Abstract

We present data-dependent learning bounds for the general scenario ofnon-stationary non-mixing stochastic processes. Our learning guarantees areexpressed in terms of a data-dependent measure of sequential complexity and adiscrepancy measure that can be estimated from data under some mildassumptions. We also also provide novel analysis of stable time seriesforecasting algorithm using this new notion of discrepancy that we introduce.We use our learning bounds to devise new algorithms for non-stationary timeseries forecasting for which we report some preliminary experimental results.

 

Quick Read (beta)