Author: Manuel Díaz-Meco Terrés
Supervisor: Fernando Berzal Galiano
University of Granada – Faculty of Sciences
Double Degree in Computer Engineering and Mathematics
Main web site
This project explores the application of Artificial Intelligence models to financial markets, aiming to enhance momentum-based investment strategies.
Using historical market data from 1990 onwards, several predictive models — both classical and deep learning — are implemented, trained, and compared to evaluate their performance in predicting asset returns and constructing more profitable investment portfolios.
A complementary web application has also been developed, allowing interactive visualization of model results and portfolio simulations.
⚠️ Important:
This project is built upon a proprietary framework developed by my thesis supervisor.
Therefore, it cannot be executed or reused by third parties without access to that original codebase.
The project includes and compares the following predictive models:
- Linear and Exponential Regression
- Support Vector Regression (SVR)
- Random Forest Regressor
- Recurrent Neural Networks (RNN)
- Convolutional Neural Networks (CNN)
- Transformers for Time Series
Each model is evaluated using both statistical metrics (MSE, MAPE, R²) and financial metrics (Sharpe Ratio, cumulative return, drawdown, etc.).
The web interface provides an interactive way to explore model outputs and portfolio results.
- Stock Visualization: displays all available stocks with trained models.
- Models per Stock: shows every trained model associated with a selected stock.
- Predictions & Metrics: visualizes model predictions, quantitative metrics, and hyperparameters.
- Portfolio Simulation: allows users to configure and run investment simulations.
- Comparative Analysis: compares simulated portfolios against the S&P500 and the Clenow momentum strategy.
The comparative study revealed that Deep Learning models, especially RNNs and Transformers, captured the temporal dependencies and nonlinear relationships in financial time series more effectively than classical approaches.
Although models like SVR and Random Forest provided solid baselines, neural architectures achieved better generalization and higher predictive accuracy.
Momentum-based portfolios constructed using these models obtained superior Sharpe ratios and cumulative returns compared to both the Clenow strategy and the S&P500 benchmark.
- Integrate macroeconomic and sentiment-based indicators.
- Explore transfer learning and reinforcement learning for dynamic investment strategies.
- Extend the web app with real-time data visualization.
- Deploy the system through containerization and cloud services.
Key literature and sources consulted include:
- Andreas Clenow — Stocks on the Move
- Fernando Berzal — Neuronal Networks & Deep Learning
- Jegadeesh & Titman (1993) — Returns to Buying Winners and Selling Losers
This project was developed for academic purposes only as part of a Bachelor's Thesis.
It cannot be executed or reused without access to the base framework provided by the thesis supervisor.
© 2025 Manuel Díaz-Meco Terrés. All rights reserved.