Öykü anlatıcısı olarak karakter : tek kişilik bir destan uyarlamasında anlatıcı

Abstract

in credit scoring statistical methods like logistic regression have been successfully used for years. in the last two decades several data mining algorithms have gained popularity and proved themselves to be useful in credit scoring. Most recent studies indicate that data mining methods are indeed good predictors of default. Additionally ensemble models which combine single models have even better predictive capability. However in most studies the rationale behind data transformation steps and selection criteria of single models while building ensemble models are unclear. in this study it is aimed to construct a fully automated comprehensive and dynamic system which gives the ability to a credit analyst to make credit decisions without any human interaction solely based on the data set. With this system it is hoped not to miss any valuable data transformation step and it is assumed that each built-in model in RapidMiner is a possible candidate to be the best predictor. To this end a model comparison engine has been designed in RapidMiner. This engine conducts almost every kind of data transformation on the data set and gives the opportunity to observe the effects of data transformations. The engine also trains every possible model on every possible transformed data. Therefore the analyst can easily compare the performances of models. As the final phase it is aimed to develop an objective method to select successful single models to build more successful ensemble models without any human interaction. However research in this direction has not been able to achieve an objective method. Hence to build ensemble models six of the most successful single models are chosen manually and every possible combination of these models are fed to another engine: ensemble model building engine. This engine tries every combination of single models in ensemble modelling and provides the credit analyst with the ability to find the best possible combination of single models and finally to reach the ultimate most successful model that can later be used in predicting thecreditworthiness of counterparties.