Two reports show a column called Sharpe. One divides average periodic excess return by its standard deviation. The other divides compounded annual growth by annualised volatility. Should they share a ranking? No: the labels conceal different calculations before any argument about strategy quality begins.
This research note is a conceptual methodology note, not a new backtest or a verdict on an archived system.
Define the number before defending it
Sharpe’s historical ratio uses the average differential return divided by the standard deviation of that differential return. The reference return therefore belongs in the definition, not in a footnote added after comparison.[1] For a conventional excess-return Sharpe, specify the cash or risk-free series, including an explicit zero-rate assumption if that is what the calculation uses.
An estimator contract is a short record of the return interval, arithmetic or logarithmic returns, reference series, sample-standard-deviation convention, annualisation rule, missing-value treatment and cost basis. It also identifies whether the observations describe a portfolio through time or only completed trades. Quietly dropping inactive periods changes the object being measured.
Compounded growth divided by volatility can be reported under its own name. It does not become a conventional Sharpe because a source file calls it one. Likewise, square-root-of-time scaling is not a universal identity: Sharpe’s simple derivation assumes summed periodic differential returns with zero serial correlation, and explicitly discusses the complications of compounding and dependence.[1]
Conceptual example
Imagine two analysts receiving exactly the same ordered return vector. Analyst A computes a periodic excess-return Sharpe. Analyst B compounds the returns before constructing a growth-to-volatility ratio. Neither has discovered a different strategy; they have asked different statistical questions. Harmonising the label requires recomputing the estimator, not relabelling B’s answer.
Now imagine that A’s vector was selected because it looked best among many alternatives. Reproducing A’s calculation would establish arithmetic consistency, not independent evidence. Bailey and López de Prado’s Deflated Sharpe Ratio addresses selection under multiple testing and non-normal returns; its inputs must describe the search rather than just the winner.[2]
Cheap check
Before opening another dataset, request the ordered periodic returns, reference returns, estimator contract and search log. Independently calculate the stated ratio and compare it with the reported value using a tolerance declared from the report’s rounding precision.
Pass: the independent calculation agrees within that tolerance; periods, units and exclusions are explicit; the search history distinguishes development choices from untouched evaluation. This is a pass for reproducibility and provenance only.
Fail: the number requires an undocumented formula, discarded observations or a supposed holdout that influenced selection. Withhold the validation claim. An incomplete trial log is unresolved evidence, not permission to assume one trial.
Limitations
A correct estimator can describe a poor investment. Search adjustments depend on assumptions about the trials and returns; they do not repair unavailable information, fictional fills or an economically inappropriate benchmark. Conversely, a failed reproduction can expose a reporting defect rather than disprove the underlying mechanism. Diagnose that distinction before writing an autopsy.
The validation gate separates these questions. Continue with choosing an honest benchmark, or use the research library to place a method in context.
Sources
[1] https://web.stanford.edu/~wfsharpe/art/sr/sr.htm — William F. Sharpe — The Sharpe Ratio (1994) [2] https://www.davidhbailey.com/dhbpapers/deflated-sharpe.pdf — Bailey and López de Prado — The Deflated Sharpe Ratio (2014)