When Square-Root-of-Time Scaling Works

Why Square-Root-of-Time Scaling Needs Conditions

Capital is often estimated over a horizon different from the period used to calibrate annual losses. A familiar shortcut multiplies annual volatility by the square root of the number of years. That shortcut can be useful, but it is not a general rule for converting any annual capital number into a multi-year one. Its validity depends on the loss measure, the distributional assumptions, and especially on how losses in different periods relate.

The essential distinction is between expected loss and unexpected loss. For independent losses with the same finite variance, expected aggregate loss grows in proportion to the number of periods, while the standard deviation grows in proportion to its square root. Under a normal model, this gives a precise square-root rule for VaR capital above expected loss. Outside such conditions, the rule may be only an approximation or may not apply.

For the standard scaling result, assume the period losses \(X_1,\ldots,X_n\) are independent and identically distributed (iid), with \(E[X_t]=\mu\) and \(\operatorname{Var}(X_t)=\sigma^2<\infty\). The losses need not be normal for the mean and variance results below. Normality is an additional assumption when we use an exact normal VaR formula.

The Variance Derivation

The aggregate mean follows from linearity of expectation. Its variance follows from independence: the variance of a sum is the sum of the variances when every pair of distinct terms has covariance zero. Since the periods are identically distributed, each contributes the same mean and variance.

The covariance terms are zero because of independence. Taking the square root gives \(\operatorname{SD}(S_n)=\sigma\sqrt{n}\). This is the source of square-root-of-time scaling: it applies directly to standard deviation under these assumptions, not to the entire loss or its expected value.

When VaR Capital Scales Exactly

Suppose, in addition, that each period loss is normally distributed, \(X_t\sim N(\mu,\sigma^2)\), and the period losses are independent. Their sum is then normally distributed with mean \(n\mu\) and variance \(n\sigma^2\). Let \(z_{\alpha}\) denote the \(\alpha\)-quantile of the standard normal distribution. The aggregate VaR and capital above expected loss are

This is an exact result under the stated normal and independence assumptions. Notice that the full VaR does not scale by \(\sqrt n\): the mean component is \(n\mu\), and only the capital above the mean has the square-root factor. For non-normal iid losses, the mean and standard deviation still obey the identities above, but the exact VaR formula need not hold. A normal approximation to the aggregate may sometimes be considered for a sufficiently large number of periods, but its accuracy depends on the loss distribution and the tail level; finite variance alone does not guarantee an accurate tail approximation.

How Dependence Changes the Scaling

Independence is a substantive assumption, not a technical detail. For any finite-variance period losses, the variance of the sum includes the covariances between periods. If these covariances are positive, aggregate variance is larger than the sum of the individual variances; if they are negative, it is smaller. The square-root rule follows only when the covariance contribution is zero and the period variances are equal.

For illustration, if each period has variance \(\sigma^2\) and every distinct pair has the same correlation \(\rho\), then each pair has covariance \(\rho\sigma^2\). The variance of the \(n\)-period sum is

This equal-correlation expression requires the stated common variance and common pairwise correlation assumptions; it is not a general formula for arbitrary dependence. It shows why persistent positive association can make multi-period volatility grow faster than \(\sqrt n\). For a time series with changing covariances, each pair’s covariance must instead be included in the aggregate variance.

When the Loss Measure Is Nonlinear

Square-root scaling describes the standard deviation of a cumulative sum under particular assumptions. It does not automatically apply to a different quantity, even when annual losses are iid. For example, the largest annual loss over a horizon is a maximum, not a cumulative sum. Its distribution is obtained from the probability that every period’s loss is below a threshold, which is a different calculation.

Limits, Checks, and Exam Traps

Before using square-root-of-time scaling, identify exactly what is being scaled and verify the model behind it. The following checklist helps separate a valid calculation from an unjustified shortcut.

A common exam trap is to convert an annual VaR directly by multiplying it by \(\sqrt n\). The correction is to decompose VaR into expected loss and capital above expected loss when the normal model applies: scale the expected loss by \(n\), and scale the standard deviation component by \(\sqrt n\). Another trap is to treat “independent” as interchangeable with “identically distributed.” Independent periods with different variances still have aggregate variance equal to the sum of those variances, but there may be no single annual standard deviation to multiply by \(\sqrt n\).