This paper provides a comprehensive overview of methods used to assess the accuracy of prediction algorithms, covering various measures such as raw percentages, quadratic error measures, correlation coefficients, and information-theoretic measures like relative entropy and mutual information. The authors discuss the advantages and disadvantages of each approach and derive new learning algorithms for designing prediction systems by optimizing the correlation coefficient. They observe and prove several results relating sensitivity and specificity of optimal systems, focusing on specific problems like protein secondary structure and signal peptide prediction. The paper emphasizes the importance of considering the degree of similarity between training and test sets to avoid overestimating predictive performance due to data redundancy. It also highlights the need for performance measures that capture the full range of prediction outcomes, such as the correlation coefficient and mutual information, which are more balanced and provide a global view of prediction accuracy. The paper concludes with a discussion on the application of these measures in multi-class prediction problems and the implications for machine learning approaches, particularly neural networks.This paper provides a comprehensive overview of methods used to assess the accuracy of prediction algorithms, covering various measures such as raw percentages, quadratic error measures, correlation coefficients, and information-theoretic measures like relative entropy and mutual information. The authors discuss the advantages and disadvantages of each approach and derive new learning algorithms for designing prediction systems by optimizing the correlation coefficient. They observe and prove several results relating sensitivity and specificity of optimal systems, focusing on specific problems like protein secondary structure and signal peptide prediction. The paper emphasizes the importance of considering the degree of similarity between training and test sets to avoid overestimating predictive performance due to data redundancy. It also highlights the need for performance measures that capture the full range of prediction outcomes, such as the correlation coefficient and mutual information, which are more balanced and provide a global view of prediction accuracy. The paper concludes with a discussion on the application of these measures in multi-class prediction problems and the implications for machine learning approaches, particularly neural networks.