E. P. DADIOS, J. SOLIS 259
parameters). There is a trade-off between performance
and simplicity. To obtain high accuracy, a large number
of free parameters are needed, which again resulted in a
very complex and thus less comprehensible or readable
model. However, often the performance of a model can
actually increase with the reduction of the number of
parameters because the generalization capabilities of the
model may increase. If the model has too many parame-
ters, it tends to over-fit the training samples, TR, and
displays poor generalization on test samples, TE.
The removal of 12 input variables found to be redun-
dant resulted to 896 extracted rules only. Eliminating the
836 noise rules, the ones with single or no “hits” in the
training set, further simplified the readability of the set of
rules down to 60 rules. These 60 rules define the credit
scoring model with a classification performance of
83.67% when tested with the test set, way above the in-
dustry standard of 74% classification performance. Fi-
nally, The HFNN model developed in this research trains
faster than the traditional NN by 16 times and has better
classification accuracy of 95.33% compared to 94.67%
of traditional NN.
6. Conclusions and Recommendations
The HFNN model developed in this research to solve
credit risk management problem is capable of self-
learning similar to the traditional neural network. Subse-
quently, once trained, it is capable of discriminating the
“good” and the “bad” accounts with better accuracy
compared to the traditional NN. Unlike the neural net-
work’s “black box” configuration, which is an undesir-
able feature for credit evaluation, the HFNN model is
capable of generating the rules behind the discrimination
of each account subjected to it. The system behaves
much like a traditional fuzzy logic system in this aspect.
However, the HFNN model is better than the traditional
fuzzy logic system because of its learning capability. The
fuzzy logic system does not have this capability.
Although, this research was done for auto loan, the
Hybrid Fuzzy-Neuro Network is easily transferable to
similar loan products like mortgage loan, salary loan, and
even for credit card grants. These types of loans are the
same because they have similar input and output vari-
ables required.
In this research, the extracted rules were just listed in
the order of their importance, i.e. the most relevant rules
were listed first in the list. For future works, it is worth to
investigate some of these rules that can be fussed to-
gether to further simplify the list of rules. The output of
the developed HFNN model is limited to 2 possible val-
ues; either good or bad. By providing the data with more
than 2 outputs, say 2 additional outputs, namely: margin-
ally good and marginally bad. Marginal accounts can be
taken for a closer look before a decision is granted. This
can be considered for future study.
REFERENCES
[1] M. Bonilla, I. Olmeda and R. Puertas-An, “Application of
Genetic Algorithms in Credit Scoring,” WASL Journal,
2000.
[2] L. P. Wallis, “Credit Scoring,” Business Credit Magazine,
March, 2001.
[3] A. Fensterstock, “Credit Scoring Basics,” Business Credit,
March, 2003.
[4] A. Jost, “Neural Networks,” Credit World, Vol. 81, No. 4,
1993, p. 26.
[5] A. Fensterstock, “The Application of Neural Networks to
Credit Scoring,” Business Credit, March, 2001.
[6] G. Vasconcelos, P. Adeodato and D. S. M. P. Monterio,
“A Neural Network Based Solution for the Credit Risk
Assessment Problem,” Proceedings of the IV Brazilian
Conference on Neural Networks, Washington, 20-22 July
1999, pp. 269-274.
[7] S. A. DeLurgio and F. Hays, “Understanding the Finan-
cial Interests in Neural Networks,” WASL Journal, 2001.
[8] BrainMaker, “Credit Scoring with Brainmaker Neural
Network Software,” 2003.
http://www.calsi.com/CreditScoring.html
[9] J. Whitehouse, “Human Capabilities + Computer Auto-
mation = Knowledge,” IT Briefing, Pacific Lutheran Uni-
versity, Parkland, 2000.
[10] E. P. Dadios and D. J. Willaims, “A Fuzzy-Genetic Con-
troller for the Flexible Pole-Cart Balancing Problem,”
Proceedings of 1996 IEEE International Conference on
Evolutionary Computation, Nagoya, 20-22 May 1996, pp.
223-228.
[11] L. A. Zadeh, “A Theory of Approximate Reasoning,” In:
Machine Intelligence, John Wiley & Sons, New York,
1979, pp. 149-194.
[12] I. Iancu, “A Mamdani Type Fuzzy Logic Controller,” In:
E. P. Dadios, Ed., Fuzzy Logic: Controls, Concepts,
Theories and Applications, InTech Croatia, Rijeka, 2012,
pp. 55-54.
[13] A. Achs, “From Fuzzy Datalog to Multivalued Knowl-
edge-Base,” In: E. P. Dadios, Ed., Fuzzy Logic: Algo-
rithms, Techniques and Implementations, InTech Croatia,
Rijeka, 2012, pp. 25-54.
[14] E. P. Dadios, K. Hirota, M. Catigum, A. Gutierrez, D.
Rodrigo, C, San Juan and J. Tan, “Neural Network Vision
Guided Mobile Robot for Driving Range Golf Ball Re-
triever,” Journal of Advanced Computational Intelligence
and Intelligent Informatics, Vol. 10, No. 4, 2006, pp. 181-
185.
[15] R. Gustilo and E. P. Dadios, “Optimal Control of Aqua-
culture Prawn Water Quality Index Using Artificial Neu-
ral Networks,” Proceedings of the 5th IEEE International
Conference on Cybernetics and Intelligent Systems and
the 5th IEEE International Conference on Robotics,
Automation and Mechatronics, Qingdao, 17-19 Septem-
Copyright © 2012 SciRes. IIM