Optimización de carteras basada en la predicción de retornos mediante modelos de series temporales y aprendizaje automático
Loading...
Official URL
Full text at PDC
Publication date
2026
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
License

Citation
Abstract
Este trabajo estudia el uso de modelos predictivos para predecir los retornos futuros de distintos activos financieros. Además, se analiza si estas predicciones pueden aportar información útil para construir carteras con mejores resultados que las obtenidas mediante métodos tradicionales más sencillos, como la media histórica o el último valor observado. El análisis se aplica concretamente a un conjunto de activos financieros conocidos como ETFs. A partir de sus datos históricos se calculan retornos diarios y semanales. Esta doble frecuencia permite comparar una visión más inmediata del mercado con otra más estable y menos afectada por el ruido diario. Sobre ambas frecuencias se aplican modelos de series temporales y modelos de aprendizaje automático. En el primer caso se aplican modelos ARIMA a cada activo de forma individual. En el segundo, se prueban modelos como Ridge, Random Forest y Gradient Boosting, incorporando información de varios activos, volumen y variables temporales. Una vez comparados los modelos, las predicciones del modelo seleccionado se incorporan a un proceso de optimización de carteras basado en la teoría de Markowitz. Los resultados obtenidos se analizan junto con los de las estrategias tradicionales mencionadas anteriormente. De este modo, el trabajo evalúa si las predicciones tienen utilidad no solo desde el punto de vista estadístico, sino también desde el económico.
This thesis studies the use of predictive models to forecast the future returns of different financial assets. In addition, it analyses whether these predictions can provide useful information for building portfolios with better results than those obtained through simpler traditional methods, such as the historical mean or the last observed value. The analysis is specifically applied to a set of financial assets known as ETFs. Based on their historical data, daily and weekly returns are calculated. This double frequency makes it possible to compare a more immediate view of the market with a more stable one, less affected by daily noise. Time series models and machine learning models are applied to both frequencies. In the first case, ARIMA models are applied to each asset individually. In the second, models such as Ridge, Random Forest and Gradient Boosting are tested, incorporating information from several assets, volume and temporal variables. Once the models have been compared, the predictions from the selected model are incorporated into a portfolio optimization process based on Markowitz’s theory. The results obtained are analysed together with those of the traditional strategies mentioned above. In this way, the work evaluates whether the predictions are useful not only from a statistical point of view, but also from an economic one.
This thesis studies the use of predictive models to forecast the future returns of different financial assets. In addition, it analyses whether these predictions can provide useful information for building portfolios with better results than those obtained through simpler traditional methods, such as the historical mean or the last observed value. The analysis is specifically applied to a set of financial assets known as ETFs. Based on their historical data, daily and weekly returns are calculated. This double frequency makes it possible to compare a more immediate view of the market with a more stable one, less affected by daily noise. Time series models and machine learning models are applied to both frequencies. In the first case, ARIMA models are applied to each asset individually. In the second, models such as Ridge, Random Forest and Gradient Boosting are tested, incorporating information from several assets, volume and temporal variables. Once the models have been compared, the predictions from the selected model are incorporated into a portfolio optimization process based on Markowitz’s theory. The results obtained are analysed together with those of the traditional strategies mentioned above. In this way, the work evaluates whether the predictions are useful not only from a statistical point of view, but also from an economic one.
Description
El código desarrollado se encuentra disponible en el repositorio de GitHub asociado al trabajo: https://github.com/aleellord/TFM











