Mitigating Long-Term Forecasting Bias in Time-Series Neural Networks via Ensemble of Short-Term Dependencies

Abstract

Time-series forecasting is essential for predicting future trends based on historical data, with significant applications in meteorology, transportation, and finance. However, existing models often exhibit unsatisfactory performance in long-term forecasting scenarios. To address this limitation, we propose the Time-Series Neural Networks via Ensemble of Short-Term Dependencies (TSNN-ESTD). This model leverages iTransformer as the base predictor to simultaneously train short-term and long-term forecasting models. The vanilla iTransformer’s linear decoding layer is optimized by replacing it with an LSTM layer, and an additional long-term model is introduced to enhance stability. The ensemble strategy employs short-term predictions to correct the bias in long-term forecasts. Our extensive experiments demonstrate that TSNN-ESTD reduces the MSE and MAE by 9.17% and 2.3% on five benchmark datasets.

Keywords

• time-series forecasting
• time-series neural networks
• long-term bias
• short-term dependency
• ensemble learning