Towards better time series prediction with model-independent, low-dispersion clusters of contextual subsequence embeddings

摘要

Many applications in financial investment management, environmental pollution reduction, energy resource scheduling, and consumer sales promotion need to forecast future values in multivariate tie series. Despite extensive research efforts laid on the task for decades from the perspective of prediction models, this paper explores boost the prediction performance with model-independent, low-dispersion clusters. The intuition is that appropriately splitting training data could lower the prediction uncertainty, resulting in better accuracy. In the proposed multi-horizon time series prediction framework (TSP-LDC), autoencoders generate contextual embeddings that capture variables’ nonlinearity and temporal dynamics. A discriminative clustering process on the contextual embeddings produces homogeneous clusters, intending to bring better performance of prediction models. Each cluster has a corresponding prediction model responsible for making inferences for subsequences belonging to that cluster, where the discriminative probabilities of the subsequence indicate its membership. Extensive experiments demonstrate that TSP-LDC improves the performance of popular approaches of time series prediction on various datasets, proving the effectiveness of the proposed approaches.