Multi-step time series forecasting

Abstract

Time series forecasting has come a long way over the years, transitioning from statistical models, to machine learning models, and now to deep learning models. Transformer-based models have held the top spots for the state-of-the-art time series benchmarking, but recent trends have been deviating from this, with the introduction of Segmented Recurrent Neural Network (SegRNN) and Patch Time Series Transformer (PatchTST). This study aims to assess the effectiveness of these models in multi-step forecasting, specifically in the ETT datasets that are commonly used in time series benchmarking tests. A comparative study was done which highlighted their nearly equivalent performance in terms of predictive accuracy. SegRNN however, distinguished itself with a substantially faster processing speed when training on the datasets. A novel hybrid model was also introduced that aims to combine the strengths from each of these models, with 10-fold cross-validation performed to ensure robustness and generalisability. Despite not achieving better benchmarks than the current models, it brought up interesting avenues for future research areas, including exploring diverse ensemble methods and refining architectures. This should ideally help to capture and learn the complex relationships between the outputs of the component models in the hybrid model. This study thus provided some insights for advancing long-term forecasting methodologies.