Gated Recurrent Unit Model Based on Expectile Regression for Time Series Data Prediction
Wisnowan Hendy Saputra, Hasri Wiji Aqsari · 2025
The development of machine learning models, especially neural networks on time series data that contain moving average elements, has many advantages. One of them is the Gated Recurrent Unit (GRU) which has advantages in terms of estimation speed and model simplicity. The GRU models are highly sensitive to outliers, which can lead to overfitting in predictions. Therefore, this research proposes the development of a GRU model that applies the expectile regression concept, which is then called Expectile Regression Gated Recurrent Unit (ER-GRU). This means applying the expectile function to the GRU architecture so that it can solve prediction problems that are sensitive to outliers. In addition, the ER-GRU prediction results are obtained through an asymmetric expectation function so that there is an adjustment to the actual robust term (adjusted robust). Based on modeling results for predicting energy futures prices, including BRE, WTI, NG, and HO, the ER-GRU model produces a good level of accuracy based on MAPE which is less than 10 %. In addition, the results of comparing the accuracy of the ER-GRU model with GRU based on RMSE and MAPE as well as RNN and LSTM show that the GRU model has the best level of accuracy.