Calmar Ratio and Information Ratio Explained
The Calmar ratio and the Information ratio sit at opposite ends of the performance-metric spectrum: Calmar divides return by the single worst drawdown — the most sample-length-dependent statistic in finance — while the Information ratio divides active return by tracking error and inherits the Sharpe ratio's sampling behavior. Both are routinely quoted without acknowledging that one is severely biased by the length of the track record and the other needs a decade of data to be distinguishable from zero. This guide builds both estimators precisely, quantifies their bias and variance, and covers how each is gamed.
The Calmar ratio and its dependence on sample length
The Calmar ratio is annualized return (CAGR) over the absolute maximum drawdown:
Calmar = CAGR / |max_drawdown|
The numerator is a reasonably behaved statistic; the denominator is not. Maximum drawdown is an extreme order statistic of the equity path, and its expected magnitude grows with the length of the observation window. Under a geometric Brownian motion with drift, the expected maximum drawdown increases roughly with the logarithm (and, for low-drift processes, closer to the square root) of the number of periods. Magdon-Ismail and Atiya (2004) give the closed-form expectation; the practical consequence is blunt:
- A short backtest has not yet sampled the strategy's true worst case, so its
max drawdown is too small and its Calmar is biased upward.
- A longer backtest sees deeper drawdowns, so the same strategy's Calmar
mechanically declines as you add data — even with an unchanged return process.
This makes raw Calmar ratios non-comparable across track records of different lengths. A Calmar of 1.5 over 2 years and 0.8 over 10 years can describe the same underlying strategy. Any honest comparison must hold the window fixed (the traditional convention is a trailing 36 months; the related MAR ratio uses the full record) or, better, compare against the expected drawdown for that horizon rather than the realized one.
A worked example
A strategy with 18% CAGR and a -24% deepest drawdown has Calmar 0.18/0.24 = 0.75. Now compare three strategies on identical windows:
| Strategy | CAGR | Max drawdown | Calmar | Recovery gain needed |
|---|---|---|---|---|
| A | 18% | -24% | 0.75 | +32% |
| B | 12% | -10% | 1.20 | +11% |
| C | 30% | -55% | 0.55 | +122% |
C has the highest return and the worst Calmar; its -55% drawdown requires a +122% gain to recover, and most allocators redeem long before that. B is the most tradeable on a drawdown-adjusted basis. But remember the bias caveat: if C's record is only one year long, its true Calmar is likely worse still, because its real worst case has not printed yet.
The Information ratio and its sampling error
The Information ratio is mean active return over tracking error:
IR = mean(r_p - r_b) / std(r_p - r_b)
where r_b is the benchmark return. Structurally it is the Sharpe ratio of the active-return series, so it inherits the same sampling distribution. Under IID normal active returns, the standard error of the per-period IR is approximately sqrt((1 + 0.5IR^2)/T), and after annualization the standard error of an annual IR near 0.5 is close to 1. The implication, due to Grinold's "fundamental law of active management," is sobering: to achieve IR = sqrt(IC^2 breadth), you need either real forecasting skill (information coefficient) or genuine breadth of independent bets, and you need a long sample to prove you have either. An IR of 0.5 measured over three years is statistically indistinguishable from zero.
A worked Information ratio example
Annual active returns over five years: +3%, -1%, +4%, +2%, +2%.
mean active return = (3 - 1 + 4 + 2 + 2) / 5 = 2.0%
tracking error (sample std, ddof=1) ≈ 1.87%
IR = 2.0 / 1.87 ≈ 1.07
A point estimate of 1.07 looks exceptional, but with only five observations the confidence interval comfortably includes values below 0.4. Five annual data points cannot establish skill, however flattering the ratio.
| Information ratio | Conventional reading | Reality check |
|---|---|---|
| < 0.4 | Below average | Most active managers, net of fees |
| 0.4 - 0.6 | Good | Needs ~10+ years to confirm |
| 0.6 - 1.0 | Very good | Check capacity and crowding |
| > 1.0 | Exceptional | Almost always small-sample luck |
Computing both, with sample-length awareness, in Python
import numpy as np
import pandas as pd
def calmar_ratio(returns: pd.Series, periods_per_year=252):
equity = (1 + returns).cumprod()
n_years = len(returns) / periods_per_year
cagr = equity.iloc[-1] ** (1 / n_years) - 1
dd = (equity / equity.cummax() - 1).min()
return cagr / abs(dd) if dd < 0 else np.inf
def information_ratio(returns: pd.Series, benchmark: pd.Series,
periods_per_year=252):
active = (returns - benchmark).dropna()
ir_period = active.mean() / active.std(ddof=1)
ir_annual = ir_period * np.sqrt(periods_per_year)
se = np.sqrt((1 + 0.5*ir_period**2) / active.size) * np.sqrt(periods_per_year)
return ir_annual, (ir_annual - 1.96*se, ir_annual + 1.96*se)
def expected_max_drawdown_gbm(mu, sigma, years, periods=252):
"""Rough expected MDD for GBM, to normalize Calmar across horizons."""
n = int(years * periods)
mu_p, sig_p = mu / periods, sigma / np.sqrt(periods)
# crude approximation: scales with sigma * sqrt(2 * ln(n)) for low drift
return sig_p * np.sqrt(2 * np.log(max(n, 2)))
The informationratio function returns a confidence interval, not just a point estimate; the expectedmaxdrawdowngbm helper lets you compare a realized Calmar against what drawdown you should expect for that horizon, which is the only length-fair way to read it.
The four-metric family
| Metric | Numerator | Risk denominator | Primary failure mode |
|---|---|---|---|
| Sharpe | Excess return | Total volatility | Blind to tails, easy to smooth |
| Sortino | Excess return | Downside deviation | Small-N denominator |
| Calmar | CAGR | Max drawdown | Severe sample-length bias |
| Information | Active return | Tracking error | Needs long sample; benchmark-dependent |
The Sharpe and Sortino ratios penalize dispersion; Calmar penalizes the worst path realization; the Information ratio penalizes deviation from a benchmark. A complete tear sheet reports all four because each conceals a flaw the others expose — a great Sharpe with a poor Calmar means smooth returns hiding one cliff; a great Calmar with a poor IR means a strategy that makes money but never beats the index it claims to track.
Rolling estimates expose decay
A single full-sample number for either ratio hides whether the edge is stable or decaying. Compute a rolling 12-month Calmar and a rolling IR and plot them: a healthy strategy shows a noisy but level series, while a crowded or decaying edge shows a downward drift that the full-sample average conceals. For the Information ratio in particular, a rolling estimate often reveals that apparent skill was concentrated in one or two favorable years — the hallmark of luck rather than a repeatable process. Rolling windows trade statistical power for time resolution, so use a window long enough to estimate the metric (at least 12 months, ideally 24-36) and read the trend, not the point-to-point wiggle.
How these ratios are gamed
- Calmar via short windows. Reporting a young track record before the real
drawdown has been sampled is the most common Calmar inflation, and it is often unintentional. Stress the future drawdown with Monte Carlo simulation rather than trusting the single historical path.
- Calmar via mark smoothing. Suppressing the marked depth of an illiquid
drawdown shrinks the denominator directly.
- IR via benchmark selection. A soft or non-investable benchmark inflates active
return. The benchmark is part of the metric; an IR of 0.8 versus cash and 0.8 versus the S&P 500 are unrelated claims.
- IR via fee timing and frequency. Quoting gross active returns, or a monthly IR
with fewer observations, flatters the number.
Honest limitations and when to use each
Both ratios are point estimates with wide error bars; treat a Calmar of 0.9 and 1.1, or an IR of 0.5 and 0.7, as effectively the same on typical sample sizes. Calmar reduces an entire path to one extreme statistic and ignores how often and how long the strategy is underwater — always pair it with drawdown duration and the underwater curve (see maximum drawdown). The Information ratio assumes the active-return series is stationary relative to the benchmark, which crowding and regime change routinely violate.
Practically: lead with Sharpe and Calmar for absolute-return mandates; lead with the Information ratio for benchmark-relative mandates, where allocators judge you on it directly; weight Calmar and drawdown duration most heavily when trading your own capital, since your tolerance for time underwater determines survival; and add Sortino for option-like or skewed payoffs. Always derive every ratio from a clean backtest — biases that overstate return or understate drawdown corrupt all of them simultaneously — and validate with a walk-forward procedure.
Conclusion
The Calmar and Information ratios extend the toolkit that Sharpe and Sortino begin, but each carries a specific statistical hazard: Calmar is biased by track-record length because maximum drawdown is an order statistic that grows with the sample, and the Information ratio needs a long, stationary history relative to an investable benchmark before it can distinguish skill from luck. Quote both with their limitations, normalize Calmar against the expected drawdown for the horizon, report the IR with a confidence interval, and combine them with maximum drawdown, the Sharpe and Sortino ratios, and sound position sizing to judge whether a strategy is not merely profitable but actually tradeable.