Annualising a monthly Sharpe ratio: why short histories mislead

Multiplying a monthly Sharpe ratio by sqrt(12) changes its reporting horizon; it does not create twelve independent years of evidence. A short, unusually calm history can produce a striking score, while another sampling method can produce a different answer from the same investments.

Neither annualisation method identifies the portfolio with the greatest geometric wealth growth. Exact compounding and net wealth must be calculated separately; annualising a ratio does not supply a growth forecast.

The main Sharpe guide introduces the statistic. Here we keep one historical dataset fixed and change the observation window and sampling frequency, then use a separate hypothetical example to isolate serial correlation.

Key takeaways

• Multiplying a monthly Sharpe ratio by sqrt(12) does not add observations or lengthen the underlying history.
• Serial correlation can change the variance scaling assumed by that convention.
• Compounded annual wealth and geometric growth require separate calculations from an annualised Sharpe score.

Begin with 120 monthly observations

Use January 2016–December 2025 nominal US-dollar total returns for the value-weighted US stock market and the matching monthly Treasury-bill series from the Kenneth French Data Library. This is not an S&P 500 price series.

Each observation is the stock return minus that month's bill return. Their monthly mean divided by their sample standard deviation, using n − 1, gives 0.23756. Multiplying by sqrt(12) gives an annualised Sharpe ratio of 0.82295, rounded to 0.823.

This calculation uses 120 observations over ten calendar years. It does not imply 120 years of evidence, and the observations need not be independent. The monthly input CSV allows the exact period to be extracted.

Calendar-year observations answer a different statistical question

For each year, compound its twelve stock returns and separately compound its twelve bill returns. Subtract the resulting annual bill return from the annual stock return. Then calculate the arithmetic mean and sample standard deviation of those ten annual differences.

The ratio is 0.87315, rounded to 0.873. No further annualisation is necessary because these observations are annual already. The ten annual inputs and method comparison reproduce both calculations.

Neither result is an arithmetic mistake. Compounding changes the observations, and sample moments do not transform mechanically between frequencies. The annual sample also has only ten data points. The gap alone is not evidence that one method proves better future performance.

Short windows change the story far more

Within the same ten-year monthly history, calculate every consecutive twelve-month and 36-month window using the same monthly method:

The twelve-month minimum is January–December 2022; the maximum is December 2016–November 2017. The 36-month minimum is April 2017–March 2020; the maximum is January 2023–December 2025. These are retrospective extrema, selected after inspecting every eligible window.

The complete rolling-window CSV prevents a single favourable date pair from standing in for the whole exercise. Overlapping windows share most of their observations: 109 windows are not 109 independent experiments. Their range is descriptive, not a confidence interval or a forecast of future ratios.

Why serial correlation changes the square-root rule

In his 1994 paper, Sharpe derives square-root scaling under additive differential returns and zero serial correlation, then warns that compounding and serial dependence complicate the relationship.

Consider a separate hypothetical stationary monthly excess-return process with mean 0.5%, standard deviation 2%, and correlation at lag k equal to 0.3^k. These are assumed properties, not estimates fitted to our stock data.

For twelve summed monthly excess returns, the variance multiplier is:

12 + 2 × sum[(12 − k) × 0.3^k], for k from 1 to 11

It equals 21.0612, rather than 12. Using that variance gives an additive annual ratio of 0.654, compared with the naive monthly-ratio-times-sqrt(12) value of 0.866. The autocorrelation scenario CSV records the calculation.

This is an exact covariance calculation for summed excess returns under the assumed process. It is not an exact Sharpe ratio for compounded annual wealth, and it is not an empirical adjustment to the 0.823 historical result.

What a reported score needs beside it

Dates, observation frequency, benchmark, sample denominator and annualisation convention make a comparison interpretable. A high score still leaves crash losses and illiquidity unresolved. The negative-Sharpe example and benchmark sensitivity show why even a correctly computed number has limited scope.