Supervisor: Prof. Dr. Hab. inż.. Marcin Woźniak, Silesian University of Technology, Poland
Conference: IVUS 2024: Information Society and University Studies 2024
Research Overview: Conducted comprehensive research investigating the performance of Long Short-Term Memory (LSTM) neural network architectures under varying data input conditions. This study addressed critical challenges in time-series forecasting by systematically analyzing how different data volumes and architectural configurations impact model performance and predictive accuracy.
Key Contributions:
• Methodology Development: Designed and implemented extensive experiments using both TensorFlow and PyTorch frameworks, analyzing Sequential LSTM, Functional LSTM, and PyTorch LSTM architectures across multiple datasets spanning 1-20+ years of financial data
• Performance Optimization: Identified optimal data input ranges for different LSTM configurations, discovering that 10-year datasets with window size 7 achieved the best balance between accuracy and computational efficiency (MAE: 9.793, MSE: 189.373, R² score: 0.885)
• Architectural Analysis: Evaluated model performance across varying window sizes (3, 5, 7, 10, 15, 30) and data volumes, providing insights into the relationship between data quantity and prediction accuracy
• Framework Comparison: Conducted comparative analysis between TensorFlow Sequential/Functional models and PyTorch implementations, demonstrating framework-specific performance characteristics
Technical Achievements:
• Processed and analyzed over 20 years of financial time-series data (6000+ data points) from major stocks
• Implemented multiple evaluation metrics, including MAE, MSE, RMSE, MAPE, R², and accuracy to provide a comprehensive performance assessment
• Achieved peak model accuracy of 98.71% with optimal data configuration in PyTorch implementation
• Identified critical thresholds for data volume optimization, preventing overfitting while maximizing predictive performance
Impact and Applications:
The research provides valuable guidelines for practitioners implementing LSTM networks in real-world applications including stock market prediction, demand forecasting, and time-series analysis. The findings contribute to the optimization of LSTM architectures for robust predictive modeling across various domains, with specific recommendations for data preprocessing techniques and architectural configurations.