Forecasting Through a Regime Change

A model trained on five years of low-volatility, steadily rising markets learns a relationship between its inputs and future returns that is really a relationship between its inputs and that specific regime. When the regime breaks, a rate-hiking cycle begins, a crisis hits, correlations flip, the model doesn't know it's now operating outside the world it learned from, and it keeps producing confident forecasts using rules that no longer apply. Forecasting through a regime change means having a plan for exactly this moment, not just a model that worked well in the past.

Why this is a genuinely hard problem

The core difficulty is that a regime change is only cleanly identifiable in hindsight. In real time, a researcher sees a handful of unusual data points and has to decide whether that's normal noise (in which case reacting is a mistake, it will just add whipsaw) or the early signature of a real break (in which case not reacting is the mistake). No detector gets this perfectly; the practical goal is a forecast that degrades gracefully rather than confidently, during the ambiguous stretch before a regime change is confirmed.

A practical approach: widen, don't discard

Rather than trying to detect the exact moment of a regime change and switch models, many desks instead build in a mechanism that automatically widens the model's uncertainty, and shrinks its position sizes, whenever recent forecast errors start running larger than the model's own history would predict.

In words: this is an exponentially weighted estimate of the model's own recent forecast-error variance, where each period's error $e_{t-1}$ updates the estimate, and a lower $\lambda$ makes the estimate react faster to a recent run of bad forecasts. When realized errors suddenly get larger, the signature of a regime shift the model hasn't adapted to, $\hat{\sigma}_t^2$ rises quickly, and that rising number is what should drive position sizes down, even before anyone has formally declared a new regime.

Worked example

A signal's rolling forecast-error variance, tracked with $\lambda = 0.94$, has averaged around $0.0004$ for two years. Over the last ten trading days, actual squared errors have averaged $0.0018$, more than four times normal. Updating the exponentially weighted estimate: $\hat{\sigma}_t^2 = 0.94 \times 0.0004 + 0.06 \times 0.0018 \approx 0.000484$, already an 21% rise from a single update, and it will keep climbing if the elevated errors persist. A desk using this number to scale position size (inversely, by $1/\hat{\sigma}_t$) automatically starts trading smaller within days of the model's real-world performance degrading, without needing anyone to manually declare a regime change first.

What this means in practice

This approach trades a small, constant cost (slightly smaller positions during genuinely stable periods, since the estimator is never perfectly zero-noise) for protection against the much larger cost of running full size straight through a regime break. It does not solve regime detection, nothing fully does, but it keeps a bad forecast from doing maximum damage while the regime is being sorted out.