Rolling Sharpe With Confidence Bands

Plotting a strategy's Sharpe ratio calculated on, say, a trailing 60-day window and updating it every day gives a rolling Sharpe: a line that shows whether performance quality is improving, stable, or fading over time, rather than one static number computed once over the whole history. The trouble is that a Sharpe ratio computed from only 60 days of daily returns is a genuinely noisy estimate — it will bounce around meaningfully from window to window even for a strategy whose true, underlying quality hasn't changed one bit. Looking at a rolling Sharpe line without accounting for that noise invites a manager to read meaning into wiggles that are just statistical jitter.

Adding confidence bands around the rolling line fixes this. The band is built from the standard error of a Sharpe ratio estimate over that window length — which itself depends mainly on the number of observations and how volatile the strategy's returns are — and it shows the range the rolling Sharpe could plausibly occupy purely from sampling noise, even if the strategy's true skill were perfectly constant. If the rolling line wanders up and down but always stays inside its own confidence band, the strategy is behaving normally; the "decay" a manager thought they saw was noise. If the rolling line moves clearly outside the band, or the band itself keeps shifting to a persistently lower level, that's a real, statistically supported signal that something about the strategy's edge has changed.

A useful example: a strategy's trailing 60-day Sharpe drops from 1.4 to 0.9 over two months, and a nervous PM wants to cut its allocation. Plotted with confidence bands, the 90% interval around a 60-day Sharpe estimate for this strategy's typical volatility turns out to span roughly ±0.6 — so a move from 1.4 to 0.9 is comfortably inside the band of noise a strategy with a true Sharpe near 1.1–1.2 would produce by chance. The rolling number moved, but not by more than ordinary sampling variation would produce; nothing here supports a decay narrative yet. If instead the Sharpe had fallen to 0.2, clearly outside that band, the same tool would support a very different, much more urgent conclusion.

What this means in practice

Rolling Sharpe with confidence bands is one of the standard tools for separating a real change in a strategy's edge from the ordinary noise every finite sample carries, and it directly guards against the common mistake of reacting to every dip in a rolling metric as if it were a verdict. Shorter windows react faster to real changes but carry wider, noisier bands; longer windows are more stable but slower to flag genuine decay — the choice of window length is itself a trade-off a monitoring team has to make deliberately, not an afterthought.