Maestría Estadística Aplicada y Ciencia de Datos · 2026

Modelado predictivo del consumo de recursos electrónicos en el Consorcio Colombia mediante COUNTER 5, SNIES, KBART y Modelos Fundacionales

Resumen

This research proposes a predictive model for electronic resource usage in the Colombia Consortium using COUNTER 5 standard data alongside physical and digital collection variables from multiple higher education institutions. The study analyzes 20 institutions over five years, comparing foundation models (Chronos 2.0, TimesFM 2.5, and TabPFN) to auto-selected classical models using PyCaret. Employing strict temporal partitioning with training data from 2020-2024 and a test period of 2025, the results show that foundation models outperform classical ones in predictive accuracy (median MASE of 0.732) and execution time. Chronos-Bolt demonstrated the best overall performance. The study also identifies that consumption dynamics are heterogeneous, with growing series benefiting more from contextual enrichment. Integrating institutional and thematic variables improves modeling precision. These results provide a solid foundation for future methodological expansions and strategic decision-making in cooperative environments, suggesting that considering external factors could further enhance model generalizability.

Palabras clave