# Application of neural networks to business bankruptcy analysis in Thailand.

Abstract: The recent East Asian economic crisis is a lesson one can learn from the absence of effective early warning systems. To serve as a sound early warning signal, the accuracy of a failure prediction model is as important as its robustness over time. This study analyses financial and ownership variables using principal component analysis. It can reduce huge number of financial data of the business bankruptcy prediction problem. Using neural networks for bankruptcy forecasting, the obtained features are fed into neural networks as the input data. Our experiments examine the predictive performance of three neural networks: Learning Vector Quantization, Probabilistic Neural Network, and Feedforward network with backpropagation learning. All these approaches are applied to data sets of 41 Thai financial institutions for the period 1993-2003.Keywords: Bankrupcy, Neural networks, Thailand, Time series prediction, financial variables, Principal component analysis (PCA)

I. Introduction

That a lack of effective early warning systems could lead to a catastrophe is illustrated by the collapse of the Thai financial sector in 1997-1998. During the recent East Asian economic crisis, 58 out of 91 finance companies were suspended in the second half of 1997, and another 12 finance companies in 1998. Altogether, 56 finance companies were closed in 1997. On the bright side, the economic crisis enables us to examine failure prediction models for financial institutions in an emerging market economy, for which only little study has been done.

Artificial neural networks have been successfully implemented in classification and prediction problems of many fields such as business, politics, medicine and technology. The purpose of this paper is to predict the financial institution failure of Thailand by using three neural networks, i.e., Learning Vector Quantization, Probabilistic Neural Network, and Feedforward network with backpropagation learning. The results among those predictive performance are compared.

Furthermore, Principal Component Analysis (PCA) is implemented prior to neural networks as a pre-processing approach to improve the predictive performance. Basically, PCA provides a linear approximation that represents the maximum variance of the original data in a low-dimensional projection. Among a large number of financial and ownership variables, PCA can reduce the data dimensionality so that it will improve the classification performance.

The rest of the paper is structured as follows. Section 2 describes financial and ownership variables used in this study. The method of dimensionality reduction is also discussed in this Section. Section 3 reviews three neural network approaches: learning vector quantization, probabilistic neural network, and feedforward network with backpropagation learning. Section 4 shows the empirical results from our developed prediction models. Finally, Section 5 discusses the results and concludes the paper.

II. Dimensionality Reduction of Financial and Ownership Variables

A. Sample

Our sample includes all finance companies listed on the Stock Exchange of Thailand for the 1993-2003, which covers the East Asian economic crisis in 1997-1998. In total, there are 41 financial institutions in the sample. Due to the measures taken by the Thai government in order to restore the stability of financial systems, financially distressed finance companies were ordered to close or merge into other institutions in 1997.

B. Explanatory Variables: Financial and Ownership

The variables used to develop our failure prediction models consist of traditional financial and ownership variables. All financial and ownership variables are shown in Table 1-2. There are 30 financial variables and seven ownership variables in total. Thirty financial variables include eight variables in balance sheet, five variable collected from income statements, and 17 calculated variables. As for ownership variables, we use Family, which is a dummy equal to 1 if a family is the largest shareholder of the financial institution, and zero otherwise, and Control Rights, which are the percentage of votes held by the financial institution's largest shareholder, to test the effects of ownership concentration on the probability of business failure.

Our selected financial variables are based on the CAMEL-type of analysis that has been widely used in bankruptcy prediction models for financial institutions. (1) The financial variables of the CAMEL-type include Equity to Assets, Loan Growth, Operating Expenses to Revenue, Return on Assets, Interest Income to Total Income, and Loan to Assets, which are proxies for Capital, Asset, Management, Earnings, and Liquidity components of the CAMEL, respectively.

We divided the explanatory variables into five periods, one year (denoted by t - 1), two years (denoted by t - 2), three years (denoted by t - 3), four years (denoted by t - 4), and five years (denoted by t[inv!]5) prior to the failure. The composition of data sets are shown in Table 3. For example, we use the financial and ownership data of 41 institutions in 1996 to forecast the followings:

1. failure within one year (1997). Data of forty-one firms in 1996 will be grouped into two groups: 24 bankrupt firms and 17 non-bankrupt firms.

2. failure within two year (1996-1997). Data of forty-one firms in 1996 will be grouped into two groups: 30 bankrupt firms and 11 non-bankrupt firms.

3. failure within three year (1995-1997). Data of forty-one firms in 1996 will be grouped into two groups: 31 bankrupt firms and 10 non-bankrupt firms.

4. failure within four year (1995-1997). Data of forty-one firms in 1996 will be grouped into two groups: 31 bankrupt firms and 10 non-bankrupt firms.

5. failure within one year (1997). Data of forty-one firms in 1996 will be grouped into two groups: 32 bankrupt firms and 9 non-bankrupt firms.

These data sets assumed that all missing data are excluded in advance. During 1994-2001, there were 32 bankrupt financial institutions: 24 firms in 1997, six firms in 1998, one firm in 1999, and one firm in 2001. We use the financial and ownership data between 1993 and 1996 for testing and training the failure prediction models. Consequently, the data sets consist of 34 firms in 1993, 35 firms in 1994, 37 firms in 1995, and 41 firms in 1996.

C. Principal Component Analysis

Principal Component Analysis (PCA) is most appropriate for approximating multivariate normal distributions. For such distribution, the low-dimensional linear projections maximizing variance of the training data provide the best possible solution. PCA is applied to the symmetric covariance matrix, which is obtained from a set of data having a zero mean. Then, the eigenvectors and eigenvalues of the covariance matrix are calculated. To reduce dimensionality, the eigenvectors are ordered by their eigenvalues, highest to lowest. It provides the components in order of significance so the components of lesser significance can be discarded.

Let the d-dimensional vector x be a random data vector. Suppose that there are m components of the data set, which m [less than or equal to] d. The [a.sub.j] is the projection of the data vector xi onto the component vector [q.sub.j] as follows.

[a.sub.j] = [q.sup.T.sub.j] [x.sub.i]; j = 1; 2; ..., m (1)

The set of projections {[a.sub.j]|[j.sub.j] = 1; 2; ..., m} into a single vector, as shown by

a = [[a.sub.1]; [a.sub.2], ..., [a.sub.m]].sup.T] (2)

= [[q.sup.T.sub.1] x, [q.sup.t.sub.2] x, ..., [q.sup.T.sub.m]x].sup.T] (3)

III. Failure Prediction Models

Traditional failure prediction models that employed statistical techniques were pioneered by Beaver's univariate tests [6] and Altman's multivariate discriminant analysis (MDA) [1]. Statistical prediction models also include linear probability model (LPM), logit regression approach (LR), probit regression approach, cumulative sums (CUSUM) procedure, and partial adjustment process [5]. However, the most widely-used models are MDA and LR [3, 4]. The early wave of the literature documented that, to name a few, MDA models were used in [1, 7, 10, 18], while LR models were used in [12, 13, 15]. (2)

Not until 1990 have neural network (NN) approaches been introduced in the field of failure/bankruptcy prediction. (3) Coats and Fant [9], Fernandez and Olmeda [11], Salchenberger et al. [17], and Zhang et al. [19] compare between NN and traditional statistical approaches. Their experimental results show that NN significantly outperforms the other methods. In this section, we will describe three classical neural network models: Learning Vector Quantization (LVQ), Probabilistic Neural Network (PNN), and Feedforward network with backpropagation learning.

A. Learning Vector Quantization

Learning Vector Quantization (LVQ) is a supervised learning for classification problem from labeled data samples. In classification, a data vector [x.sub.k] is assigned to a class according to the class label of the closest weight vector. The training algorithm involves an iterative gradient update of the winner weight vector as defined by

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (4)

where [w.sub.c] is the winner weight vector. The direction of the gradient update depends on the correctness of the classification using a nearest neighborhood rule based on Euclidean norm. If a data vector is correctly classified, the weight vector closet to the data vector is moved towards the data; if incorrectly classified, the data vector has a repulsive effect on the weight vector. The update equation for the winner weight vector [w.sub.c] defined by the nearest-neighbor rule and a data vector x(t) are

[w.sub.c](t + 1) [member of] = [w.sub.c](t) [+ or -] [alpha](t)(x(t) - [w.sub.c](t)) (5)

where the sign depends on whether the data vector is correctly classified (+) or misclassified (-). The learning rate [alpha](t) [member of] [0; 1] must decrease monotonically in time. For different picks of data samples from our training set, this procedure is repeated iteratively until convergence.

B. Probabilistic Neural Network

Probabilistic Neural Network (PNN) is a three-layer feed-forward network consisting of an input layer, a hidden layer (pattern layer), and a summation layer. It uses a supervised training set to develop distribution functions within a hidden layer. These functions are used to estimate the likelihood of a labeled data. The prior probability, also called the relative frequency, of each class is used to weigh a given input belonging to the most likely class. If the relative frequency of the classes is unknown, then all classes can be initialized to be equally likely and the initial classes are solely based on the closeness of input data to the distribution function of each class.

In the hidden layer, there is a processing unit for each input data in the training set. Generally, there are equal amounts of processing units for each output class. Otherwise, one or more classes may be skewed incorrectly and the network will generate inefficient results. Each processing unit in the hidden layer is trained once. A unit is trained to generate a high output value when an input vector matches the training vector. The training function may include a global smoothing factor to generalize better classification results. In any case, the training vectors do not have to be in any special order in the training set, since the class of a particular vector is specified by the desired output of the input. The learning function simply selects the first untrained processing unit in the correct output class and modifies its weights to match the training vector.

C. Feedforward Network with Backpropagation learning

The last neural network model applied in this study is based on a simple feedforward network with Backpropagation learning. We build the three-layer network: an input layer, a hidden layer with a number of hidden units, and an output layer with two output units, i.e., a bankruptcy unit and a non-bankruptcy unit. The hidden layer uses the logistic sigmoid activation function, while the output layer uses the pure-linear activation function. The classical backpropagation algorithm is employed as a training approach.

IV. Empirical Analysis

A. Datasets

To implement the prediction model based on neural network approaches, the network must be trained with data of both bankrupt firms and non-bankrupt firms. Consequently, some data year observations in Table 3 are discarded. Therefore, the following five experiments are designed as follows:

1. One-year-ahead prediction model. The data set contains the data in only 1996.

2. Two-year-ahead prediction model. The data set contains the data between 1995 and 1996.

3. Three-year-ahead prediction model. The data set contains the data between 1995 and 1997.

4. Four-year-ahead prediction model. The data set contains the data between 1994 and 1997.

5. Five-year-ahead prediction model. The data set contains the data between 1993 and 1997.

Note that we maintain a fixed ratio of the number of training data to the number of testing data of about 80% for all data sets.

The neural network inputs are varied between 2-37 variables as explained in Section 2.2 and the outputs are classified into two following classes: bankruptcy firms and non-bankruptcy firms.

B. Results of Learning Vector Quantization (LVQ)

The LVQ's parameters are adjusted by experiments. The results are shown in Table 4 with the following parameter settings in Matlab 7.0:

net.trainingParam.epochs = 2,000

net.trainingParam.goal = 0.0005

Number of hidden nodes = 4-15

C. Results of Probabilistic Neural Network (PNN)

The PNN's parameters are adjusted by experiments. The results are shown in Table 5 with the parameter spread in range of [0:25; 0:75], which is set in Matlab 7.0.

D. Results of Feedforward Network with Backpropagation Learning

The backpropagation parameters are adjusted by experiments. The results are shown in Table 6 with the following parameter settings in Matlab 7.0:

net.trainingParam.epochs = 1000

net.trainingParam.goal = 0.00005

Number of hidden nodes = 4-15

Transfer functions = logsig, purelin

Training functions = trainlm (Levenberg-Marquardt backpropagation)

V. Conclusion

In this paper, we propose three bankruptcy prediction models based on neural network approaches. We compare their performances regarding prediction accuracy as shown in Tables 4-6. Moreover, Principal Component Analysis (PCA) is used in our models in order to reduce the dimensionality. The data with dimensions between 5 and 9 are the most efficient while the data with dimensions less than 5 provide the lowest correctness. Comparing among three models, Learning Vector Quantization (LVQ) outperforms two other models when considering both prediction accuracy and bias. Probabilistic Neural Network (PNN) provides consistent results every running time but its accuracy is lowest. Feedforward network with backpropagation learning provides superior accuracy results to those results but its bias is considerably higher than that of the other two methods. Moreover, in an emerging market economy where ownership concentration is common, the ownership variables prove to play an important role in determining the probability with which a financial institution failed.

Acknowledgment

The authors would like to thank Pramuan Bunkanwanicha for providing the useful data.

References

[1] E. Altman, "Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy," J. Finance, vol. 23, pp. 589-609, 1968.

[2] E. Altman, "Application of Classification Techniques in Business, Banking, and Finance," in Contemporary Studies in Economic&Financial Analysis, vol. 3, Greenwich: JAI Press, 1981.

[3] E. Altman and P. Narayanan, "An International Survey of Business Failure Classification Models," Financial Markets, Institutions & Instruments, vol. 6, pp. 1-57, 1997.

[4] A. Atiya, "Bankruptcy Prediction for Credit Risk Using Neural Networks: A Survey and New Results," IEEE Transactions on Neural Networks, vol. 12, pp. 929-935, 2001.

[5] M. Aziz and H. Dar, "Predicting Corporate Bankruptcy: Whither Do We Stand?," Unpublished working paper, 2004.

[6] W. Beaver, "Financial ratios as Predictors of failure," J. Accounting Research, vol. 4, pp. 71-111, 1966.

[7] M. Blum, "Failing Company Discriminant Analysis," J. Accounting Research, vol. 12, pp. 1-25, 1974.

[8] P. Bongini, S. Claessens, and G. Ferri, "The Political Economy of Distress in East Asian Financial Institutions," J. Financial Services Research, vol. 19, pp. 5-25, 2001.

[9] P. Coats and L. Fant, "'Recognizing financial distress patterns usinga neural network tool"', Financial Management, vol. 22, pp. 142-155, 1993.

[10] E. Deakin, "A Discriminant Analysis of Predictors of Business Failure," J. Accounting Research, pp. 167-179, Spring, 1972.

[11] E. Fernandez and I. Olmeda, "Bankruptcy prediction with artificial neural networks," Lect. Notes Comput. Sc., pp. 1142-1146, 1995.

[12] J. Gentry, P. Newbold, and D. Whitford, "Classifying Bankruptcy Firms with Funds Flow Components," J. Accounting Research, pp. 146-160, 1985.

[13] D. Martin, "Early Warning of Bank Failure: A Logit Regression Approach," J. Banking & Finance, vol. 1, 249-276, 1977.

[14] P. Meyer and H. Pifer, "Prediction of Bank Failure," J. Finance, vol. 25, pp. 853-868, 1970.

[15] J. Ohlson, "Financial Ratios and the Probabilistic Prediction of Bankruptcy," J. Accounting Research, vol. 18, pp. 109-131, 1980.

[16] R. Pettaway and J. Sinkey, "Establishing On-Site Bank Examination Priorities: An Early-Warning System Using Accounting and Market Information," J. Finance, vol. 35, pp. 137-150, 1980.

[17] L. Salchenberger, E. Cinar, and N. Lash, "Neural networks for financial diagnosis," Decision Sciences, vol. 23., pp. 889-916, 1992.

[18] J. Sinkey, "A Multivariate Statistical Analysis of the Characteristics of Problem Banks," J. Finance, vol. 30, pp. 21-36, 1975.

[19] G. Zhang, M. Hu, and B. Patuwo et al., "Artificial neural networksin bankruptcy prediction: General framework and cross-validation analysis," European J. Oper. Res., vol. 116, pp. 16-32, 1999.

(1) See for example, [8, 13, 14, 16, 18].

(2) Altman [2] provides a comprehensive survey.

(3) See Atiya [4] for a review of comparison between statistical and NN approaches in bankruptcy prediction models.

Kingkarn Sookhanaphibarn (1), Piruna Polsiri (2) Worawat Choensawat (3) and Frank C. Lin (4)

(1) Advanced Virtual and Intelligent Computing Center Faculty of Sciences, Chulalongkorn University Phayathai Rd., Pathumwan, Bangkok 10330 Thailand kingkarn@ieee.org

(2) Faculty of Business Administration and DPU International College Dhurakij Pundit University 110/1-4 Prachachuen Rd., Laksi, Bangkok 10210 Thailand piruna.poi@dpu.ac.th

(3) Faculty of Information Technology, Dhurakij Pundit University, 110/1-4 Prachachuen Rd., Laksi Bangkok 10210 Thailand worawat@it.dpu.ac.th

(4) Dept. of Mathematics and Computer Science University of Maryland Eastern Shore Princess Anne, MD. 21853, U.S.A. linIBM@ATTglobal.NET

Table 1: Financial variables Financial variables Sources Investment in securities Bal. Sheet Loans and accrued interest receivables, NET Bal. Sheet Total assets Bal. Sheet Average of total assets Calculation Deposits Bal. Sheet Total liabilities Bal. Sheet Issued & paid-up preferred stocks/subordinated convt. Bal. Sheet bond Retained earnings Bal. Sheet Shareholders' equity Bal. Sheet Interest and dividend income Inc. State. Interest on loans Inc. State. Interest & dividend income after bad dept doubtful Inc. State. accounts Income tex expenses Inc. State. Earnings per share (Baht) Inc. State. Log of total assets Calculation One-year growth rate of total loans Calculation One-year growth rate of total assets Calculation Ratio of total loans to total assets Calculation Ratio of total of equity capital to total assets Calculation Ratio of total charge-offs to total loans Calculation Ratio of total interest income to total income Calculation Ratio of operating expense to total assets Calculation Ratio of operating expense to total revenue Calculation Ratio of total deposits to total assets Calculation Ratio of total deposits to total loans Calculation Return on Assets Calculation Return on Average Assets Calculation Earnings before tax to total assets Calculation Earnings before tax to average of assets Calculation Tobin's Q Calculation Table 2: Ownership variables Ownership variables Control rights by the largest shareholder (CRIGHTS) Ownership rights held by the largest shareholder Ratio of ownership to control rights Largest shareholder Type of ownership at the cutoff of 10% for all (Banks and Finance Companies) 1=Family 2=Crown Property Bureau 3=States 4=Foreigner 5=Widely hold (no controlling shareholder) Dummy =1 if LARGE is a family (FAM) Interaction between FAM and CRIGHTS (FAM*CRIGHTS) Table 3: Composition of bankrupt financial firms in 1993-1996 divided into 5 datasets: t - 1, t - 2, t - 3, t - 4, and t - 5. Data year Bankrupt year Datasets t - 1 t - 2 t - 3 t - 4 t - 5 1993 1994 1995 1996 1997 19 19 1998 5 # of Bank. firms 19 24 # of Non-B. firms 34 34 34 15 10 Total # of firms 34 34 34 34 34 1994 1995 1996 1997 20 20 20 1998 5 5 1999 # of Bank. firms 20 25 25 # of Non-B. firms 35 35 15 10 10 Total # of firms 35 35 35 35 35 1995 1996 1997 21 21 21 21 1998 5 5 5 1999 1 1 2000 # of Bank. firms 21 26 27 27 # of Non-B. firms 37 16 11 10 10 Total # of firms 37 37 37 37 37 1996 1997 24 24 24 24 24 1998 6 6 6 6 1999 1 1 1 2000 2001 1 # of Bank. firms 24 30 31 31 32 # of Non-B. firms 17 11 10 10 9 Total # of firms 41 41 41 41 41 Abbreviations: Non-B. is Non-bankrupt. Bank. is Bankrupt. Table 4: Results for the LVQ failure prediction model. TTF nPCA Testing Training Type I Type II Correct error error Correct (%) (%) (%) (%) 1 year 20-37 100 0.00 0.00 93.75 t - 1 10-19 100 0.00 0.00 93.75 5-9 100 0.00 0.00 100 < 5 88.89 0.00 33.33 93.75 2 years 20-37 87.50 18.18 0.00 91.94 t - 2 10-19 87.50 18.18 0.00 91.94 5-9 93.75 9.09 0.00 96.77 < 5 75.00 27.27 20.00 88.71 3 years 20-37 86.96 5.26 50.00 88.89 t - 3 10-19 86.96 5.26 50.00 88.89 5-9 91.30 0.00 50.00 88.89 < 5 65.22 31.58 50.00 77.78 4 years 20-37 83.33 10.00 30.00 85.74 t - 4 10-19 83.33 10.00 30.00 85.74 5-9 83.33 0.00 50.00 85.74 < 5 66.67 25.00 50.00 76.92 5 years 20-37 83.33 9.09 37.50 85.74 t - 5 10-19 83.33 9.09 37.50 85.74 5-9 83.33 9.09 37.50 85.74 < 5 66.67 22.73 62.50 76.92 abbreviations: Type I error is the misclassification of a bankrupt firm as healthy. Type II error is the misclassification of a healthy firm as bankrupt. nPCA is the number of variables after applying PCA. TTF is Time to failure. Table 5: Results for the PNN failure prediction model. TTF nPCA Testing Training Type I Type II Correct error error Correct (%) (%) (%) (%) 1 year 20-37 88.89 0.00 33.33 87.50 t - 1 10-19 88.89 0.00 33.33 87.50 5-9 88.89 0.00 33.33 87.50 < 5 88.89 0.00 33.33 87.50 2 years 20-37 87.50 18.18 0.00 80.65 t - 2 10-19 87.50 18.18 0.00 83.87 5-9 93.75 9.09 0.00 88.71 < 5 87.50 9.09 20.00 88.71 3 years 20-37 78.26 15.26 50.00 83.33 t - 3 10-19 82.26 5.26 75.00 83.33 5-9 91.30 0.00 50.00 86.67 < 5 69.57 26.32 50.00 77.78 4 years 20-37 83.33 10.00 30.00 77.78 t - 4 10-19 86.67 10.00 20.00 80.00 5-9 93.33 0.00 20.00 85.74 < 5 66.67 25.00 50.00 76.92 5 years 20-37 83.33 9.09 37.50 77.78 t - 5 10-19 83.33 9.09 37.50 80.00 5-9 93.33 0.00 25.00 85.74 < 5 66.67 22.73 62.50 76.92 abbreviations: Type I error is the misclassification of a bankrupt firm as healthy. Type II error is the misclassification of a healthy firm as bankrupt. nPCA is the number of variables after applying PCA. TTF is Time to failure. Table 6: Results for the backpropagation failure prediction model. TTF nPCA Testing Training Type I Type II Correct error error Correct (%) (%) (%) (%) 1 year 20-37 100 0.00 0.00 100 t - 1 10-19 100 0.00 0.00 100 5-9 100 0.00 0.00 100 < 5 88.89 0.00 33.33 100 2 years 20-37 93.75 9.09 0.00 100 t - 2 10-19 93.75 9.09 0.00 100 5-9 93.75 9.09 0.00 100 < 5 87.50 18.18 0.00 100 3 years 20-37 86.96 5.26 50.00 100 t - 3 10-19 91.30 0.00 50.00 100 5-9 91.30 0.00 50.00 100 < 5 86.96 0.00 75.00 100 4 years 20-37 83.33 10.00 30.00 100 t - 4 10-19 83.33 10.00 30.00 100 5-9 90.00 0.00 30.00 100 < 5 83.33 0.00 50.00 100 5 years 20-37 83.33 9.09 37.50 100 t - 5 10-19 83.33 9.09 37.50 100 5-9 90.00 0.00 37.50 100 < 5 66.67 22.73 62.50 100 abbreviations: Type I error is the misclassification of a bankrupt firm as healthy. Type II error is the misclassification of a healthy firm as bankrupt. nPCA is the number of variables after applying PCA. TTF is Time to failure.

Printer friendly Cite/link Email Feedback | |

Author: | Sookhanaphibarn, Kingkarn; Polsiri, Piruna; Choensawat, Worawat; Lin, Frank C. |
---|---|

Publication: | International Journal of Computational Intelligence Research |

Geographic Code: | 9THAI |

Date: | Jan 1, 2007 |

Words: | 4154 |

Previous Article: | Evolution of organizational adaptability: application of Hexie management theory. |

Next Article: | New region of interest image coding using general layered bitplane shift for medical image compression. |

Topics: |