Modelling Crude Oil Price Volatility and Its Implications for the Nigerian Economy: A Comparative Study of Asymmetric Garch and Long Short-Term Memory (Lstm) Models

This study models the volatility of the global Brent crude oil price and examines its implications for the Nigerian economy, which depends heavily on crude oil exports for fiscal revenue and foreign exchange earnings. Using a monthly Brent crude oil spot price series (Europe) obtained from the U.S. Energy Information Administration via the Federal Reserve Economic Data (FRED) database, spanning May 1987 to July 2026 (471 price observations; 470 monthly log-returns), the study compares three GARCH-family conditional-variance models of increasing flexibility – symmetric GARCH(1,1), EGARCH(1,1) and GJR-GARCH(1,1), each with Student-t innovations – against a single-layer Long Short-Term Memory (LSTM) neural network built and validated from first principles. Descriptive statistics show pronounced excess kurtosis (9.06) and negative skewness (−0.49), with the Jarque-Bera statistic (719.84, p<.001) decisively rejecting normality. The Brent price level is non-stationary (ADF t = −1.72) while its return series is stationary (ADF t = −9.02), and both the Ljung-Box test on squared returns (Q = 210.00, p<.001) and the ARCH-LM test (LM = 171.78, p<.001) confirm strong volatility clustering. Formal structural-break tests at two independently documented oil-market events – the November 2014 OPEC non-cut decision and the March 2020 COVID-19 demand collapse – each find a highly significant post-break increase in return variance (F = 2.034 and F = 2.552, both p<.001) but no significant shift in the mean return (p = .616 and p = .839). Fitted on an 80% training window (n = 376), EGARCH(1,1) achieves the best in-sample fit by AIC (2653.32) while GARCH(1,1) achieves the best BIC (2675.44); neither asymmetric model's leverage coefficient is statistically significant (EGARCH γ = −0.063, t = −1.16; GJR-GARCH γ = 0.111, t = 0.98), although EGARCH's negative sign is stable across an independent post-2015 robustness re-fit. On a 94-month out-of-sample test window, EGARCH(1,1) achieves the lowest RMSE (427.77) and MAE (149.88), GJR-GARCH(1,1) the lowest Theil's U (0.529), and the LSTM the lowest MAPE (272.89%); a Diebold-Mariano test comparing EGARCH against the LSTM is not statistically significant (DM = −1.067, p = .286). The study concludes that Brent crude oil price volatility is highly persistent and asymmetric in sign, though not significantly so, and that a well-specified asymmetric GARCH model remains competitive with, and is not clearly outperformed by, a deep-learning alternative, a finding of direct relevance to Nigerian fiscal and reserve-management planning.