Prediction of Locally Stationary Data Using Expert Advice

Abstract

We address the lifelong machine learning problem. Within the game-theoretic approach, in the calculation of the next prediction we use no assumptions on the stochastic nature of a source that generates the data flow: the source can be either analog, or algorithmic, or probabilistic; its parameters can change at random times; when constructing a prediction model, only structural assumptions are used about the nature of data generation. We present an online forecasting algorithm for a locally stationary time series. We also obtain an estimate for the efficiency of the proposed algorithm. The obtained estimates for the regret of the algorithm are illustrated by results of numerical experiments.