ON ROBUSTNESS OF ARIMA-GAS MODEL TO NON-GAUSSIAN ERRORS: THE CASE OF INTRADAILY DATA

Abstract

This article investigates the robustness of ARIMA-GAS model to mis specified errors, through simulated intradaily data. Three scenarios are involved. Scenario 1 utilizes Gaussian innovations, Scenario 2 utilizes centered and scaled Student’s t while Scenario 3 introduces asymmetric shocks by drawing innovations from a skew-normal distribution. For Gaussian errors, Classical ARIMA attains the lowest mean RMSE/MAE in this benign linear–Gaussian setting, with ARIMA–GAS a close second. For student’s t innovations, ARIMA–GAS achieves the lowest RMSE/MAE/MAPE, substantially improving on Classical ARIMA, which suffers from sensitivity to outliers and mis specified (Gaussian) tails. Pure GAS performs competitively (second among econometric models) yet combining GAS with the ARIMA backbone yields a further reduction in forecast error. LSTM forecasts are competitive and outperform ARIMA’s and GARCH’s; however, ARIMA–GAS retains a measurable edge in all three metrics, reflecting the benefit of combining statistical structure with adaptive updating when tails are heavy. For the skew normal innovations, ARIMA–GAS attains the lowest average RMSE/MAE/MAPE, improving materially over Classical ARIMA, whose Gaussian/symmetry assumptions leave it vulnerable to skewed shocks. Pure GAS is competitive, but the ARIMA backbone adds structure that reduces forecast loss further; GARCH’s volatility focus helps little with asymmetric innovations affecting the conditional mean. LSTM forecasts are very close to ARIMA–GAS (slightly higher mean loss), ARIMA–GAS preserves interpretability and achieves marginally better average accuracy. The robustness broadens its applicability across different domains and datasets, enhancing its utility in practical applications in areas as finance, economics, or environmental studies.