The evaluation of binary classification tasks in economical prediction

In the area of economical classification tasks, the accuracy maximization is often used to evaluate classifier performance. Accuracy maximization (or error rate minimization) suffers from the assumption of equal false positive and false negative error costs. Furthermore, accuracy is not able to express true classifier performance under skewed class distribution. Due to these limitations, the use of accuracy on real tasks is questionable. In a real binary classification task, the difference between the costs of false positive and false negative error is usually critical. To overcome this issue, the Receiver Ope­rating Characteristic (ROC) method in relation to decision-analytic principles can be used. One essential advantage of this method is the possibility of classifier performance visualization by means of a ROC graph. This paper presents concrete examples of binary classification, where the inadequacy of accuracy as the evaluation metric is shown, and on the same examples the ROC method is applied. From the set of possible classification models, the probabilistic classifier with continuous output is under consideration. Mainly two questions are solved. Firstly, the selection of the best classifier from a set of possible classifiers. For example, accuracy metric rates two classifiers almost equiva­lently (87.7 % and 89.3 %), whereas decision analysis (via costs minimization) or ROC analysis reveal differe­nt performance according to target conditions of unequal error costs of false positives and false negatives. Secondly, the setting of an optimal decision threshold at classifier’s output. For example, accuracy maximization finds the optimal threshold at classifier’s output in value of 0.597, but the optimal threshold respecting higher costs of false negatives is discovered by costs minimization or ROC analysis in a value substantially lower (0.477).