Adaptive Architecture and Advanced Optimization in Artificial Neural Networks for Financial Time Series Forecasting

Resumen

It is challenging to predict what will happen in the stock market because financial data isn’t always the same and changes quickly over time. However, accurate predictions are important for making smart investment decisions. This study presents a novel dynamic artificial neural network that adaptively adjusts the number of hidden-layer neurons based on the recent market conditions to enhance the prediction of stock opening and closing prices. The model uses the Fletcher-Reeves method for conjugate gradient optimization, the modified stochastic MCCV fold for validation, and the dynamic K-day sliding window for updating weights and biases over and over again. The proposed model outperformed the conventional artificial neural network along with different pairs of methods, achieving a higher correlation coefficient of 0.97292. The results highlight that the effectiveness of dynamic architectures and advanced training strategies can improve the accuracy of stock market forecasts.