A team of researchers from IRTIC has published the article “An open, reproducible benchmark of daily CO₂ forecasting models with applications to GHG monitoring” in volume 196 of the January 2026 issue of the journal Environmental Modelling & Software. The study demonstrates that using models and hybrid models based on the Prophet statistical algorithm, designed to manage time series data with daily observations and issue a forecast of carbon dioxide concentrations in the atmosphere, can predict this phenomenon better than traditional statistical and deep learning models.
In this way, it will be possible to strengthen the capacity to respond to emerging trends in atmospheric CO₂. This will contribute to evidence-based climate assessment and ecosystem management at regional and continental scales.
The group of authors of the article consists of Pablo Catret Ruber, David García Rodríguez, Domingo J. Iglesias Fuente, Ernesto López Baeza, José Javier Samper Zapater, and Juan José Martínez Durá. Their research presents a reproducible modelling framework that integrates statistical, machine learning, deep learning, and hybrid approaches for daily-scale CO₂ prediction.
Using high-frequency data from 28 ICOS (Integrated Carbon Observing System) atmospheric stations spread across diverse European ecosystems and climates, the team assessed the model's performance by land cover, climate zone, elevation, and latitude.
All models were implemented in Python using open-source libraries, and both the code and the processed datasets are publicly available. Both the models and the hybrid models based on Prophet, such as ProphetTCN and ProHiTS, achieved higher accuracy than traditional statistical and deep learning models, and robust generalisation across biomes.
The predictive advantage of the proposal stems from the combination of an explicit seasonal structure, conservative learning dynamics and, in hybrid configurations, the ability to capture secondary signal components through complementary learners such as TCN or LightGBM.
The accuracy obtained was higher at sites located at high altitudes and latitudes, and lower on farmland and in mixed forests. The framework is transferable to other greenhouse gas monitoring contexts and is suitable for integration into environmental decision support systems and climate policy workflows.
