Classification accuracy and correlation: LDA in failure prediction

The relationship between canonical correlation and classification accuracy in linear discriminant analysis is explored mathematically. The discriminant score is assumed to conform to a uniform distribution on the interval (0, 1]. This distribution is used as a reference distribution to extract a minimum correlation for certain classification accuracy. Four different cases are analyzed. First, a case for equal group size is considered for an overall accuracy of 100%. Second, the results are generalized for unequal group size. Third, existence of discordant observations is allowed. Fourth, the effect of concentration is analyzed for the first case. The results are demonstrated by numerical examples. In addition, a sample of 2092 default and 63,072 non-default Finnish firms are used to empirically illustrate the results in the context of failure prediction. The results show that group size of default firms, number of discordant observations, and bipolar concentration of observations strongly affect both canonical correlation and classification accuracy.