Tracking error needs an explicit convention

Start with paired active returns

Define each active return as portfolio return minus benchmark return for the same interval. Ex-post tracking error here means the sample standard deviation of that active-return series, consistent with the active-risk definition in CFA Institute's multifactor-model reading. The definition measures fluctuation around the sample mean active return, not distance from zero. It requires matched valuation times, return units and currency conventions. A portfolio total return compared with a price-only benchmark mixes income treatment with active performance. Even impeccable standard-deviation arithmetic cannot repair that mismatch, because the economic meaning was determined before the active returns reached the formula.

Centering and the sample divisor

For n paired observations, subtract their active-return mean, square each deviation, sum and divide by n minus 1 before taking the square root. NIST's sample-scale definition uses this divisor. Under independent identically distributed observations with finite variance, the resulting sample variance is unbiased for the population variance; the sample standard deviation itself is not generally unbiased. Dividing the same centered sum by n instead describes the finite supplied population or a different estimator. Both formulas are calculable, but their names and purposes differ. With fewer than two observations, the stated sample standard deviation has no defined value.

Tracking difference is not tracking error

Mean active return preserves the direction of average underperformance or outperformance. Centered tracking error removes that mean before measuring variability. Consequently, a constant shortfall can have zero tracking error while still losing ground every period. Root-mean-square active return instead measures distance from zero and retains the effect of the mean. Algebraically, its square equals mean active return squared plus the centered sum of squares divided by n. This is not identical to sample tracking error squared plus mean squared because the sample divisor differs. Cumulative tracking difference adds another choice: a difference of compounded returns is not a sum of active returns.

Four observations expose the conventions

The original synthetic hypothetical uses four monthly active returns: 0.01, 0.02, -0.01 and 0.00. Their mean is 0.005 and the centered sum of squares is 0.000500. Sample tracking error is therefore the square root of 0.000500 divided by 3, or 0.01290994. The population-divisor version is 0.01118034, while root-mean-square active return is 0.01224745. The table reconciles these calculations without changing a single observation. None is a fabricated market estimate. The differences are entirely consequences of centering and divisor choices, and rounding takes place only after each complete calculation has been evaluated.

Annualization carries assumptions

With 12 monthly periods per year, conventional annualized sample tracking error is the monthly value multiplied by the square root of 12, giving 0.04472136 in this hypothetical. This is a scale transformation, not 12 months of observed evidence. Sharpe's discussion of time aggregation explains the role of serial dependence. For stationary active returns with variance gamma0 and lag covariance gammah, the variance of a sum over k periods is k times gamma0 plus twice the sum of (k minus h) times gammah over lags. Square-root scaling omits those covariance terms and does not exactly model compounded relative wealth.

Historical dispersion versus forecast risk

Ex-post tracking error summarizes a realized sequence that may contain changing holdings and benchmark weights. Ex-ante tracking error instead uses an estimated future risk model applied to current or proposed active exposures. For a linear return model, active weights and a forecast covariance matrix produce a quadratic-form variance, but the forecast depends on how that covariance matrix was estimated. A historical standard-deviation calculator does not acquire this forward interpretation merely through annualization. Differences between ex-ante and ex-post values can reflect changing positions, estimation error, unexpected shocks or different conventions; they are not by themselves proof of an arithmetic defect.

Interpret a low number narrowly

Tracking error ignores the common component of portfolio and benchmark losses. Both can fall together while active returns remain stable. It also does not identify the probability of a large shortfall without additional distributional assumptions. Missing observations, smoothed valuations and an opportunistically selected benchmark can produce a reassuring number with a weak interpretation. A transparent comparison states the paired sample, centering rule, divisor, frequency, annualization factor and treatment of costs. Constant fees reduce mean active return without changing centered dispersion; variable costs can change both. This distinction connects tracking error to the information ratio without turning either statistic into evidence of future skill.

Primary references

These sources support methods and conventions. The example values are original constructions, not observations extracted from these documents.

• CFA Institute: Using Multifactor Models, public summary
• NIST e-Handbook: Measures of Scale
• William F. Sharpe: The Sharpe Ratio, time dependence

Related investigations