Impact of global indices on forecasting the S&P 500 index
Year: 2025
Type: article
Abstract: This study presents a hybrid Random Forest–Long Short-Term Memory (RF-LSTM) framework for forecasting the S&P 500 Index, utilizing daily data from 25 global stock indices spanning 2003 to 2024. By combining Random Forest’s feature selection with LSTM’s deep temporal modeling, the approach reveals significant geographic asymmetries in predictive influence, with North America contributing 49%, East Asia 31%, ASEAN–Oceania 10%, Europe 6%, South Asia 3%, and Latin America 1%. At the index level, the Dow Jones Industrial Average (42%), KOSPI (18%), and Russell 2000 (15%) are identified as primary predictors, highlighting both domestic and international spillover effects. Optimized using Bayesian Optimisation, the RF-LSTM model achieves superior out-of-sample performance with an R² of 0.9952, RMSE of 16.85, MAE of 13.29, and MAPE of 4.72%, reflecting error reductions of up to 32.5% compared to baseline LSTM models. The Random Forest’s permutation importance effectively isolates high-impact indices, reducing noise and dimensionality to enhance temporal modeling accuracy. In our implementation, the residual variation left after Random Forest feature selection is further modeled using LSTM, consistent with a residual-hybrid forecasting design. These findings offer investors a robust, geographically informed tool for portfolio optimization and provide policymakers with insights into global risk transmission. The results underscore the efficacy of integrating interpretable feature selection with deep learning to advance financial forecasting in a globally interconnected market.
Authors: Seyed Mostafa Mostafavi, Ali Reza Hooman
Institution: Imperial College London
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,340
Funder: Imperial College London
OpenAlex ID: W4415071315