Comparing the Bank Failure Prediction Performance of Neural Networks and Support Vector Machines: The Turkish Case

Abstract

Experience from the banking crises during the past two decades suggest that advanced prediction models are needed for helping prevent bank failures. This paper compares the ability of artificial neural networks and support vector machines in predicting bank failures. Although artificial neural networks have widely been applied complex problems in business, the literature utilizing support vector machines is relatively narrow and their capability for predicting bank failures is not very familiar. In this paper, these two intelligent techniques are applied to a dataset of Turkish commercial banks. Empirical findings show that although the prediction performance of the two models can be considered as satisfactory, neural networks show slightly better predictive ability than support vector machines. In addition, different types of error from each model also indicate that neural network models are better predictors.

Sažetak

Iskustvo stečeno u bankarskoj krizi u posljednja dva desetljeća upućuje na potrebu korištenja naprednih modela predviđanja u svrhu prevencije propasti banaka. Ovaj rad uspoređuje sposobnost umjetnih neuronskih mreža i strojeva s potpornim vektorima da predvide propast banaka. Iako se umjetne neuronske mreže često koriste za složene probleme u poslovanju, literatura koja spominje strojeve s potpornim vektorima je relativno malobrojna a njihova sposobnost predviđanja propasti banaka nije previše poznata. U ovom radu su ove dvije inteligentne tehnike primijenjene na sklop podataka turskih komercijalnih banaka. Empirijski rezultati pokazuju da iako se predviđanje dvaju modela može smatrati zadovoljavajućim, neuronske mreže pokazuju nešto bolju sposobnost predviđanja od strojeva s potpornim vektorima. Osim toga, različite vrste grešaka u svakom modelu također ukazuju na to da su modeli s neuronskim mrežama bolji prediktori.

Key words:

• Current account
• Balance of payments
• ARDL
• Bound Test

Ključne riječi::

• Propast banaka
• ANN (umjetne neuronske mreže)
• SVM (strojevi s potpornim vektorima)
• Turska

JEL classification::

• C38
• C45
• F32
• N5