What caught my attention here is not the regression framework itself but the post-2021 underperformance. The logistic regression strategy delivered 24.61% annualised over the full sample, then stopped working precisely when algorithmic penetration and passive flows reached critical mass. That is not a model failure. That is a signal about market structure.
From an allocator perspective, the same dynamic shows up in TAA consensus data: momentum and relative strength signals that worked reliably through 2020 have produced weaker differentiation since. The models did not get worse. The market they were designed for changed around them. In my view, the more important finding in this piece is not that regression still works, it is that the conditions under which any systematic signal works are themselves regime-dependent.
That is the part most practitioners underweight when evaluating backtests.
Spot on. The performance of any systematic strategy is regime dependent. We have seen a regime shift after the COVID, not only in equities but also in volatility markets. The growing participation of retail options traders, particularly in short-dated options, has changed the market dynamics considerably.
A recent example is the spot/vol correlation in the semiconductor ETF, which was negative but has turned positive and spilled over into broader equity indices. Strategies built around the negative correlation struggled.
QVR, a well-respected HF, is another example. Their dispersion strategy reportedly suffered losses as the correlations changed.
No model can reliably predict regime shift. The key is to limit the damage, and adapt quickly.
The QVR example is the clearest illustration of the point. A strategy built on a structural relationship that held for years, then one regime shift, and the edge disappears before the model knows what happened. In my experience, the speed of adaptation matters as much as the ability to adapt.
The window between "something is different" and "the damage is done" is often too short for most systematic processes to respond cleanly.
To me, that is why regime detection is ultimately harder than strategy design.
What caught my attention here is not the regression framework itself but the post-2021 underperformance. The logistic regression strategy delivered 24.61% annualised over the full sample, then stopped working precisely when algorithmic penetration and passive flows reached critical mass. That is not a model failure. That is a signal about market structure.
From an allocator perspective, the same dynamic shows up in TAA consensus data: momentum and relative strength signals that worked reliably through 2020 have produced weaker differentiation since. The models did not get worse. The market they were designed for changed around them. In my view, the more important finding in this piece is not that regression still works, it is that the conditions under which any systematic signal works are themselves regime-dependent.
That is the part most practitioners underweight when evaluating backtests.
Spot on. The performance of any systematic strategy is regime dependent. We have seen a regime shift after the COVID, not only in equities but also in volatility markets. The growing participation of retail options traders, particularly in short-dated options, has changed the market dynamics considerably.
A recent example is the spot/vol correlation in the semiconductor ETF, which was negative but has turned positive and spilled over into broader equity indices. Strategies built around the negative correlation struggled.
QVR, a well-respected HF, is another example. Their dispersion strategy reportedly suffered losses as the correlations changed.
No model can reliably predict regime shift. The key is to limit the damage, and adapt quickly.
The QVR example is the clearest illustration of the point. A strategy built on a structural relationship that held for years, then one regime shift, and the edge disappears before the model knows what happened. In my experience, the speed of adaptation matters as much as the ability to adapt.
The window between "something is different" and "the damage is done" is often too short for most systematic processes to respond cleanly.
To me, that is why regime detection is ultimately harder than strategy design.