Enhancing forecasting performance in complex commodity markets: A TimeGAN-based approach applied to oil prices

Abstract

This study addresses the challenge of limited historical data and high volatility in oil market forecasting by integrating synthetic time series generated with the generative adversarial networks (TimeGAN) model into the training process of a multilayer perceptron (MLP) neural network. Statistical and dynamic properties of synthetic and real data are compared using spectral, autocorrelation, and t-SNE analysis. Experimental results show that supplementing real data with TimeGAN-generated sequences significantly improves the MLP’s key metrics (F1-score and ROC AUC) over models trained solely on real data. TimeGAN is shown to reliably replicate the global structural and temporal features of the original data while enhancing the diversity of training samples. The findings highlight the practical value of synthetic data for boosting neural network training efficiency in economic forecasting and underscore the need for further research on validation and generative model optimisation.