Predicción de emisiones de CO2 en países de la OCDE mediante redes neuronales e indicadores demográficos, económicos y de economía circular

Introduction: Mitigating climate change requires accurate monitoring and prediction of CO2 emissions, particularly in OECD countries, where industrialization and economic growth have historically driven high emission levels.Objective: To predict CO2 emissions in OECD countries using a neural network approach that incorporates demographic and economic indicators, circular economy legislation, and plastics trade data.Methods: A quantitative, longitudinal and predictive design was applied to a panel of 32 OECD countries (1960–2023; 2,048 country-year observations). GRU networks with one-, three- and five-year windows were optimized with Optuna, trained with a temporal split (training up to 2018, validation 2019–2021, test 2022–2023), and complemented by post-hoc calibration and Kernel SHAP. Results were compared with persistence, fixed-effects panel, Random Forest and LSTM benchmarks and checked across five random seeds.Results: Optimization and calibration raised the test R² from 0.053 for the non-optimized network to 0.968 for the calibrated three-step model (MAE = 83.7 MtCO2). This model was the most stable across seeds (MAE = 86.7 ± 4.3 MtCO2), and window length did not yield consistent differences. Simple benchmarks, however, achieved lower errors: persistence (MAE = 11.0) and Random Forest (MAE = 45.8). Plastics trade, likely acting as a proxy for economic size, and labor force were the most influential variables, whereas circular economy laws had the weakest contribution.Conclusions: Calibrated GRU networks provide an interpretable framework that links CO2 emissions to structural demographic, economic, trade and regulatory indicators without using past emissions as inputs. However, they do not outperform simple benchmarks for short-term forecasting. Their value lies in explanation and scenario analysis rather than in point prediction.