Predicting prices of the US and G7 stock indices in uncertain times: Evidence from the application of a hybrid neural network

This study investigates the application of Artificial Neural Networks (ANNs) to forecast the one-day-ahead closing price of the US and G7 indices, and makes an extended analysis of three distinct periods, namely, the pre-2008 financial crisis (2003–2007), post-crisis recovery (2009–2016), and recent economic uncertainty (2017–2022). Unlike the traditional predictive approaches, our model distinguishes itself by utilizing a hybrid ANN-based architecture that integrates variable selection and forecasting stages. The proposed model consists of two main parts: selecting relevant input variables and developing a forecasting model. In the first part, an ANN-based variable selection model is utilized to identify significant input variables based on historical market conditions that reflect economic and psychological influences over the study period. These inputs are then refined by eliminating variables with low contributions, resulting in improved model performance. In the second part, we evaluate the impact of different training algorithms, hidden layer sizes, and training data distributions on the ANN's forecasting accuracy. The findings demonstrate that ANNs can effectively forecast the S&P 500 index's and G7 indices’ prices with high accuracy, particularly when employing the Levenberg-Marquardt algorithm with a simplified model architecture. Moreover, the expanded dataset covering three distinct periods has enabled us to test the model's stability and generalization across diverse market volatility and structural conditions. The study highlights the critical role of data volume in enhancing the model's performance, confirming that extensive training data is essential for capturing the complex dynamics of market behavior.