On the chance accuracies of large collections of classifiers