Description
Resample a single time series so that every resample keeps its time pattern and can take values between and beyond the observed ones, giving confidence intervals for any statistic.
The ordinary bootstrap assumes independent observations: it shuffles away time dependence, repeats some values exactly and never produces values outside the observed range. Vinod's maximum entropy bootstrap avoids all three problems. It builds a density that is uniform between midpoints of successive sorted values and exponential in the two open tails, with interval means set so the sample mean is preserved. Each resample draws quantiles from that density, sorts them and puts them back in the original series' rank order, so peaks and troughs stay where they were while the values vary. Repeating this about 999 times gives an ensemble of plausible series on which any statistic, including tail quantiles or trend estimates, can be recomputed for percentile intervals. It needs no stationarity assumption.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Record each value with its time index and sort the values.
- Compute midpoints between successive sorted values to define intervals, with open-ended first and last intervals.
- Set interval means (weights 0.75/0.25 at the ends, 0.25/0.5/0.25 inside) so the overall mean is preserved; use uniform densities inside and exponential densities in the tails.
- Draw as many uniform random numbers as there are observations, convert each to a quantile of this density, and sort them.
- Assign the sorted draws to time points in the rank order of the original series.
- Repeat about 999 times, compute the statistic on each resample and take the 2.5th and 97.5th percentiles for a 95% interval.
Sources
- Preparing for the Worst: Incorporating Downside Risk in Stock Market Investments (Vinod & Reagle, 2005)
Also known as
- ME bootstrap
- meboot
- Vinod's maximum entropy bootstrap
Categories
- Analytics & Behaviour
- Statistics & Modelling