Universal Logarithmic Score Training

Background: A unified probabilistic framework for forecasting requires that predictive models be trained and assessed directly through the logarithmic score for both classification and regression tasks.

Question / Future Work: The authors point out that while classification naturally aligns with the logarithmic score (via cross-entropy), regression/continuous-valued forecasting does not consistently adopt this standard. A future direction is to fully establish the logarithmic score as the universal standard for training and evaluation in all probabilistic forecasting, thereby making uncertainty quantification and predictive skill central design principles for AI-based forecasting, similar to its success in classification.