Feature-based Forecast-Model Performance Prediction

This paper introduces a novel meta-learning algorithm for time series forecasting. The efficient Bayesian multivariate surface regression approach is used to model forecast error as a function of features calculated from the time series. The minimum predicted forecast error is then used to identify an individual model or combination of models to produce forecasts. In general, the performance of any meta-learner strongly depends on the reference dataset used to train the model. We further examine the feasibility of using GRATIS (a feature-based time series simulation approach) in generating a realistic time series collection to obtain a diverse collection of time series for our reference set. The proposed framework is tested using the M4 competition data and is compared against several benchmarks and other commonly used forecasting approaches. The new approach obtains performance comparable to the second and the third rankings of the M4 competition.