The historical Sharpe ratio describes mean differential return per unit of its variability. First subtract the matching-period benchmark from each strategy return, then calculate the mean and standard deviation of those differences. A higher reported number is not proof of lower future loss or a more reliable strategy.

Where the Sharpe ratio came from and what it actually measures

William F. Sharpe published the first version in 1966 in his paper "Mutual Fund Performance" in the Journal of Business, where he called it the reward-to-variability ratio. He won the 1990 Nobel Memorial Prize in Economic Sciences for his work on the Capital Asset Pricing Model, and four years later revisited the measure in the Journal of Portfolio Management, settling on the modern name. The contemporary formula: portfolio return minus the risk-free rate (usually the yield on short Treasury bills), divided by the standard deviation of the excess returns.

A benchmark makes the comparison explicit. As a hypothetical annual-period illustration, means of 30% and 50% with standard deviations of 10 and 30 percentage points give ratios 3 and about 1.67 if the benchmark is fixed at zero. A different benchmark changes the numerators. Compare only matching periods, definitions and costs; the ratio omits correlations with other holdings and does not fully describe tail losses.

The formula in plain prose and what to plug in

Form a series of strategy returns minus matching-period benchmark returns, then calculate its mean and standard deviation. The historical ratio is their quotient. State whether returns are arithmetic or logarithmic, the observation frequency and the annualisation convention. Square-root-of-time scaling needs assumptions about serial correlation and aggregation; it is not automatically valid for every return series.

Choose and disclose a benchmark matching the measurement currency and period. A policy rate or a long-term government-bond yield is not automatically the return on an investable risk-free asset over that period. Record the actual benchmark series used; do not silently substitute an annual quote for a monthly realised return.

Variable benchmark:measure the differences

Four hypothetical periods:r = 1%,3%,1%,3%; the selected variable benchmark b = 0%,2%,0%,2%. Differencesr−b are always1 percentage point. Both series vary, but the differences' standard deviation iszero. This generalized example uses a variable benchmark, not a constant risk-free rate.

Scroll sideways to see the full diagram

Changing series, constant differencesStrategy1,3,1,3% and benchmark0,2,0,2% differ by 1 percentage point in each period. Standard deviation of differences iszero; the ratio is undefined.Returns[%]32101234Strategy: 1, 3, 1,3Benchmark: 0, 2, 0,2Difference: 1, 1, 1,1 p.p.SD of differences = 0; ratio undefined.
The constant vertical gap shows why differences stay constant as both series change.

Meanr is 2%, meanb is 1%, and the mean difference is 1%. SampleSD ofr alone is about 1.1547 percentage points; using it gives about 0.866 but does not measure variability of the specified differences. First buildr−b for each matching period. A zero denominator proves neither a perfect strategy nor absence of future variability.

Assumptions: Invented net returns over matching periods. This uses Sharpe's generalized historical ratio with a variable benchmark. SampleSD with divisor3 iszero for the constant difference list. No annualization. Source check: William F.Sharpe—The Sharpe Ratio(1994).

How to interpret a reported number

A value such as 1, 2 or 3 has no universal pass/fail meaning. Ask which returns, benchmark, frequency, costs and aggregation produced it. Comparisons based on different definitions can rank strategies differently, while a high sample value can coexist with unobserved tail risk.

Record the numerator and denominator beside the ratio. A small denominator can make the result large without demonstrating that future losses are small. For a variable benchmark, measure the differences themselves rather than substituting the standard deviation of strategy returns.

Observation count and time dependence affect estimation uncertainty. A three-year result does not automatically establish skill, and six months is not automatically meaningless. Report the full window and sampling frequency, inspect dependence and selection, and avoid treating any fixed number of months as a statistical guarantee.

A worked example in prose

Hypothetical annual-period summary, without annualising monthly data: A has arithmetic mean return 24% and sample standard deviation 16 percentage points; B has 12% and 5 percentage points. Assume a constant benchmark return of 4% in each annual period and comparable net-return definitions. A gives (24−4)/16 = 1.25; B gives (12−4)/5 = 1.6. These supplied summary statistics illustrate the quotient, not actual test results or a complete risk ranking.

Scaling the entire differential-return series by a positive constant preserves its mean-to-standard-deviation ratio in the ideal cost-free model. Halving a position does not always do that: residual cash return, financing, transaction costs and non-linear execution can matter. The example’s unchanged 1.25 therefore needs matching assumptions rather than a blanket claim that leverage never changes Sharpe.

Three known limitations worth keeping in mind

Standard deviation includes deviations on both sides of the mean. Equal positive and negative returns do not necessarily contribute equally: distance from the mean matters. Sortino instead uses shortfalls below a stated target, under its own denominator convention. The Sortino ratio and side-by-side comparison explain this different question; neither statistic establishes the full risk of a strategy.

Computing a historical mean-to-standard-deviation ratio does not require normal returns. Normality may support particular interpretations, but the ratio alone does not describe skewness or tail-loss severity. Review the distribution and tail scenarios separately; a high historical ratio does not guarantee protection against an extreme move.

Sample length, observation frequency and changing market conditions affect the estimate. Six months is not automatically meaningless, and no fixed 36-month window establishes statistical reliability. Report the full assumptions and supplement the ratio with maximum drawdown, Calmar and per-trade expectancy.

What to do tomorrow

  1. Build matching-period net returns and benchmark returns in two columns, adjusting for cash flows. Subtract them row by row, then use AVERAGE divided by STDEV.S on the differences. Report the resulting period ratio before any annualisation. A published annual benchmark quote divided by twelve is not automatically the realised monthly benchmark return.
  2. If the result is below a chosen review threshold, inspect the complete observations, costs, benchmark, frequency and serial correlation before changing the method. No fixed thirty-six-month cutoff gives a dataset statistical validity.
  3. If Sortino differs from Sharpe, inspect the target return, downside-deviation convention and matching sample. A higher Sortino alone does not prove upside spikes or few deep drawdowns; the denominators measure different things. Publish definitions and relevant results together.
  4. Store the mean difference, its standard deviation, observation count and frequency alongside the ratio. If you report an annualised result, document the aggregation and time-dependence assumptions; do not mix an annual return with monthly volatility.
  5. Before sharing results, state the complete measurement window and its limitations. Check selection, missing observations and dependence, and supplement the ratio with the equity path and losses. More elapsed years alone do not ensure that the sample represents future conditions.