Risk & Metrics
Sharpe Ratio
Return per unit of volatility. The standard score for whether a strategy's returns were worth the bumpiness it took to earn them.
The Sharpe ratio compares the average return above a chosen reference return
with the variability of that difference. Define each period's differential
return as d(t) = strategy return − reference return. Then:
period Sharpe = mean(d) / standard deviation(d)
annualized Sharpe = period Sharpe × √N
N is the number of equally spaced return periods in the chosen year. For
daily, continuously traded crypto observations, an analyst may use 365; a
business-day series may use 252. The convention, return frequency, reference,
and treatment of missing days should be reported rather than assumed.
For example, daily excess returns with a mean of 0.08% and standard deviation
of 1.60% produce (0.0008 / 0.016) × √365 ≈ 0.96. Setting the reference return
to zero would change the input and should be an explicit choice.
William Sharpe's own 1994 explanation of the ratio distinguishes a historical ratio from using one to predict future performance.
Reading the number
There are no universal “bad,” “good,” or “excellent” Sharpe bands. A value's uncertainty depends on the sample length, return frequency, serial correlation, search process, and strategy. Compare like-for-like estimates, report the period and assumptions, and give out-of-sample results more weight than a number selected from many in-sample trials. A high value can still be an overfit result; a low value alone does not prove there is no useful effect.
Limitations
Sharpe counts upside and downside variation alike and compresses the return path into one mean and one standard deviation. Two strategies with equal ratios can have very different max drawdowns, skew, liquidity, and tail losses. Square-root annualization assumes comparable, uncorrelated period returns. Autocorrelation changes the scaling relationship, while irregular sampling can make the annualized number look more precise than it is.
On AlphaProve
AlphaProve's tearsheet puts the headline ratio beside the equity curve, drawdown, Sortino, Calmar, and Omega metrics, and includes a rolling 30-trade view. The rolling view can reveal concentration in one part of the sample, but thirty trades is still a small, overlapping window and should not be treated as a second independent test. The overfitting experiment shows why the selection process matters as much as the reported ratio.