Diseño y evaluación de estrategias de trading algorítmico en bitcoin mediante modelos predictivos de aprendizaje automático
Loading...
Download
Official URL
Full text at PDC
Publication date
2025
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
Este Trabajo de Fin de Grado explora el diseño y la evaluación de estrategias de trading algorítmico para Bitcoin utilizando modelos de predicción basados en aprendizaje automático. En el contexto de un mercado financiero altamente volátil y continuo, el estudio construye una canalización robusta y automatizada de procesamiento de datos, incorporando indicadores técnicos, variables macroeconómicas y características derivadas de series históricas de precios. Se desarrollan y prueban varios modelos, desde árboles de regresión simples hasta técnicas de ensamblado como Bagging, Random Forest, Gradient Boosting y XGBoost. Estos modelos tienen como objetivo predecir los movimientos direccionales del precio, en particular los impulsos y retrocesos, y se evalúan mediante métricas de rendimiento diseñadas específicamente, como el retorno medio por operación y la precisión direccional.
Además, el estudio simula estrategias de inversión tanto simples como adaptativas para evaluar el poder predictivo de los modelos en escenarios reales de trading. La selección final del modelo se basa en un esquema de validación cruzada con estructura temporal (forward-chaining), enfocándose no solo en la rentabilidad, sino también en la robustez frente a la volatilidad del mercado. Los resultados sientan las bases para desarrollos futuros, como la incorporación de modelos de aprendizaje profundo, modelos de volatilidad tipo GARCH y la segmentación de estrategias según el régimen del mercado (alcista, bajista o lateral). Este trabajo pone de manifiesto el potencial del aprendizaje automático para la predicción financiera, así como los desafíos de construir modelos consistentes e interpretables en entornos dinámicos e inciertos.
Abstract: This Final Degree Project explores the design and evaluation of algorithmic trading strategies for Bitcoin using machine learning prediction models. In the context of a highly volatile and continuous financial market, the study builds a robust and automated data processing pipeline, incorporating technical indicators, macroeconomic variables, and engineered features derived from historical price series. Several models are developed and tested, from simple regression trees to ensemble techniques such as Bagging, Random Forest, Gradient Boosting, and XGBoost. These models aim to predict directional movements of the price, particularly impulses and pullbacks, and are assessed through tailored performance metrics like average return per operation and directional accuracy. Furthermore, the study simulates both simple and adaptive investment strategies to evaluate the predictive power of the models in real trading scenarios. The final model selection relies on a forward-chaining cross-validation framework, focusing not only on profitability but also on robustness against market volatility. The results lay the groundwork for future developments, such as incorporating deep learning models, GARCH volatility models, and strategy segmentation according to market regime (bullish, bearish, or sideways). This work highlights the potential of machine learning for financial forecasting, as well as the challenges of building consistent and interpretable models in dynamic and uncertain environments.
Abstract: This Final Degree Project explores the design and evaluation of algorithmic trading strategies for Bitcoin using machine learning prediction models. In the context of a highly volatile and continuous financial market, the study builds a robust and automated data processing pipeline, incorporating technical indicators, macroeconomic variables, and engineered features derived from historical price series. Several models are developed and tested, from simple regression trees to ensemble techniques such as Bagging, Random Forest, Gradient Boosting, and XGBoost. These models aim to predict directional movements of the price, particularly impulses and pullbacks, and are assessed through tailored performance metrics like average return per operation and directional accuracy. Furthermore, the study simulates both simple and adaptive investment strategies to evaluate the predictive power of the models in real trading scenarios. The final model selection relies on a forward-chaining cross-validation framework, focusing not only on profitability but also on robustness against market volatility. The results lay the groundwork for future developments, such as incorporating deep learning models, GARCH volatility models, and strategy segmentation according to market regime (bullish, bearish, or sideways). This work highlights the potential of machine learning for financial forecasting, as well as the challenges of building consistent and interpretable models in dynamic and uncertain environments.












