by performance across a representative cycle. Third, the choice of six factor models and three states is itself an in-sample design decision — a Data-Snooping Bias surface that the out-of-sample Sharpe does not account for. Fourth, and most fundamentally, the strategy is an instance of factor timing, and the broad literature — Cliff Asness’s “siren song” critique above all — finds factor-timing signals weak, easily overfit, and largely collinear with the value factor; an HMM regime label does not exempt a strategy from that skepticism, it just relocates the timing signal.

The honest verdict is that style factor rotation, as evidenced in this vault, is a cost-free in-sample-tuned backtest claim, not demonstrated tradeable alpha. Like all regime-conditional strategies its real-world profitability depends on costs, turnover and real-time classification accuracy — none of which Wang, Lin and Mikhelson model — and it additionally inherits the structural skepticism that attaches to factor timing as a category. It is best understood as Tactical Asset Allocation applied to factor sleeves: a plausible idea with a clean backtest, no out-of-sample after-cost evidence, no independent replication, and a known prior that the underlying timing edge is fragile.

Style Factor Rotation [part-of] Factor Timing Wang Lin Mikhelson 2020 [tests_strategy] Style Factor Rotation Style Factor Rotation [relates] Hidden Markov Model Regime Detection Style Factor Rotation [relates] Tactical Asset Allocation

Connections

Sources

Graph View

Backlinks