Comparación del desempeño de modelos de Machine Learning para la predicción del crecimiento económico de Ecuador utilizando variables macroeconómicas (2000 2024)

Abstract

Abstract: This study compares Machine Learning models for forecasting Ecuador s quarterly real GDP growth over 2000 2024 using aggregated macroeconomic variables. A quarterly dataset is built from official sources (Central Bank of Ecuador and INEC), and the series are transformed into growth rates using log differences. The task is defined as a forecasting problem with a fixed two-quarter horizon (h = 2) and evaluated out-of-sample with MAE, RMSE, MAPE, and MASE. Traditional benchmarks are estimated (Naïve, SARIMA, SARIMAX, and Ridge regression). The SARIMAX specification combines GDP s own time-series dynamics (autoregressive and moving-average components) with exogenous macroeconomic predictors. Results show that tree-based models improve accuracy: Random Forest and especially XGBoost achieve the lowest forecast errors. In particular, XGBoost reduces MAE by about 7 8% relative to SARIMAX and by roughly 30% compared to the Naïve benchmark. Overall, the findings suggest that combining real-activity and external-sector indicators with Machine Learning techniques can enhance short-term GDP forecasting performance for Ecuador.