Being able to make reliable predictions is a crucial task in many Artificial Intelligence problems.
In mathematical terms, a prediction can be formulated as a classification problem, where any input data is associated with a class, or as a regression task, where we search for a suitable function that fits the data.
These two classes of prediction models serve different purposes and are useful in their own regard: classification is applied to medical diagnosis, credit rating, and test grading, while regression is used to predict blood pressure, house prices, or energy consumption.
Current metrics, such as MSE for regression and AUC (Area Under the Curve) for classification, appear largely unrelated and problem-specific.
We propose a novel family of metrics, grounded in Cramér and energy distances applicable to all types of response variables: continuous, ordinal, nominal, and extendable to multivariate settings. We show the effectiveness and versatility of our metrics in a range of real applications, from finance, health care, to human resource management. We also show that metrics can be extended to assess not only accuracy but also explainability and robustness, in a joint AI assessment framework. To derive an integrated metric, we consider alternative aggregation schemes: from weighted means to decision theoretic methods.