Presentation Information
[4Yin-A-33]Residual Learning Framework for Time Series Forecasting Based on Naive Baselines
〇Chihiro Mihara1 (1. The Graduate University for Advanced Studies (SOKENDAI))
Keywords:
Time Series Forecasting,Neural Networks,Baseline Models
Neural network--based time series forecasting methods have been extensively studied, yet whether they outperform simple naive baselines---such as the average of past cycles---has not been sufficiently examined. We propose a residual learning framework that encourages neural models to surpass naive prediction on periodic time series. The framework combines naive and neural forecasts in a residual manner and introduces a penalty term into the training loss when the neural model's error exceeds that of the naive baseline. Experiments on two datasets---hourly temperature data (JMA) and hourly road occupancy data (Traffic)---with multiple neural architectures confirm that the proposed approach, particularly the weighted combination with the penalty term, tends to slightly increase the improvement rate over the naive baseline compared to using neural models alone.