Short-term forecasting of non-stationary time series using multiple feature selection methods

Abstract

Time series forecasting has been proved to be relatively easier for stationary time series, compared to non-stationary time series. This research proposes a method to partially omit the non-stationarity of the data using prioritized sampling. Using multiple feature selection methods in combination with a random forest regressor (RFR), we aim to predict the values for a non-stationary time series. In particular, the principal component analysis (PCA), kernel PCA, incremental PCA and independent component analysis methods are used. The features extracted from these methods will be fed into an RFR both individually and combined, using the union and intersection operators. The features given by the IPCA ∪ PCA ∪ KPCA method, using prioritized sampling with multiple features per day provide the best improvement over the baseline.