Modeling Nigerian Crude Oil Prices: Comparative Evaluation of Linear, Volatility, and Deep Learning Approaches

The study focuses on determining the prediction performance of autoregressive integrated moving average (ARIMA), generalized autoregressive conditional heteroskedasticity (GARCH) and long short-term memory (LSTM) models of Nigerian crude oil prices. The comparisons among traditional linear models, volatility-based models, and deep learning approaches were done using monthly data from January 1946 to June 2025. Based on the outcomes, the ARIMA (1,1,0) model produces a fairly good linear fit for the data. However, it has relatively high prediction errors, with an MSE equal to 590.87, RMSE equal to 24.31 and MAPE equal to 28.27%. The GARCH (1,1) models exhibit successful volatility clustering capturing, with the t-distribution variant outperforming the normal specification while still yielding higher prediction error than deep learning. The MSE for the LSTM model was 112.74 with a RMSE of 10.62 and a MAPE of 15.06% which closely followed the actual price movement. The LSTM model was the best model overall in predicting the Nigerian crude oil prices. This finding was evidence that markets characterised by volatility and nonlinear dynamics are best modeled using deep learning nonlinear approaches rather than the traditional econometric models.