An open, reproducible benchmark of daily CO2 forecasting models with applications to GHG monitoring

Accurate forecasting of atmospheric CO 2 concentrations supports better representation and management of environmental systems under climate change. We present a reproducible modelling framework that integrates statistical, machine learning (ML), deep learning (DL), and hybrid approaches for daily-scale CO 2 prediction. Using high-frequency data from 28 ICOS atmospheric stations across diverse European ecosystems and climates, we assess model performance by land cover, climate zone, elevation, and latitude. All models were implemented in Python with open-source libraries, and both code and processed datasets are publicly available. Prophet and Prophet-based hybrid models (e.g., ProphetTCN, ProHiTS) achieved the highest accuracy (median MAPE <0.80 %) and robust generalisation across biomes. Accuracy was highest in high-latitude and high-altitude sites, and lowest in croplands and mixed forests. The framework is transferable to other greenhouse gas monitoring contexts, and suitable for integration into environmental decision support systems and climate policy workflows. • Long-term (1-year) daily CO 2 forecasts from ICOS stations • Evaluation of statistical, ML, DL, and hybrid forecasting models • Prophet-based hybrids reach <0.80 % median MAPE across ecosystems • Open-source scripts and processed datasets are publicly available • Results inform GHG monitoring and climate policy applications