Tracking error measures difference, not loss. It is the volatility of your active return, the series you get by subtracting the benchmark’s return from your own on every shared date. A strategy that beats the Nifty 50 by wildly varying amounts has high tracking error. So does one that loses to it by wildly varying amounts.
How it is computed
active_r = r_strategy - r_benchmark
TE = stdev(active_r) * sqrt(N)
N is the number of return periods in a year. saral.money computes this on the dates the strategy curve and the selected benchmark curve share, then annualises. Worked example: if your monthly active returns have a standard deviation of 1.5%, tracking error is 1.5 * sqrt(12) = 5.2%.
Reading the number
| Annualised tracking error | What the strategy is doing |
|---|---|
| Under 2% | Effectively hugging the index |
| 2% to 6% | A tilt, in the range of factor and smart-beta funds |
| 6% to 12% | Genuinely active, concentrated positions |
| Above 12% | Only loosely related to the benchmark |
A 20-stock portfolio drawn from the Nifty 500 will show high tracking error by construction. That is not a fault. It is the arithmetic consequence of holding 20 names when the index holds 500.
Caveats
High tracking error is not the same as high risk of loss. It is the risk of a result that looks nothing like the index, in either direction. A fund that beat its benchmark by 15 points one year and 2 points the next has high tracking error and a happy investor.
The number depends on sampling frequency. Measured monthly, a strategy will usually report lower tracking error than the same strategy measured daily, because within-month divergence is invisible to month-end samples.
Compare tracking error only against a benchmark you would genuinely hold instead. Running a small-cap strategy against the Nifty 50 produces a large tracking error that tells you about the size gap, not about the strategy.
Tracking error alone cannot say whether the difference paid. Pair it with Information Ratio, which divides active return by exactly this number.