Learn / Track III / lesson 1 of 5
Sharpe, Sortino and friends
The standard ways to score return per unit of risk — what each one rewards, and how each one can be gamed.
Raw return is meaningless without risk. Fifty percent a year with a 90% drawdown along the way is not better than fifteen percent with a 5% drawdown, because the second can be levered up to beat the first.
The Sharpe ratio
Excess return divided by the volatility of excess return. Annualise it by multiplying a per-period Sharpe by (with periods per year) — valid when returns are not autocorrelated.
A key property: Sharpe is invariant to leverage. Doubling position size doubles both mean and volatility. That is exactly why it is the right ranking metric: it measures the quality of the bet independently of its size.
Rough calibration for annualised, after-cost, out-of-sample Sharpe:
- 0.3–0.5: typical of a long-only equity index over long periods.
- around 1: a good systematic strategy.
- 2 and above: excellent, usually capacity-constrained.
- above 4 in a backtest: check for look-ahead bias before celebrating.
Sortino ratio
Sharpe penalises upside volatility as much as downside. Sortino replaces with downside deviation, which only counts returns below a target (often zero):
It helps with genuinely asymmetric strategies (long options, trend following), but it uses fewer data points and is noisier than Sharpe.
Calmar and return over max drawdown
Intuitive and close to how investors actually feel risk. Its weakness: maximum drawdown is a single extreme observation, it grows with the length of the sample, and it is very noisy. Never compare Calmar ratios measured over different periods.
Information ratio
Same structure as Sharpe, but against a benchmark instead of cash:
The denominator is the tracking error. This is how active managers are judged.
- Selling tail risk. Short out-of-the-money options or martingale-style averaging produce smooth returns and a high Sharpe until the rare blow-up, which may not be in the sample. Look at skewness and the worst month, not just the ratio.
- Illiquid marks. Assets marked to stale prices look less volatile than they are, inflating Sharpe. Autocorrelation in returns is the tell.
- Short samples. The standard error of an annualised Sharpe estimate is roughly for years of data. A 2-year Sharpe of 1.5 has a standard error of about 1.0 — the true value could plausibly be anywhere from 0 to 3.
Which one should you report?
Report Sharpe as the headline, because everyone can compare it. Show max drawdown and its date range alongside, because that is what someone holding the strategy will experience. Add skewness or the worst month if the strategy has an asymmetric payoff. Then check whether the Sharpe is statistically real.