Bag-of-Future: Time Series Forecasting by Continuous Function Candidates

Time series forecasting is one of the crucial tasks in many industrial fields and bears financial, resource, and control benefits. The leading solutions in this domain are often associated with machine learning approaches. The drawback in many approaches is the missing interpretability of the lookback and the predicted time series. Therefore, we propose a novel method named Bag-of-Future, to introduce interpretability in a backcast as well as forecast of the given series. Our approach aims to select and predict multiple patterns for the input series, based on continuous functions in a preselected bag of functions. Furthermore, stacking multiple encodings results in a continuous parameterization of the given dataset. The predicted bags are used to give precise forecasts and can unveil black box systems. Among synthetic as well as multiple industrial energy datasets we present competitive results in terms of performance, interpretability and pattern detection of the forecast as well as the given time series. Lastly, we present ablations on the main parameters of our approach, evaluated on multiple criterions to provide parameter recommendations.