Uncertainty-aware forecasting of economic and socio-economic inputs for environmental and sustainability indicator assessment

Machine learning and statistical models that forecast environmental and sustainability indicators routinely use economic and socio-economic variables as inputs, yet the field offers no systematic guidance for forecasting these covariates when genuine forward-looking predictions are required. Treating future covariate values as known, as is common practice, produces overconfident predictions by ignoring a substantial and compounding source of uncertainty. This paper benchmarks eight univariate time series forecasting models across thirteen socio-economic and economic variables for thirteen EU member states, spanning measures of human well-being, ecological pressure and demographic and economic resilience. Results show that the Theta model consistently outperforms more complex neural network approaches, a finding confirmed by Friedman and Nemenyi tests. Calibrated prediction intervals show that covariate forecast uncertainty is non-negligible and grows substantially at longer horizons, particularly for volatile economic indicators. A worked example propagates simulated covariate trajectories through a downstream model of Italian municipal solid waste generation, showing that naive point-forecast propagation understates the uncertainty of the resulting sustainability prediction. Variable-specific recommendations and calibrated uncertainty estimates are provided, forming a reproducible, uncertainty-aware input layer for environmental and sustainability indicator assessments and SDG monitoring frameworks. Adopting uncertainty-aware covariate forecasting would improve the reliability of evidence used to guide environmental policy and track progress against the SDGs.