Predicción multivariable del tipo de cambio USD/PEN mediante modelos de machine learning, deep learning y fundacionales
Cargando...
Fecha
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad de Ingeniería y Tecnología
Resumen
La predicción del tipo de cambio constituye un problema fundamental para la gestión de riesgos financieros y la formulación de política monetaria en economías abiertas. En el caso peruano, el tipo de cambio USD/PEN representa una variable de alta sensibilidad debido a la dolarización parcial del sistema financiero y la dependencia de exportaciones primarias. El objetivo de esta investigación fue diseñar, implementar y evaluar un ensemble heterogéneo de modelos de machine learning para la predicción one-step-ahead del tipo de cambio USD/PEN. La metodología integró siete modelos de cuatro familias arquitectónicas: ARX, CatBoost, LSTM, NBEATSx, TiDE y dos modelos fundacionales preentrenados (TimesFM, MOIRAI). El diseño experimental incorporó un protocolo anti-leakage mediante rezagos temporales, rolling forecast sobre un holdout de 60 días, y cinco estrategias de ensemble. El dataset comprendió 8,201 observaciones diarias del período 1994-2025 con 13 variables exógenas. Los resultados demostraron que cinco modelos superaron consistentemente al benchmark naive, con errores absolutos escalados entre 0.9083 y 0.9699. TimesFM alcanzó el mejor desempeño en error (0.9083), mientras CatBoost logró la mayor precisión direccional (60Los hallazgos permiten concluir que es viable superar el benchmark de caminata aleatoria mediante machine learning, refutando la versión fuerte de la hipótesis de mercados eficientes para este par de divisas. La disociación entre métricas de error y direccionalidad tiene implicaciones prácticas: TimesFM resulta preferible para cobertura cambiaria, mientras CatBoost ofrece ventajas para trading direccional.
Exchange rate forecasting constitutes a fundamental problem for financial risk management and monetary policy formulation in open economies. In the Peruvian context, the USD/PEN exchange rate represents a highly sensitive variable due to partial dollarization of the financial system and dependence on primary exports. The objective of this research was to design, implement, and evaluate a heterogeneous ensemble of machine learning models for one-step-ahead prediction of the USD/PEN exchange rate. The methodology integrated seven models from four architectural families: ARX, CatBoost, LSTM, NBEATSx, TiDE, and two pretrained foundation models (TimesFM, MOIRAI). The experimental design incorporated an anti-leakage protocol through systematic temporal lags, rolling forecast over a 60- day holdout period, and five ensemble strategies. The dataset comprised 8,201 daily observations from 1994-2025 with 13 exogenous variables. Results demonstrated that five models consistently outperformed the naive benchmark, with mean absolute scaled errors between 0.9083 and 0.9699. TimesFM achieved the best error performance (0.9083), while CatBoost achieved the highest directional accuracy (60The findings allow us to conclude that outperforming the random walk benchmark in USD/PEN prediction using machine learning is feasible, refuting the strong form of the efficient market hypothesis for this currency pair. The observed dissociation between error and directional metrics has practical implications: TimesFM is preferable for currency hedging where error magnitude matters, while CatBoost offers advantages for directional trading.
Exchange rate forecasting constitutes a fundamental problem for financial risk management and monetary policy formulation in open economies. In the Peruvian context, the USD/PEN exchange rate represents a highly sensitive variable due to partial dollarization of the financial system and dependence on primary exports. The objective of this research was to design, implement, and evaluate a heterogeneous ensemble of machine learning models for one-step-ahead prediction of the USD/PEN exchange rate. The methodology integrated seven models from four architectural families: ARX, CatBoost, LSTM, NBEATSx, TiDE, and two pretrained foundation models (TimesFM, MOIRAI). The experimental design incorporated an anti-leakage protocol through systematic temporal lags, rolling forecast over a 60- day holdout period, and five ensemble strategies. The dataset comprised 8,201 daily observations from 1994-2025 with 13 exogenous variables. Results demonstrated that five models consistently outperformed the naive benchmark, with mean absolute scaled errors between 0.9083 and 0.9699. TimesFM achieved the best error performance (0.9083), while CatBoost achieved the highest directional accuracy (60The findings allow us to conclude that outperforming the random walk benchmark in USD/PEN prediction using machine learning is feasible, refuting the strong form of the efficient market hypothesis for this currency pair. The observed dissociation between error and directional metrics has practical implications: TimesFM is preferable for currency hedging where error magnitude matters, while CatBoost offers advantages for directional trading.
Descripción
Citación
Palma Arzubialde, C. y Salazar Sedano, J. G. (2025). Predicción multivariable del tipo de cambio USD/PEN mediante modelos de machine learning, deep learning y fundacionales [Tesis de Maestría, Universidad de Ingeniería y Tecnología]. Repositorio Institucional UTEC. https://hdl.handle.net/20.500.12815/532
Colecciones
Aprobación
Revisión
Complementado por
Referenciado por
Licencia Creative Commons
Excepto donde se indique lo contrario, la licencia de este ítem se describe como info:eu-repo/semantics/openAccess

