A Multipurpose hybrid forecasting framework for economic stress scenarios: evidence from agriculture and energy sectors

Time series forecasting is vital across many sectors, providing critical insights for decision-making by predicting future trends from historical data. However, the complex, nonlinear nature of real-world time series and the domain-specific tailoring of existing models limit their generalizability and robustness, especially during stressed economic periods. This study aims to develop multipurpose, scenario-based hybrid forecasting models applicable to both agriculture and energy sectors, addressing the need for models that perform well under varying economic conditions. We propose two hybrid models that enhance forecasting accuracy by preprocessing data streams either through decomposition or clustering based on similarity and applying advanced forecasting techniques. Using S&P Energy (GSPE) and Agribusiness (SPGAB) indices as proxies for the energy and agriculture sectors, respectively, we conduct experiments comparing individual models such as ARIMA and LSTM with hybrid approaches. Additionally, we investigate the effectiveness of multilayer perceptron (MLP) as a post-processing tool to improve residual predictions. Our models are tested in both normal economic conditions and stressed periods, including the COVID-19 pandemic, to evaluate their robustness. Results indicate that one hybrid model consistently outperforms individual and alternative hybrid models during stable periods, while the other excels in stressed scenarios. This research contributes a novel, adaptable forecasting framework that bridges gaps in existing literature by addressing multi-domain applicability and economic stress resilience, offering practical tools for improved forecasting in agriculture and energy markets.