A Hybrid Model based on Ensemble Learning from Residuals for Time Series Prediction

Abstract

This study unveils an inventive algorithmic framework for a Hybrid Model, amalgamating the strengths of two forecasters to enhance the precision of time series predictions. The proposed model involves sequential training of two distinct models, each serving a unique purpose. The first model captures intricate patterns inherent in the data, while the second refines predictions by meticulously analyzing residuals. The synergy between these models significantly enhances the overall predictive capabilities of the system, showcasing its adaptability to the complex nature of time series data. Experimental results, precisely evaluated across various combinations of regression models, provide unequivocal evidence of the hybrid model's efficacy. Elasticnet Regressor (Enet) in combination with Random Forest Regressor (RFR), Gradient Boost Regressor (GBR), Ridge Regressor, and Decision Tree Regressor (DTR) has shown solid performance, with results falling within the 95% confidence interval. These outcomes showed that Enet can effectively capture intricate patterns that are displayed in sequential data.