Statistics5 min read

Annualisation and the Square-Root-T Rule

Annualisation is the process of scaling a statistic computed at one frequency — daily, weekly, monthly — to its annual equivalent so that strategies of different bar resolutions can be compared on a common axis. The square-root-T rule governs how volatility scales with time horizon under the assumption of independent, identically distributed returns, and it is one of the most abused identities in quantitative finance.

Mean returns scale linearly with time. Standard deviation scales with the square root of time. This asymmetry is the entire reason the Sharpe ratio gets multiplied by sqrt(T) rather than T when annualised, and it is the source of every overstated risk-adjusted return you have ever seen on a tearsheet.

sigma_annual = sigma_period * sqrt(T)

mu_annual = mu_period * T

Sharpe_annual = Sharpe_period * sqrt(T)

T is the number of periods in a year at your sampling frequency: 252 for daily bars on equities, 365 for crypto, 52 for weekly, 12 for monthly, and roughly 1560 for hourly during equity sessions. The choice of T matters less than its consistency across every statistic on the same report.

Interpreting the scaled numbers

For daily equity strategies, annualised volatility below 10% typically indicates a low-beta or heavily hedged book, 10-20% is the range of most diversified long-only equity exposure, and above 30% places the strategy in single-name concentration or leveraged-futures territory. Crypto strategies routinely report 60-100% annualised volatility, which is structurally informative rather than alarming.

Annualised Sharpe ratios above 2.0 on out-of-sample data are rare and should trigger suspicion before celebration. A 1.0 to 1.5 range is the realistic upper band for honestly evaluated systematic strategies after costs. Anything reported above 3.0 on daily-bar backtests is almost always a function of overfitting, look-ahead bias, or improper handling of execution assumptions.

The sqrt(T) multiplier rewards higher-frequency sampling. A strategy with a per-trade Sharpe of 0.1 trading 1000 times a year produces an annualised Sharpe of approximately 3.16. This is why HFT shops report numbers retail traders cannot replicate — and why annualised Sharpe is meaningless without trade count context.

What annualisation does not capture

The square-root-T rule assumes returns are independent and identically distributed. Real return series exhibit autocorrelation, volatility clustering, fat tails, and regime dependence — every one of which breaks the scaling identity. Strategies with positive return autocorrelation (trend-following, momentum) will see true annualised volatility exceed the sqrt(T) estimate. Strategies with negative autocorrelation (mean-reversion, short-volatility) will see it understated, often dramatically.

Annualised statistics also collapse the time structure of risk. A strategy that loses 40% in March and grinds back to flat by December has the same annualised return as one that drifted sideways all year, but the realised drawdown experience is incomparable. Maximum drawdown, time-to-recovery, and Calmar ratio are not subject to clean sqrt(T) scaling and must be reported on their native horizon.

Never annualise statistics from short samples. Multiplying a one-month Sharpe by sqrt(12) produces a number with confidence intervals so wide that the point estimate is essentially noise. A defensible annualisation requires at least two years of return data, and even then the standard error on Sharpe is roughly sqrt((1 + Sharpe^2/2)/N).

Higher moments do not scale by sqrt(T) at all. Skewness scales as 1/sqrt(T) and kurtosis excess scales as 1/T under the IID assumption, meaning annualised distributions appear more Gaussian than the underlying period returns. This is the mathematical reason annualised tearsheets hide tail risk: the act of aggregation itself smooths away the kurtosis you most need to see.

How Kestrel Signal handles this

Kestrel Signal reports both period-native and annualised statistics on every backtest, with the annualisation factor T displayed explicitly alongside the bar frequency. Sharpe, Sortino, and volatility figures carry their standard errors computed from the realised sample, and the tearsheet flags any annualisation performed on fewer than 504 trading days. Drawdown metrics, trade-level statistics, and tail measures are presented on their native horizon — annualisation is offered as a comparison tool, never as a substitute for the underlying distribution.