Towards Improving Multivariate Time-Series Forecasting Using Weighted Linear Stacking

In this day and age, the emergence of Big Data, has made a substantial amount of data accessible across various fields. In particular, time-series data has sparked interest, with researchers and practitioners developing approaches and models in an attempt to accurately forecast such type of data. Amongst the three main forecasting approaches, that is, the Separate Model Forecasting Approach (SMFA), the Global Model Forecasting Approach (GMFA) and the Cluster-Based Forecasting Approach (CBFA), studies have showed that GMFA is the least accurate but also the least time-consuming forecasting approach. We propose a Weighted Linear Stacking (WLS) technique for increasing accuracy in the three forecasting approaches, with the highest increase observed by the GMFA, thus becoming the most accurate and viable approach for multivariate time-series forecasting. In addition, we propose two novel forecasting models, a multivariate variant of the N-BEATS model (M-N-BEATS) and a hybrid Transformer-N-BEATS (TRANS-BEATS) model, both for multivariate multi-step ahead time-series forecasting. The proposed models outperform their rivals and their forecasting performance is evaluated by multiple evaluation metrics.