No Factor Works Forever — And That's the Point
From 2017 to 2019, momentum was almost embarrassingly reliable. Then came March 2020, and crowded momentum positions unwound so fast that the factor posted some of its worst monthly returns in decades. This wasn't a data anomaly or a model bug — it was a regime shift, and regime shifts are precisely what single-factor models are structurally unprepared to handle.
A factor is, at its core, a pattern extracted from historical data. Value investing rewards buying cheap assets because, over long periods, the market has tended to reprice them upward. Momentum works because trends persist — until they don't. Quality strategies succeed in downturns because defensive earnings hold up — except when a liquidity crisis forces indiscriminate selling and nothing holds up at all. Each factor encodes assumptions about how the market behaves, and those assumptions are only valid in the specific economic conditions that gave rise to the pattern in the first place.
Conditions Change. Factors Don't Adapt on Their Own.
The practical problem isn't that any given factor is wrong — it's that factors are stationary models applied to a non-stationary world. Rising rate environments tend to punish growth and rehabilitate value. Low volatility periods let momentum compounds quietly. Credit stress scrambles the quality premium by repricing risk across the board in ways that backward-looking factor scores simply cannot anticipate. Backtests that span multiple decades look robust precisely because they average over these regimes, hiding the stretches where a given factor was actively destroying capital.
A factor that worked in the last regime isn't predictive — it's a description of a world that may no longer exist.
This is where multi-factor approaches, and particularly machine learning-based scoring systems, have a structural edge. Rather than committing to a single thesis about what the market rewards, a well-designed ML model can weight factors dynamically — giving momentum more influence when trend conditions are favorable, leaning harder on fundamentals when volatility is elevated, and dampening signals that have historically degraded in the current macro environment. The output isn't a rigid factor bet; it's a continuously recalibrated view of which signals are actually informative right now. This kind of adaptive scoring is exactly what institutional quant desks have been building for years, and what platforms like AIStockAnalytics are designed to make available beyond that closed circle.
Robustness Is a Feature, Not a Compromise
There's a temptation, especially after a strong run for any single factor, to treat that factor as validated. It's the wrong conclusion. Strong recent performance in a given regime is evidence that a factor is regime-dependent — which means the question you should be asking is what breaks it, not how much further it can run. A model that survives across regimes probably won't post the highest return in any single one of them, but it will consistently avoid the catastrophic drawdowns that come when a crowded factor thesis unravels. Over a full market cycle, that consistency matters far more than peak-cycle outperformance.