Sharpe Ratios Have Clocks

The most dangerous part of a strategy deck is sometimes the tiny multiplier standing in front of the Sharpe ratio:

daily Sharpe * sqrt(252)

It looks like a unit conversion, so it borrows the authority of arithmetic. Daily becomes annual. Monthly becomes annual. Smooth marks look respectable. A research note gets shorter.

But the square-root rule is not a law of nature. It is a model wearing a small hat.

It says, roughly: the returns you are scaling are close enough to independent draws from a stable process that variance grows linearly with time. When that is true, the rule is a useful shorthand. When it is false, the annualized Sharpe ratio is partly a statement about the sampling clock.

This is not investment advice.

The Clean Case

Sharpe’s 1994 formulation compares differential return to the standard deviation of that differential return.¹ For excess returns \(r_t\), the one-period ex post Sharpe ratio is

\[\widehat{S}_1 = \frac{\bar r}{\hat \sigma_1}.\]

If returns are independent with constant variance, then a \(q\)-period return

\[R_t^{(q)} = r_t + r_{t-1} + \cdots + r_{t-q+1}\]

has mean \(q\mu\) and variance \(q\sigma^2\). Therefore

\[S_q = \frac{q\mu}{\sqrt{q}\sigma} = \sqrt{q}S_1.\]

That is the whole spell. Daily to annual becomes \(\sqrt{252}\) because there are roughly 252 trading days in a year. Monthly to annual becomes \(\sqrt{12}\).

The key phrase is not “square root.” The key phrase is “if returns are independent with constant variance.” Financial returns are frequently not that kind.

The General Case

For a stationary return process, write the lag-\(k\) autocovariance as

\[\gamma_k = \operatorname{Cov}(r_t, r_{t-k}).\]

The variance of a \(q\)-period sum is not merely \(q\gamma_0\). It is

\[\operatorname{Var}(R_t^{(q)}) = q\gamma_0 + 2\sum_{k=1}^{q-1}(q-k)\gamma_k.\]

The extra terms are the fine print. Positive autocovariances make multi-period variance grow faster than the IID calculation. Negative autocovariances make it grow slower. Either way, the square-root rule has smuggled a time-series model into a single number.

Lo’s 2002 paper makes this point directly for Sharpe ratios: monthly Sharpe ratios cannot generally be annualized by multiplying by \(\sqrt{12}\), except under special assumptions, and serial correlation can materially change both the level and ranking of reported Sharpe ratios.²

For long horizons, the object you want is the long-run variance

\[\Omega = \gamma_0 + 2\sum_{k=1}^{\infty}\gamma_k.\]

Then a clock-aware annualized Sharpe estimate is

\[\widehat{S}_{\mathrm{LRV}} = \frac{252\bar r}{\sqrt{252\widehat{\Omega}}} = \frac{\sqrt{252}\bar r}{\sqrt{\widehat{\Omega}}}.\]

In practice we do not know all the autocovariances. A Newey-West style heteroskedasticity-and-autocorrelation-consistent estimate truncates the sum and downweights distant lags with Bartlett weights.³ This is not magic. It is just a way to replace “pretend the lags are zero” with “estimate how much the lags contribute.”

A Toy Strategy With a Great Daily Sharpe

The lab below simulates daily excess returns from a strategy with four knobs:

a base Sharpe when the process is unsmoothed and close to IID;
serial correlation in the latent return process;
reported-return smoothing, as if marks arrive slowly or stale prices are averaged into the series;
volatility clustering, which mostly changes uncertainty around the estimate.

The point is not to mimic any fund exactly. The point is to make a measurement error visible: the same economic path can produce a high daily-clock Sharpe and a much lower long-run-variance Sharpe.

I audited the lab and the article numbers before rewriting. At the default settings, the daily-clock Sharpe is 1.47, the HAC Sharpe is 0.90, lag-1 autocorrelation is 57.9%, and the long-run variance estimate is 2.66x the daily variance. A 648-case sweep over base Sharpe, autocorrelation, smoothing, volatility clustering, sample length, and Monte Carlo path count produced finite metrics throughout and the requested number of sample paths in every case. The audit also found that the exported evaluate({}) API fell back to zeros instead of the UI defaults; the lab now cleans missing parameters the same way the sliders do.

Base Sharpe: 0.65 Autocorrelation: 25% Return smoothing: 40% Vol clustering: 40% Sample length: 6y Monte Carlo paths: 300

Sampling-clock Sharpe Daily sqrt rule HAC Sharpe Autocorrelation Variance scaling

Deterministic synthetic experiment. The HAC estimate uses a finite Newey-West/Bartlett long-run variance estimate on daily returns. The aggregation chart uses non-overlapping h-day sums, so the far-right points are deliberately noisier in short samples.

Start with the default settings. The daily-clock Sharpe is much higher than the HAC Sharpe. The reason is visible in two places: the lag-1 autocorrelation is large, and the volatility of multi-day sums rises faster than the IID square-root calculation would predict.

Now set return smoothing to zero and autocorrelation to zero. The daily-clock Sharpe and HAC Sharpe move much closer together. The Sharpe-by-clock line becomes flatter. The square-root rule is no longer doing much violence because its assumptions are closer to true.

Next make autocorrelation negative. The naive daily annualization can now be too pessimistic rather than too optimistic. That is a useful corrective to the usual moral tale. The problem is not that \(\sqrt{252}\) always exaggerates performance. The problem is that it silently assumes away the covariance structure.

Finally raise vol clustering while leaving smoothing low. The point estimate may not move much. But the Monte Carlo panel widens, because a few volatility episodes can dominate a short backtest. Conditional heteroskedasticity is often less about the numerator of the Sharpe ratio and more about the confidence you should have in the estimate.

Smooth Returns Are Not Free Risk Reduction

Return smoothing deserves special suspicion because it can make a risky stream look quiet without changing the underlying economics.

Suppose true returns are \(x_t\) but reported returns are a smoothed average:

\[r_t^{\mathrm{reported}} = (1-\theta)x_t + \theta r_{t-1}^{\mathrm{reported}}.\]

The weights sum to one, so the mean is roughly preserved. Daily volatility, however, falls, and autocorrelation appears. The naive daily Sharpe ratio can rise because the denominator has been mechanically compressed.

That is not only a toy pathology. Getmansky, Lo, and Makarov study serial correlation and illiquidity in hedge fund returns, arguing that illiquid portfolios tend to have smoother reported returns, which can understate volatility and increase risk-adjusted performance measures such as the Sharpe ratio.⁴ Their model is more serious than the one-line filter above, but the intuition is the same: a stale mark can be a volatility suppressant.

The long-run variance asks a different question. It does not ask how volatile one reported day looks. It asks how much variance accumulates when days are added together. Smoothing can hide variance at the daily clock, but it cannot make the long-horizon cumulative risk disappear.

The Ratio Is a Statistic, Not a Trophy

There are three separable questions that often get collapsed into one Sharpe ratio:

What is the mean excess return per unit of one-period volatility?
How does variance accumulate across time?
How uncertain is the estimate after seeing only this sample?

The IID annualized Sharpe answers the first question and assumes the second and third are harmless. A HAC Sharpe is better for the second question. It is not a full answer to the third, especially with fat tails, regime shifts, leverage changes, option-like payoffs, and selection bias from trying many strategies.

This is why a single “corrected Sharpe” can still be misleading. A useful research report should show the clock:

the return sampling frequency and calendar;
the autocorrelation function of excess returns;
naive and long-run-variance annualized Sharpe ratios;
Sharpe by aggregation horizon, such as daily, weekly, monthly, and quarterly;
confidence intervals or bootstrap uncertainty that respect the time series;
whether marks are exchange-traded, model-priced, stale, interpolated, or otherwise smoothed;
transaction costs, capacity, drawdowns, skew, and tail exposure beside the ratio.

The last bullets are not decorative. A strategy with a clean Sharpe and hidden left-tail exposure is not clean. A strategy with a high Sharpe because it marks illiquid positions slowly is not high quality. A strategy whose Sharpe vanishes when sampled weekly was partly a clock effect.

A Better Habit

I like reporting Sharpe ratios. They compress a real tradeoff into a number people understand. The mistake is to let compression turn into erasure.

The habit I want in a quant notebook is simple:

never report a Sharpe ratio without reporting the clock that made it

If the daily, weekly, and monthly versions agree, great. The square-root rule earned its convenience. If they disagree, that disagreement is not a nuisance to hide in an appendix. It is information about the strategy, the marks, and the market microstructure underneath the return series.

The Sharpe ratio is not one number floating in space. It is a measurement taken through a clock.