MANUSCRIPT OF AN ENGINEERING ARTICLE

• Olumhense Benedict Adoghe

  Achievers University Owo

  Author

The financial industry has undergone a transformative evolution over the last decade, fueled largely by the adoption of artificial intelligence (AI) and, more specifically, deep learning models. With an explosion in data availability and the growing complexity of market behaviors, traditional statistical models are no longer sufficient. Deep neural networks (DNNs) have emerged as a promising solution due to their ability to learn complex nonlinear relationships and patterns in data. In 2023, global investment in AI for financial services exceeded $11.2 billion, with 41% directed toward machine learning applications such as fraud detection, risk scoring, and predictive analytics (McKinsey, 2024).

Previous studies have explored the use of machine learning in financial modeling, particularly using support vector machines (SVM), decision trees, and ensemble models. However, these models often suffer from limitations in scalability and feature extraction when faced with unstructured or high-dimensional financial data. A study by Zhang et al. (2022) showed that DNNs outperformed SVMs by over 15% in forecasting stock market trends using news sentiment data. Similarly, Gu, Kelly, and Xiu (2020) found that multilayer perceptrons provided superior return predictions compared to linear models across 94 equity characteristics. Despite their promise, challenges in explainability, overfitting, and data imbalance remain active areas of research.

This study employed a deep feedforward neural network architecture for financial time series prediction, particularly for stock price movement classification. The dataset comprised daily trading data (open, close, high, low, volume) from 30 companies listed on the S&P 500 index over a five-year period (2018-2022), resulting in over 38,000 records. Preprocessing steps included min-max normalization and windowed time-series transformation. The network structure included three hidden layers with ReLU activation and dropout layers (20%) to prevent overfitting. The Adam optimizer was used with a learning rate of 0.001, and binary cross-entropy served as the loss function.

Model performance was evaluated using a stratified 10-fold cross-validation strategy. Accuracy, precision, recall, and F1-score were computed for each fold, and results were averaged.

• PDF