Predicción de la producción de rosas mediante aprendizaje automático: Una comparación entre modelos fenológicos y autorregresivos

Abstract

This research compares the effectiveness of two modeling approaches for predicting rose production: one based on traditional phenological counts and another purely autoregressive, using classical Machine Learning (ML) algorithms as a low-cost alternative. Due to the limited availability of historical data, synthetic datasets were generated to preserve the seasonal and cyclical patterns of Colombia’s floriculture sector. MLP, LSTM, and XGBoost models were evaluated under a reproducible experimental design, applying cross-validation and standard error metrics (MSE, R²). Results indicate that the autoregressive XGBoost model achieved the best performance (R²=0.82), outperforming models based on phenological information (R²=0.809). These findings demonstrate that production history provides a stronger predictive signal than manual field counts, reducing dependence on subjective and labor-intensive procedures. The study offers a replicable predictive framework that enhances production planning and strengthens the competitiveness of the floriculture industry through efficient use of existing data.