Forecasting circular economy indicators: A Machine Learning study of European Union member states

Accurate forecasting of Circular Economy (CE) indicators is essential for supporting evidence-based policy development and long-term strategic planning across the European Union (EU). Reliable projections enable policymakers to anticipate future resource needs, assess the impact of interventions and design measures that accelerate the transition towards a more circular economy. This study applies Machine Learning (ML) algorithms to predict official CE indicators published by Eurostat, covering four thematic areas: production and consumption, waste management, secondary raw materials and competitiveness. 25 member states of the EU are individually modelled, using country-specific time series data to train and evaluate five ML algorithms for regression: Ridge regression, Lasso regression, Random forest, XGBoost and support vector regression. A replicable framework for CE indicator forecasting is presented to support national and EU-level policy planning and early interventions. Best practice in ML-based forecasting is demonstrated, addressing challenges such as data sparsity, non-stationarity and model overfitting. No single model consistently outperforms others, though linear models tend to provide more reliable uncertainty estimates for structurally predictable indicators. Two features was determined optimal across models, as including additional features provided minimal improvement in MAE, reflecting the constraints imposed by the limited size of the training datasets. The results show the potential and limitations of current forecasting methodologies when applied to CE indicators, emphasising the importance of representative training data and careful uncertainty quantification in policy-relevant forecasts. • ML forecasting framework developed for EU Circular Economy indicators. • Forecasting accuracy varies substantially across indicators and countries. • Structural indicators show stronger forecasting performance than waste indicators. • Limited training data constrains gains from additional predictive features.