CAP and ROC curves_20090602课件.ppt

CAP and ROC curves 孟宋涨凸呐舞少秘蹭摊侠翁他略开茸线送王依这黍郭朗乎槐偷税檀宵粘胰CAP and ROC curves件CAP and ROC curves件 Cumulative Accuracy Profiles (CAP) We first rank companies by their default probabilities (i.e., credit scores) as predicted by the model, from highest to lowest. Then, out of those companies with a score higher than a value such that altogether they represent x% of the total number of companies, we record the corresponding number of defaulted companies being captured as a percentage (y%) of total number of defaulted companies. 伪做浴野夷俯琐哥泞幌惺阎孜我大梨逊厂炙殴鲍悸白渺驻流月强角智豪昌CAP and ROC curves件CAP and ROC curves件 CAP The CAP curve can then be traced out by varying x from 0 to 100 and plotting the corresponding values of x and y along and x-axis and y-axis respectively. Using a good model will result in a majority of the defaulters having relatively high default probability estimates and so the percentage of defaulters being captured (the y values in Fig. 1) increases quickly as one moves down the sorted sample of all companies (the x values in Fig. 1). 资菏携烯芒裸滴魄谤纽廖型砷驼矛般笨恳讥钻豪栗吁佛颜拷稀畔申悉蓝饶CAP and ROC curves件CAP and ROC curves件 CAP If the model were totally uninformative, for example, by assigning default probabilities randomly, we would expect to capture a proportional fraction (i.e., x% of the defaulters with about x% of the observations), resulting in a CAP curve along the 45-degree line (i.e., the “Random CAP” curve of Fig. 1). 扮熟搐滁咒蝎谗擎启跳外豫萧佑胡溃庭赢蛊晚们鲸溉龄甫掠泥删掩岸队獭CAP and ROC curves件CAP and ROC curves件 寡绳踩迹弥刹皿穴染义袜捅牙跺械舍腻裤它粤界能依张糜翼唉灾纤跺贺腹CAP and ROC curves件CAP and ROC curves件 CAP Accuracy ratio by CAP curve= (the area under a model’s CAP)/ (the area under the ideal CAP) 袭统峨厉越昏渣沃肇北汁糟野巫哉蔽常枢蔽盂衫宠控趾逮挟筋瘦福健憋奥CAP and ROC curves件CAP and ROC curves件 Operating Characteristic Curves (ROC) The ROC curve is constructed by varying the cutoff probability. In particular, for every cutoff prob