Introduction The high volatility and non-linear dynamics of emerging financial markets, such as the Nigerian Stock Exchange (NSE), pose significant challenges to traditional linear forecasting models. This study developed and evaluated a residual-based hybrid forecasting framework designed to predict short-term stock prices for a stratified sample of three Nigerian banks. The primary objective was to enhance predictive accuracy by integrating the linear efficiency of Auto-Regressive Integrated Moving Average (ARIMA) with the non-linear learning capabilities of Support Vector Regression (SVR) and Long Short-Term Memory (LSTM) networks. Methods The study utilized daily stock data and key macroeconomic indicators spanning a 10-year period (April 2015-April 2025). A Walk-Forward Validation protocol was implemented to simulate real-world trading conditions, strictly preventing look-ahead bias. The framework operated on an additive decomposition principle, where ARIMA modeled the linear trend and the machine learning algorithms modeled the complex residuals. Predictive performance was assessed using Mean Absolute Error and Root Mean Square Error. Furthermore, the study applied SHapley Additive exPlanations (SHAP) to deconstruct the "black box" models and identify the primary drivers of volatility. Results Empirical results demonstrated that the ARIMA-LSTM hybrid consistently outperformed both the standalone ARIMA and the ARIMA-SVR models across all bank tiers. For the Tier 1 institution, the ARIMA-LSTM model achieved a Mean Absolute Error of 2.35, effectively capturing market trends despite high trading volumes. SHAP analysis revealed a distinct hierarchy of influence: large-cap stocks were predominantly driven by global macroeconomic factors, specifically the USD/NGN exchange rate and crude oil prices, while mid-cap stocks showed higher sensitivity to domestic fixed-income signals, such as bond yields and technical momentum indicators. Conclusions The study concludes that hybrid deep learning architectures significantly reduce forecasting errors in emerging markets and provides a transparent mechanism for investors to identify tier-specific economic triggers.
Durodola et al. (Fri,) studied this question.