A HYBRID AUTOREGRESSIVE-LONG SHORT-TERM MEMORY TIME SERIES MODEL FOR FORECASTING STOCK PRICES

Abstract

This study proposed a hybrid autoregressive–long short-term memory (AR–LSTM) model for forecasting Airtel’s daily adjusted closing prices from July 2002 to July 2025. The approach integrates the linear modelling capability of ARIMA with the nonlinear pattern recognition strength of LSTM to address the limitations of standalone methods in capturing complex financial time series dynamics. The Autoregressive Integrated Moving Average (ARIMA) component models the series’ linear dependencies, while the LSTM network learns the residual nonlinear structures, producing a combined forecast. Model performance was evaluated against ARIMA and standalone Long Short-Term Memory (LSTM) benchmarks using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE), alongside Diebold–Mariano tests for predictive accuracy. Empirical results showed that the AR–LSTM achieved the lowest error metrics, indicating modest predictive improvements. However, the DM tests revealed that these gains were not statistically significant at conventional levels, suggesting that improvements may reflect sample variability rather than consistent superiority. The study highlights the potential of hybrid modelling in emerging markets like Nigeria, where volatility and structural breaks are common, while noting the need for volatility-sensitive extensions such as GARCH-based hybrids to improve responsiveness during high-volatility periods.