Pronóstico multi-horizonte del ITCER en Ecuador mediante aprendizaje profundo con variables macroeconómicas

Abstract

El presente trabajo desarrolla y evalúa un esquema de pronóstico multi-horizonte para el Índice de Tipo de Cambio Efectivo Real (ITCER) del Ecuador, utilizando información histórica del índice y variables macroeconómicas relacionadas. La metodología incluyó modelos de referencia y modelos de aprendizaje automático y aprendizaje profundo. Se implementaron Naive Persistence, Seasonal Naive, ETS, SARIMAX, LightGBM, LSTM multi-horizonte, Temporal Fusion Transformer (TFT) y Time-series Dense Encoder (TiDE)...This study develops and evaluates a multi-horizon forecasting framework for Ecuador’s Real Effective Exchange Rate Index (REER), using historical information from the index and related macroeconomic variables. The methodology included benchmark models, as well as machine learning and deep learning approaches. Naive Persistence, Seasonal Naive, ETS, SARIMAX, LightGBM, multi-horizon LSTM, Temporal Fusion Transformer (TFT), and Time-series Dense Encoder (TiDE) models were implemented. Model performance was evaluated using MAE, RMSE, MAPE, sMAPE, and R² metrics, with a temporal data partition to prevent the use of future information during training and testing...