Early-Warning Models for Drawdowns

By the time a large drawdown is visible on an equity curve, it's already too late to avoid it — the goal of an early-warning model is to flag rising risk of one using signals that tend to move before the drawdown itself does, not to predict the drawdown's exact size or timing. That's a fundamentally different, more modest target than forecasting returns: the model doesn't need to know a crash is coming next Tuesday, only that current conditions resemble the conditions that have historically preceded bad stretches.

An analogy: a hospital's early-warning score

Hospitals use early-warning scores that combine heart rate, blood pressure, breathing rate, and temperature into a single number that flags a patient at risk of deteriorating — not a diagnosis of what's wrong, just a signal that "something in this combination of readings looks like the pattern that precedes trouble." A drawdown early-warning model does the same job for a portfolio: it combines several leading indicators — volatility clustering, correlation spikes across previously uncorrelated assets, deteriorating liquidity — into a single score that flags elevated risk, without claiming to know exactly what will go wrong or when.

What goes into the score

The model is typically a classifier trained to predict "will a drawdown beyond some threshold begin within the next $k$ days" using a feature set built entirely from information available today: realized volatility over recent windows, the average pairwise correlation across the portfolio's holdings (correlations tend to spike toward 1 right before and during systemic stress, a phenomenon sometimes measured via a turbulence statistic), bid-ask spreads or other liquidity proxies, and sometimes macro indicators. The label is constructed after the fact by looking forward from each historical date and marking whether a qualifying drawdown began within the window — a supervised learning setup built entirely on leading features and a forward-looking label, which must never be allowed to leak backward into the features.

Worked example: reading a turbulence-style score

Define a simple two-feature score: 20-day realized volatility percentile and average pairwise correlation percentile (each 0–100, based on trailing five-year history), averaged together. In a calm period, volatility sits at the 30th percentile and correlation at the 25th, giving a score of about 28 — low, unremarkable. Two weeks before a known historical drawdown episode, volatility has risen to the 70th percentile and correlation has jumped to the 85th percentile (previously uncorrelated sector exposures starting to move together, a classic pre-stress signature) — a score of about 78. Backtesting this simple score across ten years shows that scores above roughly 70 preceded a >8% drawdown within the following month about 40% of the time, versus roughly a 5% base rate for that outcome in any given month — a meaningful lift, though far from a certainty, which is the expected character of an early-warning signal rather than a forecast.

What this means in practice

Early-warning scores feed derisking triggers, hedge sizing, and position-limit adjustments — a portfolio manager might cut gross exposure or add tail hedges when the score crosses a threshold, without waiting for a drawdown to already be underway. They are deliberately probabilistic, not deterministic: a well-calibrated score should raise the odds of a bad stretch meaningfully above the base rate, and it will still be wrong plenty of the time, exactly as a smoke detector goes off occasionally with no fire.