Temporal-spatial dependencies enhanced deep learning model for time series forecast

Abstract

Forecasting financial time series can be challenging because of the inherent complex interplay of temporal and spatial dynamics in the data. Specifically, time series for different geographical locations often have different yet inter-dependent patterns; even for the same-location time series, patterns can vary over time. Therefore, an accurate forecasting model has to take both temporal and spatial patterns into consideration. This study proposes a new model named Temporal-Spatial dependencies ENhanced deep learning model (TSEN), which consists of multiple RNN-based layers and an attention layer: each RNN-based layer learns the temporal pattern of a specific time series at the presence of multivariate exogenous series, and the attention layer learns the spatial correlative weight and obtains the global representations simultaneously. We illustrate the use of TSEN in forecasting household leverage time series in China. We show that clustering and choosing correlative series are necessary steps to obtain accurate forecast. The results show that the proposed method captures the temporal-spatial dynamics of household leverage and generates more accurate prediction than traditional methods.