Please use this identifier to cite or link to this item:
https://repository.unad.edu.co/handle/10596/81189Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor.advisor | Pineda Martinez, Esneider Dejesus | |
| dc.coverage.spatial | cead_-_josé_acevedo_y_gómez | |
| dc.creator | Herrera Beleño, Juan Sebastián | |
| dc.date.accessioned | 2026-05-23T16:22:54Z | |
| dc.date.available | 2026-05-23T16:22:54Z | |
| dc.date.created | 2026-02-09 | |
| dc.identifier.uri | https://repository.unad.edu.co/handle/10596/81189 | |
| dc.description.abstract | Se desarrolló un pipeline integral que incluyó descarga de datos históricos, limpieza, transformación, ingeniería de características y modelado mediante ARIMA y XGBoost. Ambos modelos fueron evaluados con RMSE, MAE y MAPE, y sus predicciones se tradujeron en un sistema de señales validado mediante backtesting frente a Buy & Hold. XGBoost mostró un desempeño predictivo superior (MAPE ≈ 2.7%), superando ampliamente a ARIMA. No obstante, el modelo híbrido con ponderación fija no mejoró estos resultados, evidenciando que un ensamble mal balanceado puede degradar al componente más fuerte. A pesar de la alta precisión estadística, la estrategia de trading generó pérdidas, principalmente por sobre-trading, ausencia de filtros de tendencia y desalineación entre métricas predictivas y objetivos financieros. Se concluye que optimizar solo el error no garantiza rentabilidad y se requiere incorporar gestión de riesgo, control operativo y costos transaccionales. | |
| dc.format | ||
| dc.title | Modelo híbrido ARIMA–XGBoost para la predicción de precios y generar señales de compraventa de activos en mercados bursátiles internacionales | |
| dc.type | Proyecto aplicado | |
| dc.subject.keywords | Maching learning | |
| dc.subject.keywords | Series de tiempo financieras | |
| dc.subject.keywords | Backtesting | |
| dc.subject.keywords | ARIMA-XGBosst | |
| dc.description.abstractenglish | A comprehensive data pipeline was developed that included historical data download, cleaning, transformation, feature engineering, and modeling using ARIMA and XGBoost. Both models were evaluated using RMSE, MAE, and MAPE, and their predictions were translated into a signal system validated by backtesting against Buy & Hold. XGBoost showed superior predictive performance (MAPE ≈ 2.7%), far outperforming ARIMA. However, the fixed-weight hybrid model did not improve these results, demonstrating that a poorly balanced ensemble can degrade the strongest component. Despite high statistical accuracy, the trading strategy generated losses, mainly due to overtrading, lack of trend filters, and misalignment between predictive metrics and financial objectives. It is concluded that optimizing error alone does not guarantee profitability and that risk management, operational control, and transaction costs must be incorporated. | |
| dc.subject.category | Ingenieria | |
| Appears in Collections: | Especialización en Ciencia de Datos y Analítica | |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| jsherrerabel.pdf | 1.33 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.