A Scalable Learning Model for Multi-seasonal Time Series Forecasting

Time series analysis and prediction are vital for various applications in diverse domains. However, producing high-quality predictions would not be straightforward due to the complex nature of data and the requirement of human experts with domain knowledge. Conventional statistical techniques have long been employed, but they are often limited with the incompetent capability for capturing multiple seasonality. Recently, the learning-based approach using machine intelligence has gained greater attention with the feasibility of complex, multi-seasonal time series analysis. In this study, we present a machine learning model for multi-seasonal time series forecasting using deep learning structures including Long Short-Term Memory (LSTM) and Deep Neural Network (DNN). In particular, our approach is based on a stacked model that incorporates the outputs of multiple recurrent prediction components configured for distinct periods, with a data projection strategy based on seasonality information for greater accuracy and scalability. Our experimental results performed with three public time series datasets from different domains show that the proposed model is at least comparable to conventional techniques for multi-seasonal time series forecasting with a much smaller time complexity for the model build-up.