Computationally Efficient Forecasting Methods for Large-Scale Real-Time Applications

Forecasting is a key activity for any business to operate efficiently. It is becoming increasingly common to have to regularly forecast many millions of time series. The dramatic increase in the availability of large collections of time series raises the need for developing reliable efficient and automatic algorithms for forecasting. This thesis presents three such algorithms for large-scale applications based on the meta-learning approach. Each of these algorithms uses vectors of features computed from time series to guide the way the forecasts are computed. The results show the proposed algorithms perform well compared to competitive methods in large forecasting competitions.