Hyperparameter Optimization over Data Hierarchy for Model Selection in Time Series Forecasting
Time series data is sometimes associated with a hierarchy. The data at the bottom level is often sparse and incoherent (for example, in the retail domain), which makes it hard to obtain optimized forecasting models. We propose a novel model selection method for time series forecasting, that exploits the data hierarchy during the hyperparameter optimization (HPO) of the model being trained at the bottom level, by leveraging the better predictability of the higher-level aggregated time series and incorporating the prediction errors at all levels in the HPO objective. Experiments on several public hierarchical time series datasets demonstrate the efficacy of the proposed method over standard model selection techniques.