We present online prediction methods for univariate and multivariate timeseries that allow us to handle nonstationary artifacts present in most realtime series. Specifically, we show that applying appropriate transformations tosuch time series can lead to improved theoretical and empirical predictionperformance. Moreover, since these transformations are usually unknown, weemploy the learning with experts setting to develop a fully online method(NonSTOP) for predicting nonstationary time series. This framework allows forseasonality and/or other trends in univariate time series and cointegration inmultivariate time series. Our algorithms and regret analysis subsumes recentrelated work while significantly expanding the applicability of such methods.For all the methods, we provide sub-linear regret bounds using relaxedassumptions. We note that the theoretical guarantees do not fully capture thebenefits of the nonstationary transformations, thus we provide a data-dependentanalysis of the follow-the-leader algorithm for least squares loss thatprovides insight into the success of using nonstationary transformations. Wesupport all of our results with experiments on simulated and real data.
展开▼
