AutoFAR: An Intelligent Hybrid System for Boosting Autoformer Forecasts

• Filipe C. de L. Duarte UFPB
• Ruam E. R. C. Pastor UFPE
• Luís G. B. Amaral UFPE
• Paulo S. G. de Mattos Neto UFPE
• Jair P. de Sales UFCA

Resumo

Time series forecasting is essential for many real-world applications, spanning healthcare, climatology, finance, energy, manufacturing, etc.. Machine Learning (ML) models have garnered significant attention for their ability to capture complex, nonlinear patterns in temporal data, an area where traditional linear statistical methods often face limitations. The literature highlights that Autoformers have gained significant relevance due to their ability to effectively model autocorrelation structures in time series while performing trend and seasonality decomposition. However, recent studies have presented mixed results regarding its performance. This can occur because nonlinear models often struggle to simultaneously capture linear and nonlinear temporal patterns effectively. This paper proposes a novel intelligent hybrid system named AutoFAR to address this limitation. The proposal enhances Autoformer forecasting by modeling the residual series using an Autoregressive Integrated Moving Average (ARIMA) model. Initially, Autoformer is used to forecast the time series under study. Subsequently, its residuals are analyzed to identify any underlying linear patterns. If significant linear patterns exist, an ARIMA model is fitted to these residuals. The final forecast is obtained by adding the outputs from the Autoformer and ARIMA models. An experimental evaluation was conducted with five well-known datasets. The results show that AutoFAR consistently outperforms both Autoformer and well-established forecasting models, Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), Random Forest, eXtreme Gradient Boosting (XGBoost), and Transformers. For four out of five time series, AutoFAR achieves significant improvements, indicating that intelligently modeling linear components left by the Autoformer is a promising strategy.