Drawdown, volatility and the limits of historical risk

Every risk measure covered elsewhere in this section — VaR, CVaR, Sharpe, Sortino — is computed from a finite window of historical returns, and every one of them inherits the same underlying limitation: the past window did not contain everything that can happen. This article looks at drawdown and volatility specifically, because they are the two measures most often treated as settled facts about a strategy rather than as estimates with their own margin of error, and at where a purely historical approach to risk runs out.

Drawdown, precisely

A drawdown at time $t$ is the decline from the running peak of cumulative wealth up to $t$ :

DD_t = \frac{W_t - \max_{s \le t} W_s}{\max_{s \le t} W_s}

where $W_t$ is portfolio value at time $t$ . Maximum drawdown is the most negative value $DD_t$ takes across the sample. Drawdown analysis reports three related quantities: the maximum drawdown itself, the average drawdown across all peak-to-trough episodes in the sample (not just the worst one), and the duration — how long the portfolio spent below its previous peak before recovering.

A worked example. A portfolio’s wealth index moves $100 \to 118 \to 96 \to 105 \to 130$ across five periods. The peak before the decline to 96 was 118, so that drawdown is $(96-118)/118 = -18.6\%$ . The portfolio does not exceed its prior peak of 118 again until it reaches 130, so if this took, say, 14 periods, the drawdown duration for that episode is 14 periods — the portfolio was “underwater” the entire time, even during the recovery from 105 back past 118.

Drawdown’s chief virtue is that it is intuitive in a way volatility is not: an investor feels a 30% decline in account value directly, in a way they do not feel “18% annualised standard deviation.” Its chief limitation is sample dependence — a five-year backtest that happened not to include a genuine crisis will report a maximum drawdown that understates what the strategy would show over a longer or differently timed history. The Calmar ratio, which divides return by maximum drawdown, inherits this same fragility, as covered in risk and return metrics that matter.

Volatility: an estimate, not a constant

Volatility is usually reported as a single annualised number, which invites treating it as a fixed property of an asset. It is not. Realised volatility varies through time — periods of calm and periods of turbulence cluster, a well-documented empirical pattern — and the standard annualisation $\sigma_d\sqrt{252}$ silently assumes returns are independent and identically distributed across the sample. When volatility clustering is present, that assumption understates the risk of experiencing an unusually turbulent stretch, because the scaling treats every day as drawn from the same distribution rather than from a distribution that shifts with the recent past.

A worked example. Suppose a security’s daily returns have a full-sample standard deviation of 1.2%, giving an annualised volatility of $1.2\% \times \sqrt{252} \approx 19.0\%$ . If the same security’s most recent 20 trading days show a standard deviation of 2.1% — a genuine regime shift, not noise — the annualised figure implied by that recent window is $2.1\% \times \sqrt{252} \approx 33.3\%$ , nearly double the full-sample estimate. A risk report using the full-sample number would materially understate current risk during exactly the period when that understatement matters most.

This is why the covariance matrix feeding portfolio construction is not simply “the sample covariance”: an EWMA estimator (with a default decay of 0.94) weights recent observations more heavily than distant ones specifically to track this kind of drift, and shrinkage estimators such as Ledoit-Wolf shrinkage address a different problem — noisy, unstable estimates when the number of assets is large relative to the sample length — rather than time-variation itself. Choosing among sample, EWMA, shrinkage and denoised covariance estimators is a decision about which kind of estimation error a researcher is more worried about, not a decision with one universally correct answer.

Where purely historical risk measures run out

Historical VaR, drawdown statistics and sample covariance all share a structural limit: they can only describe what is in the sample. A historical VaR computed from five years that happened to avoid a genuine tail event will report a smaller loss threshold than a longer sample would, not because the risk was actually smaller but because the sample never realised it.

Extreme value theory exists specifically to address this gap for the tail. Rather than relying on the empirical quantile of observed returns, it fits a generalised Pareto distribution to the exceedances beyond a chosen threshold — the losses that are already unusually large — and uses that fitted distribution to extrapolate into the tail beyond what the sample has actually shown. This is a genuinely different exercise from historical VaR: it is modelling the shape of the tail rather than reading off an empirical quantile, which lets it say something about losses more extreme than any observed in the sample, at the cost of depending on how well the generalised Pareto distribution actually fits that tail.

The broader point, which applies to drawdown, volatility, VaR and covariance alike, is that a purely historical risk measure describes the regime the sample was drawn from. It has no mechanism for anticipating a regime the sample never saw — a genuinely new kind of correlation breakdown, a liquidity event with no precedent in the loaded data, a structural change in the instruments themselves. This is precisely the gap that scenario and stress testing is built to probe from a different angle, by specifying moves and replaying episodes rather than reading statistics off ordinary historical returns; see scenario and stress testing your allocation for how historical replay and hypothetical shocks complement the measures covered here.

Putting it together

A researcher assessing an allocation’s risk should read maximum drawdown alongside average drawdown and duration, not in isolation; check whether recent-window volatility diverges materially from the full-sample figure; and treat any VaR or CVaR figure as conditional on the sample it came from, using extreme value theory or historical replay of a genuine crisis episode as a check on what the ordinary sample-based numbers might be missing. None of these steps produces a guarantee — they produce a more complete picture of what is known and what is not, which is the most a historical risk measure can honestly offer.

Key takeaways

Maximum drawdown depends on the single worst episode in the sample; average drawdown and duration describe the more typical experience of holding the strategy and should be read alongside it.
Annualised volatility assumes returns are independent and identically distributed, and understates current risk whenever volatility clustering means the recent window is more turbulent than the full-sample average.
Covariance estimators — sample, EWMA, Ledoit-Wolf shrinkage, denoised — each trade off differently between tracking recent changes and avoiding noise, and the choice matters more as the asset count grows.
Extreme value theory extrapolates tail risk beyond the observed sample using a fitted generalised Pareto distribution, which can say more about deep-tail losses than an empirical quantile, at the cost of depending on a distributional assumption.
Every measure in this article describes the historical sample it was computed from; none of it is a forecast, and none of it is investment advice.