Sharpe Ratio Under Autocorrelation and Non-Normal Returns

The Sharpe ratio is only easy to annualize and compare when returns behave far more cleanly than trading strategies usually do.

The Intuition

The Sharpe ratio looks simple:

$$ \widehat{SR} = \frac{\bar{r}}{s}, $$

where $\bar{r}$ is the mean excess return per period and s is the sample standard deviation.

That simplicity hides two fragile assumptions.

First, the familiar annualization rule

$$ SR_{\text{annual}} \approx \sqrt{q}\, SR_{\text{periodic}} $$

is valid only when returns are serially independent. Second, the usual confidence-interval logic is cleanest under nearly Gaussian returns. Many trading strategies violate both conditions. They show serial dependence from smoothing, overlapping holdings, or staggered exits, and they show skewness and fat tails from crash exposure or stop-loss behavior.

So the Sharpe ratio is not wrong. It is easy to misuse.

The Fixed-Strategy Layer

This primer is about fixed-strategy Sharpe uncertainty: what a reported Sharpe means when the strategy itself is held fixed. That is different from the search problem handled by the Deflated Sharpe Ratio. DSR adds a multiple-testing penalty on top of the uncertainty described here.

If you do not understand this fixed-strategy layer, search-aware inference becomes opaque.

The Math

For a single return series, an approximate variance term for the Sharpe estimator is

$$ \operatorname{Var}(\widehat{SR}) \approx \frac{1 - \hat{\gamma}_3 \widehat{SR} + \frac{\hat{\kappa}-1}{4}\widehat{SR}^2}{T-1}, $$

where:

• T is sample size
• $\hat{\gamma}_3$ is sample skewness
• $\hat{\kappa}$ is sample kurtosis

Negative skew and fat tails increase the estimator's uncertainty. Two strategies with the same sample Sharpe do not carry the same statistical evidence if one has smooth, symmetric returns and the other has crash-prone left tails.

Why Autocorrelation Breaks Naive Annualization

Under IID returns, the variance of a q-period cumulative return is just q times the one-period variance. That is why the $\sqrt{q}$ rule works.

With autocorrelation, the q-period variance becomes

$$ \operatorname{Var}\!\left(\sum_{t=1}^{q} r_t\right) = q\sigma^2 \left[ 1 + 2 \sum_{k=1}^{q-1}\left(1-\frac{k}{q}\right)\rho_k \right], $$

where $\rho_k$ is the lag-k autocorrelation.

This yields Lo's autocorrelation-adjusted annualization:

$$ SR_q = SR_1 \sqrt{ \frac{q} {1 + 2 \sum_{k=1}^{q-1}\left(1-\frac{k}{q}\right)\rho_k} }. $$

The denominator is the correction term.

• Positive autocorrelation makes it larger, so naive annualization overstates Sharpe.
• Negative autocorrelation makes it smaller, so naive annualization can understate Sharpe.

The same logic also changes effective sample size. Positive dependence means fewer independent bets than the raw count of observations suggests.

Worked Example

Suppose a daily strategy has a native daily Sharpe of 0.08. The naive annualized Sharpe is

$$ 0.08 \times \sqrt{252} \approx 1.27. $$

Now assume all higher-order autocorrelations are negligible and only $\rho_1$ matters.

Case 1: Positive autocorrelation

If $\rho_1 = 0.25$, then

$$ 1 + 2\left(1-\frac{1}{252}\right)\rho_1 \approx 1.50, $$

so the correct annualization factor is about

$$ \sqrt{\frac{252}{1.50}} \approx 12.97, $$

and the adjusted annualized Sharpe is

$$ 0.08 \times 12.97 \approx 1.04. $$

Case 2: Negative autocorrelation

If $\rho_1 = -0.25$, the denominator is about 0.50, so the annualization factor rises to roughly 22.4, producing an adjusted annualized Sharpe near 1.79.

The strategy's native-period Sharpe did not change. The dependence structure did. That is why comparing annualized Sharpe across strategies with different rebalancing frequencies, holding periods, or overlap patterns can be misleading.

This also explains why daily and monthly Sharpe ratios are not automatically commensurable. A monthly strategy marked daily does not suddenly create 21 independent monthly bets per month. The native economic decision frequency and the serial dependence induced by holding periods matter more than the sheer number of marks in the backtest.

Where Dependence Comes From in Trading

Autocorrelation in strategy returns is common for practical reasons:

• overlapping positions smooth returns mechanically
• stale pricing or illiquid marks create artificial smoothness
• stop-loss and rebalance rules create short-run reversal patterns
• monthly or quarterly holdings compared to daily marks induce overlap effects

Positive autocorrelation often appears in smoothed or slowly rebalanced strategies. Negative autocorrelation often appears in mean-reversion or over-correcting execution logic. The sign matters because it changes both confidence intervals and annualization.

Overlapping returns create a related problem. If you evaluate a 20-day holding-period strategy using daily overlapping portfolio returns, adjacent observations share most of their underlying PnL path. The sample size looks large, but the effective number of independent bets is much smaller. That inflates naive confidence if dependence is ignored.

In Practice

Always start from the native decision frequency. Compute Sharpe where the strategy actually makes economic bets, then annualize only after checking dependence.

Inspect the autocorrelation function of returns, not just the signal. Portfolio construction, rebalancing rules, and cost accounting can introduce dependence even when the underlying alpha forecast is simple.

Use the non-normality correction when returns are visibly skewed or heavy-tailed. A crash-prone strategy with the same sample Sharpe as a symmetric one deserves a wider confidence interval.

Do not confuse this topic with DSR. A fixed-strategy Sharpe can be estimated carefully and still be unconvincing after search adjustment. The two corrections address different problems.

Common Mistakes

• Annualizing with $\sqrt{q}$ without checking serial dependence.
• Comparing Sharpe across strategies with different holding-period overlap or smoothing conventions.
• Ignoring skewness and kurtosis when the strategy clearly has asymmetric tail exposure.
• Treating daily marks on a monthly strategy as if they created 252 independent bets.
• Jumping straight to search-aware inference before getting the fixed-strategy uncertainty right.

Connections

This primer underpins Chapter 16's discussion of Sharpe credibility and feeds directly into the DSR primer. It also matters in Chapter 17 because allocators often compare strategies at different turnover or rebalancing frequencies. Chapter 26's monitoring topics are distinct: they ask how live performance drifts over time, whereas this primer asks how much evidence a fixed backtest Sharpe contains in the first place.