The Sharpe Ratio, Properly Derived

Every quant knows the formula. Most cannot derive the standard error, explain why the geometric mean Sharpe differs from the arithmetic mean Sharpe, or tell you why their reported “Sharpe 2.3” is biased upward by 15%. This article derives the whole thing.

We will start from a return series, build up the Sharpe ratio, derive its expectation and variance under non-normality, derive the bias correction, and finish with the annualization argument. By the end, you will know what every term in the formula costs you and why a reported Sharpe of 1.5 is not the same thing as a true Sharpe of 1.5.

The Definition

Let $r_t$ be the simple return at time $t$, for $t = 1, \ldots, T$. The arithmetic mean return is:

The sample standard deviation is:

The Sharpe ratio, as William Sharpe originally defined it (1966), is:

That is the entire formula. Two numbers divided. The rest of this article is about the things that go wrong when you take those two numbers from a finite sample and try to draw conclusions.

Arithmetic vs Geometric Mean: The Forgotten Term

Most implementations use the arithmetic mean return. The geometric mean return (the one that compounds correctly) is:

The approximation holds for small returns and small variance. The second term, $\hat{\sigma}^2/2$, is the variance drag. For a strategy with 10% annualized volatility, the drag is $0.10^2/2 = 0.005$, or 50 basis points per year. For a 20% volatility strategy, the drag is 200 basis points. That is the cost of being wrong about your mean.

If you compute Sharpe using the arithmetic mean, you are crediting yourself with the geometric mean’s variance drag. This is the standard practice. It is also slightly dishonest. The honest computation uses the geometric mean, which gives the Sharpe ratio in the compounded sense:

For a high-volatility strategy, the gap between arithmetic and geometric Sharpe is non-trivial. A 2.0 arithmetic Sharpe on 30% annualized vol becomes a 1.85 geometric Sharpe. Most reporting systems hide this. López de Prado’s Advances in Financial Machine Learning (2018, Ch. 14, ISBN 978-1-119-48208-6) discusses this in the context of strategy reporting.

Expected Value of the Sharpe Estimator

Suppose returns are i.i.d. with mean $\mu$, variance $\sigma^2$, skewness $\gamma_3$, and excess kurtosis $\gamma_4$. The sample mean $\bar{r}$ is unbiased: $E[\bar{r}] = \mu$. The sample variance $s^2 = \frac{1}{T-1}\sum (r_t - \bar{r})^2$ is also unbiased: $E[s^2] = \sigma^2$.

But the ratio $\bar{r}/s$ is biased. The bias arises because $s$ is a noisy estimate of $\sigma$, and ratio distributions are not ratios of expectations.

To derive the bias, expand $\bar{r}/s$ around $\mu/\sigma$ using the delta method. Define $g(\bar{r}, s) = \bar{r}/s$. The first-order Taylor expansion around $(\mu, \sigma)$ gives:

Taking expectations:

The first expectation is zero. The second is $E[s] - \sigma$, which is negative because $s$ is biased downward in small samples (a square root of a chi-squared is biased down). To leading order:

So $E[s - \sigma] \approx -\sigma/(4T)$. Plugging in:

The bias is positive and proportional to $1/T$. For $T = 252$ (one year of daily data), the bias is roughly $1/(4 \cdot 252) \approx 0.001$, or 0.1% of the true Sharpe. For $T = 60$ (one year of weekly data), the bias is $1/240 \approx 0.004$, or 0.4%. For very short samples, the bias compounds with the multiple testing bias (covered in the Deflated Sharpe article) and the autocorrelation bias (covered in the Haircut Sharpe article).

Variance of the Sharpe Estimator

The variance of $\widehat{SR}$ under the null $\mu = 0$ is the foundation of hypothesis testing on Sharpe. Start from the same Taylor expansion but now compute the second moment:

For i.i.d. returns, $\text{Var}(\bar{r}) = \sigma^2 / T$, $\text{Var}(s) \approx \sigma^2 / (2T)$, and $\text{Cov}(\bar{r}, s) \approx 0$ to leading order. So:

For small $SR$, the variance is approximately $1/T$. This is the textbook result: under the null of zero mean, the sample Sharpe has variance $1/T$. The standard error is $1/\sqrt{T}$.

But this is only valid for Gaussian returns. With skewness and excess kurtosis, the variance picks up additional terms. The full formula, including non-normality corrections, is:

This is the formula in Bailey and López de Prado (2014), “The Deflated Sharpe Ratio: Correcting for Selection Bias, Non-Normality, and Multiple Testing,” and the formula López de Prado reproduces in Advances in Financial Machine Learning (2018, §14.7.2).

The non-normality corrections matter. A strategy with negative skewness (typical of short volatility strategies) has larger Sharpe variance than a Gaussian strategy with the same mean and standard deviation. The intuition: a few large negative returns add more noise to the mean estimate than the standard deviation can offset. A typical short-vol strategy with $\gamma_3 = -1$ and $SR = 1$ has roughly 1.5x the Sharpe variance of a Gaussian strategy with the same parameters.

Lo’s Bias Correction

Andrew Lo (2002), “The Statistics of Sharpe Ratios,” derived a sharper bias correction that accounts for autocorrelation in returns. The corrected Sharpe ratio is:

Wait, the more commonly cited correction from Lo (2002) is simpler: it adjusts the standard error of the Sharpe ratio to account for serial correlation, but the bias correction he derives for the Sharpe ratio itself is roughly:

where $\rho_1$ is the first-order autocorrelation of returns. The first term is the standard small-sample bias. The second is the additional bias from positive autocorrelation: positively autocorrelated returns have lower effective sample size, which inflates the Sharpe ratio. The Haircut Sharpe article in this module derives this in detail.

Annualizing the Sharpe

The annualization formula is straightforward but routinely misapplied. If returns are measured at frequency $f$ (with $f = 252$ for daily, $f = 52$ for weekly, $f = 12$ for monthly), the annualized Sharpe is:

The reasoning: variance scales linearly with time, standard deviation scales with $\sqrt{T}$, and Sharpe is mean over standard deviation. If you scale time by $f$, variance scales by $f$, standard deviation scales by $\sqrt{f}$, and Sharpe scales by $\sqrt{f}$.

But this assumes returns are independent across periods. For monthly returns of an equity strategy, the assumption is roughly correct. For daily returns with significant autocorrelation, the assumption breaks. For high-frequency returns with strong mean reversion, the assumption breaks badly.

There is also the question of which frequency to use. Annualizing a daily Sharpe by $\sqrt{252}$ is conventional but obscures the bias. A 5-year daily backtest gives $T = 1260$, with bias $1/(4 \cdot 1260) \approx 0.0002$. A 5-year monthly backtest gives $T = 60$, with bias $1/240 \approx 0.004$. The monthly bias is 20x larger. The standard error also changes: a daily Sharpe of 0.1 has standard error $1/\sqrt{1260} \approx 0.028$, while a monthly Sharpe of 0.1 (annualized from $\sqrt{12} \approx 0.346$) has standard error $\sqrt{12}/\sqrt{60} \approx 0.346$.

The practical implication: use daily data when you can. The bias is smaller, the standard error is smaller, and the sampling distribution is closer to Gaussian (by the central limit theorem).

A Worked Example

Consider a strategy with 5 years of daily returns ($T = 1260$), mean return $\mu = 0.0005$ per day, standard deviation $\sigma = 0.01$ per day, no skewness or excess kurtosis, no autocorrelation. The sample Sharpe is:

Annualized: $\widehat{SR}_{annual} = 0.05 \cdot \sqrt{252} \approx 0.794$.

Standard error under the null: $1/\sqrt{1260} \approx 0.028$ (daily). Annualized: $0.028 \cdot \sqrt{252} \approx 0.446$.

So the 95% confidence interval for the true annualized Sharpe is approximately $0.794 \pm 1.96 \cdot 0.446 = [-0.08, 1.67]$.

That confidence interval contains zero. It contains 1.0. It contains “you have no edge.” Reporting a Sharpe of 0.79 with a star next to it would be misleading.

import numpy as np

T = 1260
mu, sigma = 0.0005, 0.01
sr_daily = mu / sigma
sr_annual = sr_daily * np.sqrt(252)
se_daily = 1.0 / np.sqrt(T)
se_annual = se_daily * np.sqrt(252)

print(f"Daily Sharpe: {sr_daily:.4f}")
print(f"Annual Sharpe: {sr_annual:.4f}")
print(f"95% CI (annual): [{sr_annual - 1.96*se_annual:.3f}, {sr_annual + 1.96*se_annual:.3f}]")

A Sharpe of 0.79 over 5 years looks respectable. The confidence interval tells the real story. This is why the Probabilistic Sharpe Ratio (PSR), covered in the next article, is a better way to summarize a backtest.

What This All Means for Reporting

If you report a Sharpe ratio, you should also report:

1. The sample size $T$ and frequency of returns.
2. The bias correction (Lo or AFML formulation).
3. The standard error under non-normality (using $\gamma_3$ and $\gamma_4$).
4. The confidence interval, not just the point estimate.
5. Whether you used arithmetic or geometric mean.

A “Sharpe 2.0” without a confidence interval is not information. It is marketing. The next article in this module derives the Probabilistic Sharpe Ratio, which gives you a single number that summarizes how confident you should be in your reported Sharpe relative to a benchmark.

The Information Ratio as a Special Case

The Information Ratio (IR) is the Sharpe ratio applied to active returns, the returns of a portfolio relative to a benchmark:

The derivation is identical. The bias, variance, and standard error formulas apply directly to the IR with $\mu$ replaced by the mean active return and $\sigma$ replaced by the active return standard deviation.

For a portfolio with tracking error 5% and active return 3%, the IR is 0.6. The standard error under Gaussian assumptions is approximately $1/\sqrt{T}$ (for IR in the same frequency as the returns). For monthly active returns over 5 years ($T = 60$), the standard error is $1/\sqrt{60} = 0.129$. The 95% CI is $0.6 \pm 1.96 \cdot 0.129 = [0.35, 0.85]$.

IR has the same multiple testing problem as Sharpe. A search over many active strategies will produce a high-IR strategy by chance. The Deflated Information Ratio (analogous to DSR) corrects for this.

The Sharpe Ratio Under Stable Paretian Distributions

Most financial return series have heavy tails. The Gaussian assumption is convenient but often wrong. The Stable Paretian (or $\alpha$-stable) distribution is a generalization that includes heavy tails.

For an $\alpha$-stable distribution with characteristic exponent $\alpha$ (where $\alpha = 2$ is Gaussian and $\alpha < 2$ has heavy tails), the Sharpe ratio is still defined as $E[r]/\sigma[r]$, but the standard error formula changes.

For $\alpha < 2$, the variance may be infinite (if $\alpha < 2$). The Sharpe ratio is still well-defined, but the variance-based standard error breaks down. Alternative formulations use the scale parameter of the stable distribution.

In practice, most backtests assume Gaussian or near-Gaussian returns and use the variance-based Sharpe. This is approximately correct for moderate tail thickness. For strategies with extreme tail risk (e.g., crisis-period short vol), the Gaussian approximation is poor.

The Sharpe Ratio in Different Return Definitions

The choice of return definition affects the Sharpe ratio. Common choices:

Simple returns: $r_t = (P_t - P_{t-1}) / P_{t-1}$. Used for most equity strategies.

Log returns: $r_t = \log(P_t / P_{t-1})$. Used for theoretical work; approximately equal to simple returns for small returns.

Excess returns: $r_t = R_t - R_f$. Used for strategies evaluated against the risk-free rate.

Active returns: $r_t = R_t - R_b$. Used for active portfolio management relative to a benchmark.

For small returns (say $|r_t| < 5\%$), all four definitions give similar Sharpe ratios. For large returns (cryptocurrencies, leveraged ETFs, leveraged FX), the differences are non-trivial.

The log return formulation has nicer aggregation properties (log returns sum across time, simple returns do not). For computing the Sharpe ratio of multi-period returns, log returns are preferred.

The Sharpe Ratio as a Frequentist Estimator

The Sharpe ratio estimator is a function of the data. It has a sampling distribution under the assumed data-generating process. The properties derived above (bias, variance, asymptotic normality) are properties of this estimator.

A Bayesian would treat the Sharpe ratio as a parameter with a prior distribution and update the prior using the data to get a posterior. The frequentist Sharpe ratio estimator is the posterior mean under a flat prior.

The choice between frequentist and Bayesian framing is mostly philosophical for the Sharpe ratio. Both give similar point estimates. They differ in how they quantify uncertainty (confidence interval vs credible interval vs posterior probability of exceeding a benchmark).

For most quant use cases, the frequentist framing is sufficient. The Bayesian framing is useful when you have strong prior information about the strategy class (e.g., “all our momentum strategies have Sharpes in [0.3, 1.5]”).

Sharpe Ratio Criticisms

The Sharpe ratio has been criticized on several grounds:

Criticism 1: Penalizes upside volatility. The Sharpe ratio penalizes both upside and downside volatility equally. A strategy with large positive returns and small negative returns has a lower Sharpe than a strategy with the same mean return but smaller absolute volatility. The Sortino ratio (which uses downside deviation) addresses this.

Criticism 2: Ignores path dependence. A strategy with steady returns and a strategy with volatile returns that compound to the same endpoint have the same Sharpe. The Calmar ratio (which uses max drawdown) addresses this.

Criticism 3: Not robust to outliers. One extreme return can dominate the Sharpe ratio calculation. Robust estimators (median absolute deviation, winsorized standard deviation) address this.

Criticism 4: Pathological behavior under non-stationarity. If the return distribution changes over time, the single Sharpe ratio is misleading. Rolling Sharpe ratios address this.

These criticisms are valid for specific contexts. The Sharpe ratio is a useful summary statistic for stationary, approximately symmetric return distributions. For non-stationary or asymmetric distributions, it should be supplemented with other metrics (Sortino, Calmar, conditional Sharpe).

The Sharpe Ratio and Capital Allocation

If you believe your backtest Sharpe is the true Sharpe (a strong assumption), the Kelly criterion gives the optimal capital allocation:

For a strategy with Sharpe 1.0 and volatility 15%, $f^* = 6.67$. So allocate 6.67% of capital to the strategy. In practice, half-Kelly or quarter-Kelly is used to be conservative: 1.67% to 3.33%.

The Haircut Sharpe should be used for Kelly sizing, not the reported Sharpe. If your reported Sharpe is 1.5 and Haircut Sharpe is 1.3, the Kelly size based on 1.5 over-allocates by 15%.

Most quant shops do not use Kelly directly. They use risk-based allocation (e.g., target vol of 10% means position size is 10%/strategy_vol). The Sharpe ratio informs the allocation but does not determine it directly.

A Final Note on Computing the Sharpe

Practical implementation issues:

1. Use daily returns when possible. Monthly or weekly returns have larger bias and standard error. Daily is the standard for liquid strategies.

2. De-mean the returns carefully. If you compute Sharpe using excess returns ($r_t - r_f$), subtract the risk-free rate consistently. Mixing risk-free rate inclusion and exclusion gives nonsense.

3. Handle missing data. If some days have no trades (holidays, half-days), the return is 0, not missing. Including 0-return days in the Sharpe calculation lowers the mean and inflates the standard deviation. Decide upfront whether to include them.

4. Annualize correctly. Daily Sharpe $\times \sqrt{252}$ for annual. Weekly $\times \sqrt{52}$. Monthly $\times \sqrt{12}$. Do not mix frequencies.

5. Report sample size. A Sharpe from 60 daily observations is qualitatively different from a Sharpe from 1260 daily observations. Always report $T$.

If you follow these five rules, your reported Sharpe will be defensible. If you do not, it will not.

The Sharpe Ratio Under Regime Switching

Real financial returns often come from a mixture of regimes (high vol, low vol, trending, mean-reverting). The Sharpe ratio computed across regimes is a weighted average of regime-specific Sharpes:

where $\pi_j$ is the probability of regime $j$, $\mu_j$ and $\sigma_j$ are the regime-specific mean and standard deviation, and $\rho_{jk}$ is the correlation between regimes.

For a two-regime model with high-vol regime ($\pi_1 = 0.2$, $\mu_1 = 0$, $\sigma_1 = 0.02$) and low-vol regime ($\pi_2 = 0.8$, $\mu_2 = 0.001$, $\sigma_2 = 0.005$):

$SR_{overall} = \frac{0.2 \cdot 0 + 0.8 \cdot 0.001}{\sqrt{0.2 \cdot 0.0004 + 0.8 \cdot 0.000025}}$ $= \frac{0.0008}{\sqrt{0.0001}} = \frac{0.0008}{0.01} = 0.08 \text{ daily}$

Annualized: $0.08 \cdot \sqrt{252} = 1.27$. The strategy appears to have a 1.27 Sharpe, but this is dominated by the low-vol regime. In the high-vol regime, the strategy has zero expected return. If the high-vol regime occurs more frequently in the future than in the backtest, the strategy’s true Sharpe is lower.

The honest Sharpe for a regime-switching strategy is the regime-weighted average, with explicit uncertainty about the regime probabilities. Most reporting systems hide this.

The Sharpe Ratio for Multi-Asset Strategies

For a portfolio of multiple assets or strategies, the Sharpe ratio depends on the covariance matrix:

where $\mathbf{w}$ is the weight vector, $\boldsymbol{\mu}$ is the mean return vector, and $\boldsymbol{\Sigma}$ is the covariance matrix.

The Sharpe ratio is maximized by the tangency portfolio:

with maximum Sharpe:

This is the foundational result of mean-variance portfolio theory (Markowitz 1952). The Sharpe ratio of any other portfolio is less than $SR_{max}$.

In practice, $\boldsymbol{\mu}$ and $\boldsymbol{\Sigma}$ are estimated from data, with estimation error. The realized Sharpe is less than the estimated maximum Sharpe. The shrinkage estimator (Ledoit-Wolf) is the standard approach for reducing estimation error.

The Sharpe Ratio and the Nobel

William Sharpe won the 1990 Nobel Prize in Economics (jointly with Harry Markowitz and Merton Miller) for the Capital Asset Pricing Model and the Sharpe ratio. The Sharpe ratio remains the standard risk-adjusted return measure in finance.

The Nobel was awarded for the CAPM, of which the Sharpe ratio is a component. The CAPM itself has been extensively tested and largely rejected by empirical evidence. The Sharpe ratio has survived because it is a useful descriptive statistic even if the CAPM is wrong.

The persistence of the Sharpe ratio is not because it is the best measure. It is because it is simple, intuitive, and computable from a single return series. Better measures exist (Sortino, Calmar, conditional Sharpe) but they require more data or assumptions. The Sharpe ratio is the lowest-friction way to summarize risk-adjusted return.

The Sharpe Ratio in Practice

In industry, the Sharpe ratio is used for:

• Strategy comparison across asset classes.
• Risk budgeting and capital allocation.
• Performance attribution.
• Marketing and investor reporting.

Each use has caveats:

• Strategy comparison: must adjust for autocorrelation (Haircut), multiple testing (DSR), and non-normality.
• Risk budgeting: must use the Haircut Sharpe, not the reported Sharpe.
• Performance attribution: must decompose the Sharpe into signal components.
• Marketing and reporting: must include confidence intervals and statistical diagnostics.

A serious research report includes all of these. A marketing report usually includes only the bare Sharpe. The difference is the difference between honest research and dishonest marketing.

The Final Word

The Sharpe ratio is a useful summary statistic for risk-adjusted return. It is not a perfect measure. It assumes normality, independence, and stationarity. None of these assumptions hold exactly for real financial returns.

The honest user of the Sharpe ratio:

1. Reports the sample size and frequency.
2. Reports the standard error with non-normality correction.
3. Reports the confidence interval.
4. Adjusts for autocorrelation (Haircut).
5. Adjusts for multiple testing (DSR).
6. Computes PSR for significance vs benchmark.
7. Computes MBL for sample size adequacy.
8. Computes PBO for overfitting risk.

Any one of these is incomplete. The combination gives the full picture. The honest Sharpe ratio is a multi-number statistic, not a single number.

References

• Sharpe, W. F. (1966). “Mutual Fund Performance.” Journal of Business, 39(1), 119–138.
• Lo, A. W. (2002). “The Statistics of Sharpe Ratios.” Financial Analysts Journal, 58(4), 36–52.
• Bailey, D. H. and López de Prado, M. (2014). “The Deflated Sharpe Ratio: Correcting for Selection Bias, Non-Normality, and Multiple Testing.” Journal of Portfolio Management, 40(5), 94–107.
• López de Prado, M. (2018). Advances in Financial Machine Learning, Ch. 4 and 13. Wiley. ISBN 978-1-119-48208-6.