Evaluation of credit risk using evolutionary-fuzzy logic scheme

Abstract

Credit risk evaluation systems are playing an important role in the financial decisions making by enabling faster credit decisions and diminishing possible risks. Credit scoring is the most common used technique for evaluating the creditworthiness of the credit applicants. Its objective is to classify the credit applicants according to their likely payment behavior into good or bad customer groups.

Accuracy and transparency are two important criteria that should be satisfied by any credit scoring system. Good accuracy enables correct assessment and avoids any heavy losses associated with wrong predictions while transparency is important for financial analysis and for understanding the decision process.

In this research, the transparency and the accuracy of credit scoring model have been investigated and optimized using two different fuzzy model types, namely Takagi- Sugeno (TS) and Mamdani type. The former fuzzy credit scoring model is generated using a clustering method while the latter is extracted using neural networks learning technique. The transparency and accuracy of both the resulting fuzzy credit scoring models have been simultaneously optimized using two multi-objective evolutionary techniques. The potential and the main features of the proposed modelling approaches are illustrated using two benchmark real world data sets, namely German and Australian Credit Data Sets.

Experimental results show that (TS fuzzy system) compares favorably with the other methods like NNs, RBF, “GA+SVM” while it is superior to some methods like CART, Rough sets and the popular C4.5. Moreover, unlike black box methods like ANN and GA, this method has some levels of transparency, and can be useful to do some analysis like defining the customers’ attributes that influence the system decision and approximate values of these attributes. For generating a completely transparent credit scoring model, Mamdani fuzzy system should be chosen.

Finally, a generic software called EvoFNS (Evolutionary-Fuzzy-Neuro-System) was developed which can be used for fuzzy identification (generation and optimization), prediction or classification and knowledge extraction (data mining tool). As the TS fuzzy system is more accurate than Mamdani fuzzy system and it can be used for predicting the customers’ reditworthiness while the latter, being more transparent,can be better utilized for data analysis and knowledge discovery.