Aplicación de modelos de machine learning para predicciones de exportaciones para el mercado floricultor colombiano

Abstract

This applied research project implemented, evaluated, and compared two machine learning models –Random Forest and XGBoost– to predict the monthly volume of Colombian flower exports to its three main destination markets: the United States, the United Kingdom, and Canada. The study integrated macroeconomic variables from Colombia and each importing country (gross domestic product, inflation, and exchange rate), together with seasonality and time-lag variables, over a destination-country/month panel of 396 observations covering the 2015–2025 period. The development was carried out in Python on Google Colab, following a methodology that included cross-validation for hyperparameter optimization, chronological data partitioning, and formal statistical validation through the Wilcoxon signed-rank test. Results show that XGBoost achieved a slightly better point performance than Random Forest on three of the four evaluated metrics (MAE, RMSE, and R²), although the Wilcoxon signed-rank test, applied to a 72 observation test set, confirmed that this difference is not statistically significant. Variableimportance analysis, complemented with SHAP values, showed a statistically significant agreement (ρ = 0.663; p = 0.0135) that both models rank the importing market’s economic size and demand seasonality as the most explanatory factors, while Colombia’s own macroeconomic conditions carry comparatively less weight. The project demonstrates the technical feasibility of applying ensemble learning algorithms to an agricultural foreigntrade problem, and its main contribution is a complete, reproducible methodology that can be reused by PROCOLOMBIA, Colombia’s Ministry of Commerce, Industry and Tourism, or by future research within the Colombian flower-growing sector.