Machine Learning for Forecasting: Size Matters

Machine learning has been increasingly applied to solve forecasting problems. Classical forecasting approaches, such as ARIMA or exponential smoothing are being replaced by machine learning regression algorithms, such as XGBoost, Gaussian processes or deep learning. However, despite the increasing attention, there are still doubts about the forecasting performance of machine learning methods. Makridakis, one of the most prominent names in the forecasting literature, has recently presented evidence that classical methods systematically outperform machine learning approaches for univariate time series forecasting [1]. This includes algorithms such as the LSTM, multi-layer perceptron or Gaussian processes.