Fuzzy Gradient-Based Assessment for Robust Forecasting in Uncertain Environments

Accurate forecasting in noisy, non-stationary settings often forces a trade-off between performance and interpretability. We propose Fuzzy Gradient-Based Assessment (FGBA), which treats fuzzy membership degrees as an interpretable basis and learns a gradient-weighted aggregation over these linguistic concepts. We examine FGBA on two datasets: NASDAQ forecasting with economic indicators and sectoral CO 2 emissions forecasting. Using rolling-origin validation, we report RMSE, MAPE, and rRMSE, and we also review calibration plots and residual diagnostics. The results indicate that FGBA performs competitively across metrics. In the emissions dataset, a tree-based baseline shows slightly lower MAPE, while FGBA remains close in performance. FGBA produces concept-level attributions that can be used to summarize which linguistic concepts contribute more to predictions and how these contributions vary over time. We also include a carbon-policy scenario to illustrate how the model could be used in practice. We discuss limitations related to membership specification and scalability in real-time settings, and we note possible extensions such as adaptive memberships and lower-latency updating.