Crisis-Period Carve-Out Testing

A strategy's overall backtest Sharpe ratio is an average across the entire sample period, and averages hide a lot. A strategy that performed brilliantly in calm markets for eight years and then lost half its capital in a three-week crisis can still show a respectable average Sharpe ratio over the full period, because the calm years contribute far more data points than the brief crisis. Crisis-period carve-out testing pulls specific, well-known stress periods — the 2008 financial crisis, the March 2020 COVID crash, the 1998 LTCM-era turmoil, the 2013 taper tantrum — out of the full backtest and evaluates the strategy's performance during each one individually, rather than letting it get averaged away into the overall number.

Why crises specifically, not just "bad periods"

Known crisis periods are useful test cases because they share a specific, well-understood mechanism: liquidity dries up suddenly, correlations across normally-unrelated assets spike, and everyone is trying to reduce risk at the same time. A strategy's ordinary backtest performance says nothing about how it behaves under those specific conditions unless a big historical crisis happens to fall inside the sample, which for many strategies with shorter histories it doesn't. By explicitly carving out these periods and analyzing them separately, a researcher gets a direct look at how the strategy behaves in exactly the kind of environment where forced deleveraging and correlation breakdown are most likely to hurt it, rather than hoping the aggregate statistic happens to capture that risk.

A concrete case

A merger-arbitrage strategy shows an excellent long-run Sharpe ratio when backtested over fifteen years. Carving out just the 2008 crisis period and looking at it in isolation, the strategy shows a sharp, concentrated loss as deal spreads widened dramatically and several pending mergers were called off amid the market turmoil — a pattern completely invisible in the smooth, long-run average. This carve-out doesn't necessarily disqualify the strategy, but it gives a much more honest picture of what a real future crisis might do to it, and it's exactly the kind of information a capital allocator needs before sizing the position.

What this means in practice

Carve-out testing is a targeted complement to full-sample backtesting, not a replacement for it — it answers a specific and important question ("how bad could this get in a real crisis") that an aggregate statistic across the whole sample period is poorly suited to answer on its own, since a handful of extreme weeks get diluted by years of ordinary ones.