Multi-Task Learning for Time Series Forecasting Using NARMAX-LSTM

Traditional time series forecasting models based on neural networks are usually designed for a single task, which ignores the source of knowledge or information from other related tasks. On the other hand, multitask learning (MTL) offers a new perspective where one model is not only for one forecasting case or task but several tasks at the same time. The MTL model proposed in this study is based on the Nonlinear AutoRegressive Moving Average with eXogenous input (NARMAX) model and Long-Short Term Memory (LSTM) Neural Networks. The NARMAX-LSTM model is tested and validated using three different time horizons and different cases/tasks. The model forecasting performance is compared with that of three other linear and nonlinear models. The comparison results show that the multi-task NARMAX-LSTM model does not only have better prediction performance but also avoid suffering from negative transfer and overfitting.