Prediction of Locally Stationary Data Using Expert Advice

Predicting data coming from a "black box" is one of the main tasks of machine learning. In this case, no stochastic assumptions about data source is used. The data comes online as a time series consisting of pairs of the form ("signal", "response"). The data source can be an analog, deterministic (algorithmic) or stochastic process. In this case, we will use simple structural assumptions about the source of the data. In this paper, an approach is proposed in which training is performed on small subsamples of the main sample, forecasts of the constructed predictive models are combined into one common forecast based on the known aggregation methods. The general scheme of the online learning process is as follows. The learning process occurs at discrete times in steps t = 1,2,.... At the next step t, according to the data from the subsample, from the data observed in the past, a local predictive model (expert predictive strategy) is defined to obtain a response to the signal.