An attention-guided hybrid statistical and deep learning modeling for enhanced time series forecasting: A case study of South African telecommunication companies

Abstract

Accurate stock volatility forecasting is critical for informed investment decisions and effective risk management. This study proposes an attention-guided hybrid modeling framework that integrates Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and Exponential GARCH (EGARCH) models with Long Short-Term Memory (LSTM) networks to improve volatility prediction for South African telecommunication companies (MTN SA, Telkom SA, and Vodacom SA). While GARCH-based models capture volatility clustering and asymmetries and LSTMs learn non-linear dependencies, each method on its own has limitations. We therefore develop GARCH-LSTM and EGARCH-LSTM hybrids enhanced with attention mechanisms that dynamically weight time-series features. The models are trained on one year of stock price data (July 2023–July 2024), incorporating technical indicators such as Simple Moving Average (SMA), Exponential Moving Average (EMA), and Relative Strength Index (RSI) to enrich the feature sets. Using an 80/20 train–test split and rolling window validation, performance is evaluated via Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The attention-enhanced EGARCH-LSTM achieves the lowest MAE values: 0.0780 (MTN), 0.0901 (Telkom), and 0.0905 (Vodacom). For RMSE, the lowest errors are model-specific: LSTM-Attn performs best for MTN (0.1036), GARCH-LSTM (without attention) for Telkom (0.1190), and EGARCH-LSTM-Attn for Vodacom (0.1259). These models reduce forecasting errors by up to 4.84% in MAE and 3.67% in RMSE compared to non-attention counterparts. Volatility projections show Telkom exhibits the most significant upward volatility trend rising from 2.66 to 2.99 (GARCH) and from 0.0249 to 0.0306 (EGARCH) over 10 days, while MTN and Vodacom remain more stable. Correlation analysis confirms that attention-based models provide more consistent forecasts across stocks. These findings suggest that combining econometric volatility models with deep learning (DL) and attention mechanisms yields a robust forecasting strategy, particularly well-suited for volatile emerging markets.