Soft computing system for bank performance prediction

Abstract

This paper presents a soft computing based bank performance prediction system. It is an ensemble system whose constituent models are a multi-layered feed forward neural network trained with backpropagation (MLFF-BP), a probabilistic neural network (PNN) and a radial basis function neural network (RBFN), support vector machine (SVM), classification and regression trees (CART) and a fuzzy rule based classifier. Further, principal component analysis (PCA) based hybrid neural networks, viz. PCA-MLFF-BP, PCA-PNN and PCA-RBF are also included as constituents of the ensemble. Moreover, GRNN and PNN were trained with a genetic algorithm to optimize the smoothing factors. Two ensembles (i) simple majority voting based and (ii) weightage based are implemented. This system predicts the performance of a bank in the coming financial year based on its previous 2-years’ financial data. Ten-fold cross-validation is performed in the training sessions and results are validated with an independent production set. It is demonstrated that the ensemble is able to yield lower Type I and Type II errors compared to its constituent models. Further, the ensemble also outperformed an earlier study [P.G. Swicegood, Predicting poor bank profitability: a comparison of neural network, discriminant analysis and professional human judgement, Ph.D. Thesis, Department of Finance, Florida State University, 1998] that used multivariate discriminant analysis (MDA), MLFF-BP and human judgment.

Introduction

The prediction of bankruptcy for financial firms especially banks has been the extensively researched area since late 1960s [3]. Creditors, auditors, stockholders and senior management are all interested in bankruptcy prediction because it affects all of them alike [58]. The most precise way of monitoring banks is by on-site examinations. These examinations are conducted on a bank's premises by regulators every 12–18 months, as mandated by the Federal Deposit Insurance Corporation Improvement Act of 1991. Regulators utilize a six part rating system to indicate the safety and soundness of the institution. This rating, referred to as the CAMELS rating, evaluates banks according to their basic functional areas: capital adequacy, asset quality, management expertise, earnings strength, liquidity, and sensitivity to market risk. While CAMELS ratings clearly provide regulators with important information, Cole and Gunther [15] reported that these ratings decay rapidly. Fraser [18] noted that banks performed better by holding relatively more securities and fewer loans in their portfolios. Gady [19] and Fraser [18] show that core deposit funding is beneficial for banks, particularly demand deposits, which are non-interest bearing. Gady [19] has indicated that high-performance banks were able to generate more interest or non-interest income than underperforming banks. Wall [57] observed that higher profit banks relied more on equity funding. Brewer et al. [11] observed that firms used the derivative instruments to change their risk exposure. They also concluded that there was a negative correlation between risk and derivatives usage. Haslem et al. [21] determined the impact of types of strategies followed by individual banks related to the relative profitability performance. Kwast and Rose [25] employed statistical cost accounting techniques to examine the relationship between bank profitability and two dimensions of operating performance—pricing and operating efficiency.

In what follows, a brief review of the applications of statistical and intelligent techniques to bankruptcy prediction problem in banks is presented. Altman [3] pioneered the work of using financial ratios and MDA to predict financially distressed firms. He used the following financial ratios: (i) working capital/total assets, (ii) retained earnings/total assets, (iii) Earnings before interest and taxes/total assets, (iv) market value of equity/total debt and (v) sales/total assets. Other statistical techniques such as regression analysis [24], logistic regression [36] were also employed in the past. These techniques typically make use of the company's financial data to predict the financial state of the company (healthy, distressed, high probability of bankruptcy). However, the usage of MDA or statistical techniques, in general, relies on the restrictive assumption on linear separability, multivariate normality and independence of the predictive variables [23], [33], [34]. Unfortunately, many of the common financial ratios violate these assumptions.

Tam [52] explored a neural network approach for this problem and compared its performance with that of MDA, logistic regression, k-nearest neighbour (k-NN) method and ID3. He concluded that neural network outperformed all of them. Tam and Kiang [53] found that a neural network outperformed statistical methods and decision trees. As a result, many researchers view the neural network as an attractive alternative to statistical techniques for bankruptcy prediction. Salchenberger et al. [45] reported that the neural network produced fewer or equal number of total errors, type I errors, and type II errors for each of the forecast periods in consideration compared to the logit model. Wilson and Sharda [58] compared the performance of the neural networks vis-à-vis the discriminant analysis (DA) on the same data set used by Altman [3]. They found that MLFF-BP outperformed the DA. Lee et al. [26] applied three different hybrid neural network architectures viz., MDA assisted neural network, ID3 assisted neural network and a self-organizing map assisted neural network. The hybrid neural networks performed much better than the stand-alone prediction models. Further, they concluded that SOM assisted feed forward neural network outperformed other hybrids. Jo et al. [22] used MDA, case-based forecasting system (CBFS) and neural networks for predicting the bankruptcy of Korean firms. They demonstrated that neural networks outperformed MDA and CBFS. This study also revealed that CBFS was inappropriate for bankruptcy prediction. Bell [8] reported that neural networks and logistics regression performed equally well in the prediction of commercial bank failures.

Olmeda and Fernandez [35] solved the bankruptcy prediction problem for Spanish banks by considering the following financial and economic ratios, viz., (i) current assets/total assets, (ii) current assets-cash/total assets, (iii) current assets/loans, (iv) reserves/loans, (v) net income/total assets, (vi) net income/total equity capital, (vii) net income/loans, (viii) cost of sales/sales and (ix) cash flow/loans. They employed MLFF-BP, logistics regression, multivariate adaptive splines (MARS), C4.5 and MDA as stand-alone models as well as in various combinations in the multiple voting scheme devised by them to construct an ensemble system. They found that neural networks outperformed all other models in the stand-alone mode and the combination of neural network, logistic regression, C4.5 and MDA performed the best among all the combinations. In another study, Alam et al. [2] used fuzzy clustering and two self-organizing neural networks to identify potentially failing banks. The results showed that both the fuzzy clustering and self-organizing neural networks are promising tools in the identification of potentially failing banks. McKee [30] employed rough set theory to predict corporate bankruptcy and concluded that it significantly outperformed a recursive-partitioning model. Atiya [5] reviewed the applications of the prediction techniques including neural networks to the bankruptcy prediction problem and proposed new financial indicators, which he used in the design of a new neural network model. Shin and Lee [47] applied SVM to the problem of corporate bankruptcy prediction and concluded that SVM outperformed the MLFF-BP in terms of accuracy and generalization, as the training dataset size got smaller.

Ahn et al. [1] proposed hybrid models combining rough sets and MLFF-BP for bankruptcy prediction in Korean firms. They observed that the hybrid models with feature selection and sample size reduction aspects yielded better solutions compared to MLFF-BP and DA. Park and Han [37] employed k-nearest neighbour (k-NN) weighted with analytical hierarchy process (AHP) for predicting bankruptcy in Korean firms. They used case-based reasoning for indexing and retrieving similar cases. They concluded that the weighted k-NN model outperformed other models. Shin and Lee [46] proposed a genetic algorithm based method for firm bankruptcy prediction. The rules generated by genetic algorithm were comprehensible and yielded an accuracy of 80.8% and could learn the linear relationships between input variables. Baek and Cho [6] proposed the auto-associative neural network (AANN) for bankruptcy prediction in Korean firms and observed that AANN outperformed MLFF-BP. Cielen et al. [13] employed data envelopment analysis (DEA) for predicting bankruptcy in Belgian banks and concluded that DEA outperformed C5.0 and a combination of linear programming and discriminant analysis. Tung et al. [55] proposed a new neuro-fuzzy system, viz., generic self-organizing fuzzy neural network based on compositional rule of inference to predict bankruptcy in banks and concluded that the MLFF-BP outperformed this neuro-fuzzy system. Andres et al. [4] used a variant of additive fuzzy systems with Gaussian membership functions and normalized consequents for predicting commercial and industrial firms of Spain and concluded that they outperformed discriminant analysis and logistic regression. Ryu and Yue [44] introduced isotonic separation for predicting firm bankruptcy and concluded that it outperformed MLFF-BP, logistic regression and probit method.

The present work uses the data set used by Swicegood [51]. Hence it forms the background for the present work. He predicted the underperformance of small community banks and large regional banks based on an MLFF-BP, MDA and professional human judgment. For both the regional and community bank data, he observed that the neural network models outperformed the MDA models especially for Type I errors. He reported that the prediction for small (community) banks was less accurate as compared to large (regional) banks.

He defined Type I error as the number of ‘actually poor performance banks’ predicted as ‘adequate performance banks’ expressed as percentage of total poor performance banks and Type II error as the number of ‘actually adequate performance banks’ predicted as ‘poor performance banks’ expressed as a percentage of total adequate performance banks.

The objective of the present study is primarily to develop a new hybrid architecture in the soft computing paradigm to achieve low Type I error as well as high overall accuracy for small community banks. From the business perspective, Type I error is more detrimental than Type II error. Thus, a good prediction model should have lower Type I errors.

In order to develop a robust prediction system, a number of models taken from neural networks, statistics, decision trees and fuzzy rule based classifiers will have to be seamlessly integrated, implemented, tested and validated. Given the amount of data used by Swicegood [51], it is quite obvious that data mining can be applied to predict the bank performance with the business objective of identifying failing banks. Therefore, in this study, the CRISP-DM [16], the industry standard methodology for data mining projects, was used in the design, implementation and evaluation of the various prediction models and the ensemble.