Rolling Performance Windows and Stability

Report a single Sharpe ratio for a five-year backtest and you get one number that averages over everything that happened in those five years — including a possible regime where the strategy stopped working entirely, as long as it's outweighed by strong performance elsewhere in the sample. A rolling performance window fixes this by recomputing the metric — Sharpe, drawdown, hit rate, whatever's relevant — over a moving slice of the data (say, a trailing 12-month window, stepped forward one month at a time) and plotting how that metric evolves through time, rather than collapsing everything into one static summary.

What a rolling window reveals that a single number hides

A strategy with a strong overall Sharpe of 1.2 could be: (a) a consistently good strategy earning roughly 1.2 in every rolling window, which is what a stable, trustworthy edge looks like, or (b) a strategy that earned a Sharpe of 3.0 for its first three years and then a Sharpe near zero or negative for the last two, averaging out to 1.2 overall. Those are completely different situations for a live-trading decision — case (b) is a strategy whose edge has likely decayed and shouldn't be trusted to continue at the historical average going forward, while case (a) has actual evidence of persistence. A single full-sample number cannot distinguish these; a rolling window plot makes the difference immediately visible.

What to watch for in a rolling plot

Three patterns matter most: a steady downward trend in rolling Sharpe suggests genuine alpha decay, as more participants find and trade away the same edge; a sudden regime break — good performance abruptly replaced by flat or negative performance — suggests something structural changed (a market regime shift, a venue rule change, a crowding event) rather than gradual decay; and high variance in the rolling metric with no clear trend suggests the strategy's Sharpe estimate is simply noisy and the full-sample number, while not necessarily wrong, has wide uncertainty around it that a single point estimate doesn't convey.

Worked example: two strategies with the same full-sample Sharpe

Both strategies show a full-sample annualized Sharpe of 1.0 over four years. Strategy M's rolling 12-month Sharpe, stepped quarterly, reads: 1.1, 0.9, 1.0, 1.2, 0.8, 1.1, 0.9, 1.0, 1.1 — bouncing narrowly around 1.0 the whole time. Strategy N's rolling 12-month Sharpe reads: 2.4, 2.1, 1.8, 1.0, 0.3, −0.2, −0.4, −0.1, 0.1 — clearly declining from strong early performance to flat-to-negative recent performance. Both average to roughly the same full-sample number, but only Strategy N shows a warning sign that its edge, whatever it was, is no longer present in the most recent data — exactly the information a decision to allocate capital going forward needs.

What this means in practice

Any backtest report worth trusting should include a rolling metric plot alongside the full-sample summary statistics, not instead of them. Before allocating capital to a strategy based on a strong full-sample Sharpe, check whether that Sharpe is a stable average or the residue of strong performance that has since faded — the two situations warrant very different confidence about what happens next.