Problem
Analysts often need a confidence interval for a statistic of a time series: a mean level, or the difference between the means of two series. The classic bootstrap draws single points at random and so destroys the dependence between neighbouring observations. When values are dependent, the covariance between and is not zero, and the naive bootstrap estimate is inconsistent: it does not converge to the true parameter. The intervals you get can be wrong, and so is any decision built on them.
Illustrative example: a retailer compares the average hourly sales of two stores over a month and resamples hours independently. Daily and weekly cycles are lost, so the interval does not reflect the real uncertainty.
Solution in brief
Resample in a way that keeps the dependence. Either resample blocks of consecutive points (block bootstrap), or fit a model and resample its residuals (model-based bootstrap). A third family, bootstrap in the frequency domain, is rarely usable. The talk gives pros and cons of each and four ways to choose the block length.
Questions this case answers
- How to build valid confidence intervals for a statistic of a dependent time series when resampling single observations breaks the dependence
- How to choose the block length for a resampling scheme on dependent data
- How to build valid confidence intervals for a metric measured on a time series
- How to bootstrap data where observations depend on each other
Tools and methods
- Block bootstrap (non-overlapping, moving, stationary, circular) — resample blocks of consecutive points.
- Residual bootstrap — fit a model, resample its residuals, add them back to the predictions.
- Frequency-domain bootstrap — resample the periodogram; an idea only, no reliable implementation found.
- arch (Python) — ready implementations and automatic block-length selection.
- Dickey–Fuller test — check for stationarity.
Google or a one-time email code. No password.
