Quant Analytics: from scratch to pro · Statistics for markets · Pro
Bootstrap and robust statistics for heavy-tailed data
Market returns have fat tails and some serial dependence, which breaks the textbook formulas for standard errors. The bootstrap rebuilds the sampling distribution from the data itself, the block bootstrap respects dependence, and robust statistics such as the median and MAD limit the damage from outliers.
In this lesson
- Run an i.i.d. bootstrap and a block bootstrap and say when each is appropriate
- Build a percentile confidence interval from bootstrap replicates
- Compute a robust scale estimate with MAD x 1.4826
- Choose between mean, median and winsorised statistics on data with outliers
Read first: How sure is an average? Standard error and confidence intervals, Hypothesis tests and p-values for market data, Return distributions and fat tails
Textbook inference assumes well-behaved, independent data. Daily returns are neither: they have fat tails, volatility that clusters and a little dependence from day to day (see fat tails). The tools in this lesson make fewer assumptions. They cost some precision when the textbook case does hold, and they protect against being misled when it does not.
The bootstrap
Efron (1979) proposed a way to learn how a statistic varies without a formula for it. Treat the sample as if it were the population and draw from it again and again.
- From
nobservations, drawnwith replacement. This is one resample. - Compute the statistic (a mean, a Sharpe ratio, a drawdown) on the resample.
- Repeat several thousand times.
- The spread of the results is the statistic's standard error. The 2.5th and 97.5th percentiles form a 95% percentile interval.
For the mean the answer can be written down without simulation: the exact i.i.d. bootstrap standard error is sd x sqrt((n - 1) / n) / sqrt(n), almost the textbook figure. The bootstrap earns its place for statistics with no formula, such as the maximum drawdown, a Sharpe ratio or a rule's best parameter.
Dependence: the block bootstrap
The plain bootstrap treats the data as exchangeable. Shuffling destroys any autocorrelation and volatility clustering, so the resampled series look more random than the real one and the interval comes out too narrow. With a daily autocorrelation of 0.1 the true standard error of a mean is about sqrt((1 + 0.1) / (1 - 0.1)) times the i.i.d. figure, which is 10.6% larger.
The block bootstrap (Kunsch, 1989) resamples blocks of consecutive observations and joins them. Dependence inside a block is kept. The stationary bootstrap (Politis and Romano, 1994) draws blocks of random length with a geometric distribution, so the resampled series is itself stationary.
Block length matters. Too short and dependence is lost; too long and there are too few distinct blocks. A length of the order of n^(1/3) is a common starting point, which is about 14 days for 2,520 days of data. Check that the answer does not swing with the choice.
Worked example
Twenty daily returns, in percent, sorted: -6.0, -0.6, -0.5, -0.3, -0.2, -0.2, -0.2, 0.0, 0.1, 0.1, 0.1, 0.2, 0.3, 0.3, 0.3, 0.4, 0.4, 0.5, 0.5, 0.6. The first is a crash day.
| Statistic | All 20 days | Without the crash day |
|---|---|---|
| Mean | -0.21 | 0.09 |
| Median | 0.10 | 0.10 |
| Standard deviation | 1.40 | 0.35 |
One day moves the mean from positive to negative and quadruples the standard deviation. The median does not move.
The textbook standard error of the mean is 1.40 / sqrt(20) = 0.314. The exact bootstrap figure is 0.306. They differ only by the (n - 1) / n factor. Neither tells you that the mean is dominated by one observation. The median and the MAD do.
Robust statistics
- Median. The middle value. It ignores the size of outliers. On normal data it is less efficient: its standard error is about 1.25 times the mean's, which is 64% of the information.
- MAD (median absolute deviation). The median of
|x - median|. For the 20 days it is 0.30. To compare with a standard deviation, multiply by 1.4826, the reciprocal of the normal 75th percentile (Rousseeuw and Croux, 1993, discuss this constant and alternatives). That gives 0.445, a scale not driven by the crash day. - Winsorising. Replace the most extreme values with the nearest value kept: for example, the lowest and highest 5%. Here it moves the mean to 0.055 and the standard deviation to 0.36.
- Trimming. Drop the extreme values instead of replacing them.
A statistic is robust if a small share of bad data cannot move it far. The median tolerates up to half the data being corrupt; the mean tolerates none.
Choosing
Outliers in returns are often real, and they are the thing that hurts. Discarding the crash day makes a risk estimate look better than the world is. Use robust statistics to describe typical behaviour and to find data errors, and use the full data, with a bootstrap, to measure risk and the uncertainty of any estimate that the tail affects.
Common mistakes
- Using the i.i.d. bootstrap on autocorrelated or clustered data.
- Winsorising or trimming and then reporting the result as the risk.
- Choosing the block length to get the interval you wanted.
- Bootstrapping a statistic that depends on the extreme (the maximum, the worst day) and treating the interval as reliable. Resampling cannot produce values larger than any seen.
- Mistaking the MAD for the standard deviation without the 1.4826 factor.
In practice
Report the mean and the median together. If they disagree, look for the observation responsible before doing anything else. Use a block bootstrap for intervals on anything computed from a time series, and show how the interval changes with block length. Research on FinUties is where studies built on methods like these are published, with their method and limits stated.
Worked numbers
Each result is recomputed from the formula shown.
| What | Formula | Result |
|---|---|---|
| sample standard deviation of 20 daily returns, percent | 1.40409 | |
| textbook standard error of the mean | sd20 / sqrt(20) | 0.314 |
| exact i.i.d. bootstrap standard error of the mean | sd20 * sqrt(19 / 20) / sqrt(20) | 0.306 |
| MAD of 0.30 on the standard deviation scale | 0.30 * (1 / phi_inv(0.75)) | 0.445 |
| MAD scale constant | 1 / phi_inv(0.75) | 1.4826 |
| relative efficiency of the median on normal data | 2 / pi | 64% |
| median standard error relative to the mean's | sqrt(pi / 2) | 1.25 |
| standard error inflation with autocorrelation 0.1 | sqrt((1 + 0.1) / (1 - 0.1)) - 1 | 10.6% |
| block length of order n^(1/3) for 2,520 days | 2520 ** (1 / 3) | 14 |
Check yourself
Why is MAD multiplied by 1.4826 to estimate a standard deviation?
For normal data the MAD equals 0.6745 standard deviations (the 75th percentile of the standard normal). Dividing by 0.6745, which is multiplying by 1.4826, puts it on the same scale as the standard deviation.
Twenty daily returns include one of -6%. The mean is -0.21% and the median is +0.10%. Which describes a typical day?
The median. One outlier moves the mean by about 0.3 points (the other 19 days have a mean of 0.09%); it moves the median not at all.
Why does the i.i.d. bootstrap understate uncertainty on daily returns with positive autocorrelation?
It shuffles observations and so destroys the dependence. With autocorrelation 0.1 the true standard error of a mean is about 10% larger than the i.i.d. figure.
What does a block bootstrap do differently?
It resamples consecutive stretches of data, so within a block the order and dependence are preserved. Block length trades bias against variance.
What is the cost of using the median instead of the mean on normal data?
Efficiency. The median's standard error is about 1.25 times that of the mean (2/pi = 64% of the information).
Key terms
BootstrapStandard errorConfidence intervalStandard deviation
Sources
- Bootstrap Methods: Another Look at the Jackknife. The Annals of Statistics 7(1), 1-26 (Bradley Efron), 1979. Accessed 9 Nov 2026
- The Jackknife and the Bootstrap for General Stationary Observations. The Annals of Statistics 17(3), 1217-1241 (Hans R. Kunsch), 1989. Accessed 9 Nov 2026
- The Stationary Bootstrap. Journal of the American Statistical Association 89(428), 1303-1313 (Dimitris N. Politis, Joseph P. Romano), 1994. Accessed 9 Nov 2026
- Alternatives to the Median Absolute Deviation. Journal of the American Statistical Association 88(424), 1273-1283 (Peter J. Rousseeuw, Christophe Croux), 1993. Accessed 9 Nov 2026
- The Econometrics of Financial Markets. Princeton University Press (John Y. Campbell, Andrew W. Lo, A. Craig MacKinlay), 1997. Accessed 9 Nov 2026