Regime Classification for Risk Scaling

The idea behind regime-based risk scaling is straightforward: identify when the market has shifted into a high-risk state — a crash, a liquidity crunch, an unusually correlated sell-off — and cut position sizes before the worst of the damage happens, then scale back up once conditions normalise. A classifier trained on history can look excellent at this task when evaluated the ordinary way: label every historical day as calm or stressed with the benefit of hindsight, and check how often the model's classification matches. That evaluation quietly assumes the model knows the regime as it happens, which is a much harder problem than knowing it in hindsight — most crises are only cleanly identifiable as crises a few days or weeks after they start.

The lag between "identifiable" and "identified in time"

A regime label built after the fact — say, using a centred rolling window of realised volatility, or a change-point algorithm that reports the change-point date only once enough post-change data has accumulated to be confident — has access to information a live system wouldn't have on the day itself. A classifier trained to reproduce those labels, then tested by feeding it only trailing information as it would see live, typically performs much worse, because it is now solving the genuinely hard problem: distinguishing "today is the start of a real regime shift" from "today is an ordinary volatile day that will revert by next week," using only data up to today.

This gap matters enormously for risk scaling specifically, because the entire value of the classification is in acting before the damage compounds. A classifier that correctly identifies a crash regime three weeks after it started, using only information a live system would have had, has correctly labelled history but delivered close to zero risk-management value — most of the drawdown it was meant to help avoid has already happened by the time it fires.

A worked example

A team builds a two-state hidden Markov model on daily returns, trained to distinguish "calm" and "stressed" regimes, and evaluates it by comparing its filtered state probabilities against a hindsight-labelled history of known stress periods (2008, 2020, 2022). Measured this way, classification accuracy looks strong — the model's stressed-state probability rises sharply during each known crisis.

The team then re-examines when exactly the model's stressed-state probability crossed a risk-scaling threshold (say, 70%) relative to when each crisis actually began, using only the filtered probability as it would have been computed live, day by day, with no future data. In the 2020 case, the model crosses the threshold several trading days after the sharpest part of the initial sell-off — a real, publishable improvement over doing nothing, but a much more modest one than the hindsight-accuracy number implied, because a large fraction of the drawdown occurred in those first few days before the model reacted.

Scaling down risk on this real-time signal rather than the hindsight one is the only honest way to estimate what risk-scaling actually would have saved — and that number is nearly always smaller, sometimes dramatically so, than a hindsight backtest suggests.

Picture two overlapping versions of this distribution — a calm regime with a small standard deviation and a stressed regime with a much larger one. A classifier's real job is separating draws from the two as early as possible using only a handful of recent observations, which is genuinely difficult when the tails of the calm distribution already overlap the body of the stressed one.

Where it breaks in practice

A common shortcut is to use a feature that is itself subtly forward-looking — a "trailing" volatility computed with a window that, due to an implementation detail, includes today's return in a way that leaks same-day information the classifier wouldn't truly have at the moment scaling decisions are made. This produces a classifier that appears to react instantly to regime shifts in backtest and reacts with real, unavoidable lag once deployed live.