Relative Predictive Accuracy of Machine Learning-Enhanced Long Memory Volatility Models for Modeling Nigeria Energy Data

Abstract

This study investigates the relative predictive efficacy of machine learning-augmented long memory volatility models in analyzing Nigeria's energy statistics from January 1960 to December 2024. Utilizing monthly energy pricing information sourced from the Central Bank of Nigeria, the research employs thorough econometric and computational methods to evaluate the persistence of volatility, significant structural changes, and long-term dependencies existing within the energy data of Nigeria. Initial assessments show that raw energy data display non-stationarity, structural volatility, and clustering phenomena, with the Augmented Dickey-Fuller, Phillips-Perron, and KPSS tests affirming non-stationarity in the level data and stationarity in the returns data. The analysis of structural breaks utilizing the ruptures algorithm discovers eight notable breakpoints that align with key policy changes, worldwide oil crises, and organizational transformations in Nigeria’s energy sector. To represent the noted persistence and long memory, established econometric models such as ARFIMA, FIGARCH, and HYGARCH are estimated and then combined with Artificial Neural Networks (ANN) and Support Vector Regression (SVR). The outcomes demonstrate that hybrid models significantly exceed the performance of their isolated versions, with ARFIMA–ANN and FIGARCH–ANN showing the lowest Mean Squared Errors at 0.034 and 0.035 respectively. The ANN consistently shows a greater capability to capture nonlinear volatility patterns, while SVR demonstrates a moderate level of success. The results highlight that integrating long-memory stochastic models with machine learning frameworks provides strong predictive performance for complex energy series that depend on different regimes. These findings have significant consequences for the formulation of energy policies, management of volatility, and investment strategies in Nigeria's transforming energy sector. The study concludes that econometric models enhanced by machine learning are vital for developing adaptable forecasting systems in emerging markets facing structural and policy changes.