Annualizing the Sharpe ratio: serial correlation and the √12 rule

A one-period Sharpe ratio is mean excess return divided by the standard deviation of returns for that same period. Excess return subtracts a risk-free return measured at a matching frequency. A monthly Sharpe therefore uses monthly excess returns and monthly return volatility. Subtracting a quarterly risk-free rate from monthly returns, or mixing simple and log returns, changes the quantity being measured. A reported Sharpe is an estimate based on a chosen sample and definition. It is not the probability of a profitable next trade, a maximum-loss measure, or a guarantee of future performance. William Sharpe’s original explanation presents the basic reward-to-variability measure. Comparisons also require consistent return construction and risk definitions.

Let SR₁ be a one-period Sharpe and let one year contain q periods. If the periodic mean and variance are stable and returns have zero covariance across periods, the mean of an additive q-period excess return is q times the one-period mean and its standard deviation is √q times as large. Under those conditions, SR_q=√q×SR₁. IID returns are a common sufficient assumption. Monthly frequency alone does not make monthly returns independent. Trend persistence, reversals, overlapping holding periods, and smoothed marks can link adjacent observations. Sharpe’s original discussion calls out the zero-serial-correlation condition behind this aggregation. Changing data frequency and applying √q without checking that condition hides an important assumption.

Write the periodic excess return as xₜ, its variance as γ₀, and its lag-k covariance as γₖ=Cov(xₜ,xₜ₋ₖ). Under a time-invariant covariance structure, the variance of a q-period sum is Var(Σxₜ)=qγ₀+2Σₖ₌₁^{q−1}(q−k)γₖ. Two monthly returns with the same variance can still produce different aggregate risk when they move together. Looking at lag 1 alone does not automatically account for longer-lag covariance. With ρₖ=γₖ/γ₀ and SR₁=μ/σ, the Sharpe for an additive q-period sum is SR_q=√q×SR₁/√(1+2Σₖ₌₁^{q−1}(1−k/q)ρₖ). This is Lo’s (2002) time-aggregation correction. The adjustment factor summarizes all lag correlations that enter the q-period variance. Positive autocorrelation generally raises aggregate variance and lowers the Sharpe relative to naive √q scaling; negative autocorrelation can work in the other direction.

Suppose mean monthly excess return is 0.4% and monthly standard deviation is 2%. The monthly Sharpe is 0.004/0.02=0.20. If monthly returns are treated as uncorrelated, naive annualization gives 0.20×√12≈0.693. For comparison, take a wholly hypothetical AR(1) example with lag-1 autocorrelation 0.25 and, under that model, ρₖ=0.25ᵏ. The 12-period adjustment denominator is √(1+2Σₖ₌₁¹¹(1−k/12)×0.25ᵏ)≈1.262, so the adjusted figure is 0.693/1.262≈0.549. The naive value is higher in this example. These numbers illustrate sensitivity to a correlation assumption; they are not a forecast or the measured result of any strategy. The powers apply because this example assumes AR(1), not because a lag-1 estimate determines every later lag.

A strategy with persistent positions, momentum, or reversal may have economically persistent or reversing returns. Rolling three-month returns calculated every day also overlap, mechanically linking observations that should not be treated as independent. These mechanisms can leave similar statistical traces while requiring different interpretation and inference. Illiquid assets with infrequent marks or model valuations anchored to earlier prices can appear smoother and show positive autocorrelation. Trade-price reversals or short-term inventory effects can create negative correlation. Before adjusting the statistic, inspect sampling frequency, holding period, price source, and how positions are marked. Removing observed correlation mechanically can hide risk that belongs to the strategy; treating a recording artifact as predictive skill can be misleading too.

Align the return and risk-free frequencies. Choose either gross or net results and apply commissions, slippage, funding, and borrowing costs consistently. Leverage, cash balances, dividend reinvestment, and portfolio weights change the excess-return series. State whether observations are capital-weighted portfolio returns or equally weighted trade returns. The formula above applies to additive aggregation of q periodic excess returns under its assumptions. A compounded annual simple return is not generally equal to the sum of monthly simple returns. Mixing a monthly arithmetic mean, a compounded annual return, and an annualized volatility can make the numerator and denominator refer to different time conventions. Disclose how each series and annualization was constructed so another reader can reproduce the result.

Lo’s correction does not make sample autocorrelations known constants. With only a few years of monthly data, correlations at longer lags such as six through eleven months can be very noisy. A positive lag-1 estimate does not establish an AR(1) process with ρₖ=ρ₁ᵏ. If you use an AR(1) model, state why, inspect its residuals, and report sensitivity to alternatives. Show the sample length and frequency, the autocorrelation profile, the correction formula, and the estimation method. Check whether changing the sample window or correlation estimates changes the result and strategy ranking. The Sharpe itself also estimates a mean and variance, so its sampling uncertainty may need a standard error or interval. Lo’s 2002 paper discusses the statistical distribution and uncertainty of Sharpe estimates as well as serial-correlation adjustment.

If many strategies or parameter settings were tried and only the highest Sharpe was kept, data snooping remains after time-aggregation correction. Regime changes, ruin risk, extreme losses, illiquidity, execution costs, and leverage are not repaired by this formula. A higher corrected Sharpe does not guarantee safer or better future performance. For model selection and backtest overfitting, see [Bias–variance trade-off and financial-model overfitting](/en/learn/bias-variance-tradeoff-financial-model-overfitting-explained). For the path of losses, see [maximum drawdown and recovery](/en/learn/maximum-drawdown-and-recovery-percentage-explained). For portfolio return compounding, continue with [arithmetic and geometric average returns](/en/learn/arithmetic-vs-geometric-average-returns-explained).

A reproducible comparison should include: - Return frequency, observation count, sample dates, and source and conversion of the risk-free rate - Gross or net basis, portfolio weights, leverage, and whether aggregation is compounded or additive - Periodic Sharpe, naive √q annualization, the adjustment formula, and estimated lag correlations - Rationale and sensitivity to alternatives if a model such as AR(1) is assumed - Sampling uncertainty for Sharpe, supplemented by drawdown, tail-loss, and liquidity measures Do not highlight only the highest Sharpe. State the data-selection process and performance-comparison rules. This article explains statistical performance measurement; it is not trading advice or a return promise.