Group for Automation in Signals and Communications

World Automation Congress © 2010 TSI Press.

MODELING LOGIC AND NEURAL APPROACHES TO BANKRUPTCY PREDICTION MODELS AMPARO MARIN DE LA BARCENA, ALEXIS MARCANO, J.A. PIN-VELA, DIEGO ANDINA

Group for Automation in Signals and Communications, Technical University of Madrid, Spain Avenida Complutense n030, 28040 Madrid, Spain

ABSTRACT- The guiding principle of process automation and soft computing is to achieve more robust, traceable and low cost solutions which incorporate the required intelligence to information technologies, thus enabling human centered functionalities. The application of Artificial Intelligence

( IA)

and Neural systems to the financial and banking industries has performed well in the areas of

Risk Management improvement and Bankruptcy prediction. This paper contributes to analyze the synergies between logic and neural based approaches as the basis to enhance bankruptcy prediction models development.

Key Words: Bankruptcy Prediction, Artificial Neural Networks Applications, Risk Management

1. INTRODUCTION

Process automation and soft-computing are changing the way people and corporations do business today. Conscious of their potential benefits (e.g. increasing accuracy, increasing efficiency, lower costs, resemblance of human capabilities), fmancial institutions have shown increasing interest towards the adoption of novel soft computing intelligent techniques that can support business management and decision making processes. Financial institutions' core business is the intermediation of funds which can shortly be conceptualized as Risk Management. Although there is no one accepted definition for Risk Management, in the fmancial community it is widely accepted that this concept involves risk taking, risk transformation and risk shifting [1]. For this reason, appropriate Risk Management is at the center of Banks' interests. In the last years, Risk Management has been gaining momentum, particularly in periods of economic downturn. Indeed, nowadays, reliance on risk prediction and assessment models that can support decision making processes has become a strategic lever to achieve competitive advantage and performance improvement [2]-[4]. As the trend towards process automation and the bet for predictive models has been gaining pace, researchers and practitioners have strived to develop and improve Risk Management tools which facilitate the evaluation of customers' credit worthiness, prediction of customer behavior and propensity of default [5]-[10]. In general, significant amount of investigation efforts within the financial environment are towards reducing manual processes and providing assistance to banking agents on their daily decisions [4], which are mainly related to the approval! rejection of credit loans, the extension/ constriction of lines of credit, and the definition of client interest rates. Nevertheless, automation levels remain low and there is still lack of incorporation of intelligence to information technologies, which limits the inclusion of human centered functionalities that could result in significant improvements regarding risk management and predictive modeling. In this context, neural based systems have outperformed many other techniques, particularly in bankruptcy prediction, but the identification of a suitable approach that helps maximize results is often considered a major challenge. Assuming that an appropriate reference framework can facilitate finding the right approach, this paper is aimed to analyze to what extent neural based modeling techniques can be framed within the long­ established framework defmed by the types of modeling techniques that exist in logic. The underlying philosophy of this study is that, potential synergies identified between the logic and neural based approaches, could set the basis for enhanced bankruptcy prediction models development via further exploitation and inclusion of human intelligence and reasoning capabilities on neural based systems. This paper is organized as follows. Section 2 highlights why bankruptcy prediction is important and proceeds in subsequent sub-sections to provide an overview of the different approaches that can be pursued

and how they can be classified, not only within the logic arena, but also from the risk management perspective. Section 3 focuses on the application of Neural systems to bankruptcy prediction and applies the logic framework to the neural based approach to assess to what extent this is a valid reference framework. For illustrative purposes, some practical examples and main results have been included in subsection 3.2. Finally, Section 4 summarizes our main conclusions, and envisioning the potential contribution of neural systems as an emerging soft computing technique, next steps have also been discussed. 2. APROACHING BANKRUPTCY PREDICTION 2.1 Bankruptcy Prediction as a Main Pillar of Risk Management

One of the most important objectives of Risk Management is to pre-empt credit lending to individuals and corporations with high probability of default, which brings us to the identification of Bankruptcy prediction as a main pillar within the Risk Management arena. Bankruptcy is a clear situation in which a person (physical or juridical) is highly prone to not paying back its debt, and therefore, Bankruptcy prediction models can tum very helpful to optimize loan grant decisions whilst improving Risk Management. Bankruptcy is envisioned at least partially as a consequence of failure of Risk Management tools and this largely explains why individuals, companies and even regulators, are more conscious than ever about finding accurate techniques to overhaul the Bankruptcy process [11]. 2.2 Methodologies Overview

After reviewing the importance of Bankruptcy prediction, especially for companies and financial institutions (e.g. banks), the next step is to reason how to tackle this field. In logic, there are two main methods for reasoning, known as the inductive and deductive approaches [12].

;

V Fig. 3. Inductive reasoning

Fig. 4. Deductive reasoning

Inductive reasoning (see Fig. 3), also known as "bottom-up" approach, works moving from specific observations to broader generalizations and theories. This way of reasoning begins with specific observations and measures, tries to detect patterns and regularities, formulates some tentative hypotheses that can be explored, and fmally ends up developing some general conclusions or theories. Deductive reasoning (see Fig. 4), also known as "top-down" approach, works from the more general to the more specific. This method begins with thinking up a theory about the topic of interest. The next step consists in narrowing the initial theory into more specific hypotheses that can be tested. Narrowing down even further, observations to address the hypotheses are collected. This ultimately leads to the ability of testing the hypotheses with specific data, and thus a confirmation (or not) of the original theories.

2.3 Categorization of Bankruptcy Prediction Techniques

The objective is to define a conceptual and holistic method that serves as a framework of reference to predict bankruptcy. In this paper we pretend to bring the aforementioned approaches that exist in logic (Inductive vs. Deductive), to the field of bankruptcy prediction, whilst analyzing and assessing their potential fit, specially as concerns to Neural Networks based techniques. Whereas Inductive reasoning could be assembled to an empirical or statistical approach, Deductive reasoning could be understood as a structural approach, based on modeling the underlying dynamics of a set of parameters [13]. In essence, the starting point of these methods is the information provided by a representative sample of data and they mainly differ on how this information is used.

2.3.1

Statistical methods

During the 1950-s and 1960-s, statistical methods were the dominant techniques. Their main advantage is that they allow the use of estimators, confidence intervals and hypothesis contrary to fact testing to infer the probability of default and identify parameters that could be relevant for bankruptcy prediction and decision making. These models mainly differ amongst themselves in their probability distribution function: • Linear probability model: assumes the probability of default! bankruptcy varies linearly with selected factors (e.g. liquidity, profitability, productivity, solvency, sales-generating ability) [14]-[16]. • Log it model: assumes that the probability of default! bankruptcy is logistically distributed [17][18]. • Probit model: assumes that the probability of default has a (cumulative) normal distribution [19][20]. • Classification or regression trees: segments candidates based on observations about their attributes [21]. 2.3.2

Non-Statistical methods

The next generation of credit scoring and bankruptcy prediction models comes by hand of techniques that are not only based on historical data, but also incorporate forward looking information: Options-pricing theory [22], Expert Systems [23], Genetic Algorithms [24]. The above mentioned are assembled to deductive reasoning techniques due to the fact that, assuming a set of rules and theories, the systems are complex and intelligent enough to formulate hypothesis, test them with real data and based on their observations, confirm whether the reasoning deducted from the initial theories is acceptable. 3. APPLICATION OF NEURAL SYSTEMS 3.1 Overview

For the purpose of this paper, we would like to make emphasis on Neural Networks [25][26], which, as adaptive systems for data processing and analysis[27], differ from the rest of the methods in their flexibility and ability to reproduce human learning and reasoning capabilities (e.g. self-learning, formalization and clustering of unclassified information, historical data based forecasting, addition of new information to past experiences) without an already known model. Research studies on using Neural Networks for bankruptcy prediction started in 1990, and are still active now. As already stated in the previous section, from the logic perspective, there are two main possible approaches to develop bankruptcy prediction models (inductive vs. deductive). In this context, our study has focused on investigating to what extent this logic perspective can be applied and can contribute towards the identification of suitable and accurate Neural based techniques. With the objective remaining the same, i.e. provide a strong mathematical basis for risk related decision making, we have worked under the following assumption: • Inductive approach: Mainly based on scorecards and tables that associate a given candidate its probability of being "healthy", i.e. eligible for being granted a credit, or "unhealthy", prone to go bankrupt. In a simplified view, this approach can be understood a matter of moving from the specific (cases compiled in the tables, previous examples, etc.) to the general. This approach is foreseeable to consist in matching or mapping data, and thus, many linear methods will be included in scope (since linear relationships among data are generally the most intuitive and desired for modeling). • Deductive approach: the starting point is a set of historical data which can indeed be complemented by external sources (e.g. European Central Bank, Specialized Credit Agencies and Bureaus) [28]-[33]. This information is used to make assumptions on client relationships and the target's risk of default or bankruptcy. This approach is considerably centered on non-linear methods. 3.2 Practical examples and main results

We have analyzed the potential for application of Neural Networks towards Bankruptcy prediction from different perspectives.

3.2.1

Business Perspective

Most of the Bankruptcy prediction models are based on companies' key financial indicators [13]: Total assets (TA): include current assets and long term assets. The total assets give some indication of the size of the firm. Therefore it is frequently used as a normalizing factor. The current assets can or will typically be turned into money fairly fast and are financed by the company's total liabilities and shareholders' equity. • Total liabilities (TL): consist of current liabilities and long term debt. Current liabilities include short term loans (less than one year due), accounts payable, taxes due, etc. • The shareholders' equity (BV) consists of the capital raised in share offerings and the retained earnings. The shareholders' equity is also called the book value of the firm. Even though it is based on the historical costs (plus adjustments through depreciation/amortization) of the firm' s assets and liabilities, rather than market values, it has been a very useful indicator in assessing the financial health of a firm. • Retained earnings (RE): means the accumulation of the firm's earnings since the firm's inception. The magnitude that is often considered is the retained earnings before interest and taxes. The firm's earnings are an important indicator as highly negative earnings indicate that the firm is losing competitiveness, which jeopardizes its survival. • Working capital (WC): is obtained subtracting current liabilities to current assets. Indicates the firm's ability to pay its short term obligations. If it is too negative, the company might default on some payments. • Cash Flow (CF): it is widely used as is less prone than earnings to management manipulation. It measures directly the ability of the firm to generate cash to retire debt. Taking the ratios derived from the aforementioned key financial indicators as the input for Neural Networks prediction models, non-linear approaches have proven better results than linear approaches, being saturation one of the main problems. Additionally, the multiplicative factors differing amongst each type of model have also an impact on the accuracy of results. Taking some figures to illustrate the aforementioned statements, linear models can be assembled to single layer Neural Networks where the output is normally binary (1 or 0) and is equivalent to a certain amount. Suppose that an input to a Neural Networks based model is the ratio earnings/ total assets. With a linear approach, if this ratio changes from -0.1 to 0.1, the difference (0.2) would probably be treated in the same way as if the ratio increased from 1.1 to 1.3. Nevertheless, in the case of a non-linear approach, the model would better resemble reality, as the impact from increasing from -0.1 to 0.1 (e.g. -1.0) is lower than when increasing from 1.1 to 1.3 (e.g. 1.2). •

3.2.2

Time and Local Variants

Existing literature reveals that the time and local components have an impact on the accuracy of results. For instance, when prediction is extensive to shorter periods of time, the accuracy of results often improves (see Table II). Table II. Examples of Data Accuracy depending on the Time Horizon
24 months

87,04%

80,95%

65,63%

57,14%

Since the benchmarked cases proceed from studies conducted in different countries (France, Italy, US, Finland, Belgium, Spain, UK), further analysis would be required in order to determine the strength of the influence of the local component on the results obtained. It has been demonstrated the existence of a correlation between a firm's brokerage, its key performance indicators and fmancial ratios linked to the economic situation (see Table III).

Table III. Correlation Matrix Inl!ut

BV/TA

CF/TA

ROqCF}

BV/TA CF/TA ROC(CF) ROA

1.000

0.386

0.146

0.386

1.000

0.442

0.697

0.146

0.442

1.000

0.348

0.316

0.697

0.348

1.000

ROA 0.316

Although to date not many Neural Networks are considering macroeconomic indicators, performance of the predictive models has shown significant improvement when they have been used. Nevertheless, it remains unclear to what extent the impact of economic conditions on bankruptcy predictability is subject to the country. As a next step, it should be investigated if these accuracy improvements due to the inclusion of economic variables are pretty similar or largely vary from country to country. If this is considered in order to compare the performance of inductive vs. deductive approaches, deductive methods are likely to be again better off than inductive ones, as they are more flexible to incorporate new information (e.g. new financial ratios, macroeconomic indicators, ... ). Moreover, as the time variant concerns, inductive methods would imply that tables and scorecards compile a range of cases broad enough to allow appropriate induction levels. Unfortunately, in this arena, data is often unavailable and non-existing; problem which, in the case of deductive methods can be counter-parted by the contemplation of additional and the inclusion of diversified categories of complementary data within the reference sample that will serve as the basis to extract a conclusion through deductive reasoning. 4. CONCLUSIONS AND NEXT STEPS

In this paper we have approached bankruptcy prediction models from the logic perspective. Additionally we have analyzed to what extent neural based systems can be classified within this reference framework and are more appropriate for bankruptcy prediction. The main takeaways are that it is indeed possible to apply the logic categorization to neural based systems and that deductive reasoning results are more appropriate for Neural Networks predictability models. In spite of this, several points are still open issues and should desirably be addressed by the research community. For instance, although in general Neural Network models are superior to other techniques, the accuracy of the results obtained calls for further enhancement and can be sometimes highly constrained by the availability of data sets. Another key point is that, although deductive reasoning has proven better performance than the one of inductive reasoning, in the case of Neural Networks applied to bankruptcy prediction, models in the foreseeable future should combine both approaches. There are already several examples of hybrid models that achieve significant improvements as compared to pure standalone techniques. However, when developing these new models, it is critical not to forget that the key success factor of the next generation of neural based models for bankruptcy prediction relies on an appropriate balance between complexity and efficiency. REFERENCES

1. 2.

3. 4. 5. 6.

Fundamentals of the Banking Business (Grupo Santander, Spain, 2009). 1. Yoon, Y. S. Kwon, T. H. Roh, Performance Improvement of Bankruptcy Prediction using Credit Card Sales Information of Small & Micro Business, in 5th ACIS International Conference on Software Engineering Research, Management & Applications (SERA 2007), pp.503-512. K. Kuester, S. Mittnik and M. S. Paolella, Value-at-Risk Prediction: A Comparison of Alternative Strategies in Journal of Financial Econometrics 4(1) (2006) 53- 89. M. Aiken, Competitive Intelligence through Neural Networks, Wiley Interscience 10(1) (Mississippi, 2001), pp.49-53 1. Sun and H. Li, Data mining method for listed companies' financial distress prediction, Knowledge-Based Systems, 21(1) (2008) 1-5. H. Li and 1. Sun, Majority voting combination of multiple case-based reasoning for financial distress prediction, Expert Systems with Applications, 36(3) (2009) 4363-4373.

7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18.

19. 20. 21. 22. 23.

24.

25. 26. 27. 28. 29. 30. 31. 32. 33.

Sun and H. Li, Financial distress early warning based on group decision making, Computers & Operations Research, 36(3) (2009) 885-906. H. Li and J. Sun, Predicting business failure using multiple case-based reasoning combined with support vector machine, Expert Systems with Applications, 36(6) (2009) 10085-10096. H. Li and J. Sun, Gaussian case-based reasoning for business failure prediction with empirical data in China, lriformation Sciences, 179(1-2) (2009) 89-108. H. Li and 1. Sun, Business failure prediction using hybrid2 case-based reasoning, Computers & Operations Research, 37(1) (2010) 137-151. K. M. Wong, Bankruptcy as a Risk Management Tool: Economic and Social Implications, Review of Business, 24 (2003) W. Trochim, Deduction & Induction, Research Methods Knowledge Base, 2006 A. Atiya, Bankruptcy Prediction for Credit Risk Using Neural Networks: A Survey and New Results, IEEE Transactions on Neural Networks, 12(4) (2001) 929-935 E. Altman, Financial ratios, discriminant analysis and the prediction of corporate bankruptcy, Journal of Finance, 23, (France 1968), pp. 589-609. E. Altman, Corporate financial distress and bankruptcy: a complete guide to predicting and avoiding distress and profitingfrom bankruptcy, (N. Y., Wiley, 1993) E. Altman, R. G. Haldeman and P. Narayan, Zeta analysis: a new model to identify bankruptcy risk of corporations, Journal of Banking and Finance 10, (1977), pp. 29-54 1. Ohlson, Financial ratios and the probabilistic prediction of bankruptcy, Journal of Accounting Research 18(1), (Wiley, 1980), pp. 109-131. S. Pongsatat, J. Ramage and H. Lawrence, Bankruptcy Prediction for Large and Small Firms in Asia: A Comparison of Ohlson and Altman, Journal of Accounting and Corporate Governance 1 (2), (Wiley 2004), pp. 1-13 1. M. Rizzo, The economic determinants of the choice of an exchange rate regime: A probit analysis, in Economic Letters 59(3), (France 1998), pp. 283-287 The probit model, (Indira Ghandhi Institute of Development Research IGIDR, India, 2008) C. Mballo, E. Diday, The criterion of Kolmogorov-Smimov for binary decision tree: Application to interval valued variables, IDS Press 10(4), (France, 2006), pp. 325-341 B. Fisher and M. Scholes, The pricing of options and corporate liabilities, in Journal of Political Economy 81(3), (Chicago, 1973), pp. 637-659 H. A. Abdou, Genetic programming for credit scoring: The case of Egyptian public sector banks, in Expert Systems with Applications: An International Journal 36(9), (New York 2009), pp. 1140211.417 V. Desay, D. G. Conway, 1. N. Crook and G. A. Overstreet, Credit-scoring models in the credit­ union environment using neural networks and genetic algorithms, in IMA Journal of Management Mathematics 8(4), (US 1997), pp.323-346 R. L. Wilson, R. Sharda, Bankruptcy prediction using neural networks, Special issue on neural networksfor decision support 11(5), (Amsterdam, 1994) S. Chakraborty, S. K. Sharma, Prediction of corporate financial health by Artificial Neural Network, in International Journal of Electronic Finance 1(4), (Geneva, 2007) pp.442-459 Alyuda Research Inc. Credit Scoring and Neural Networks(Las Vegas, 2009). Experian 2009. www .experian.co.uk Equifax 2009. www .equifax.co.uk Callcredit 2009. www .ca11credit.co.uk Understanding your Credit Report and Credit Score (Financial Consumer Agency of Canada, 2009) Federal Trade Comission, 2009. www .ftc. gov Credit Information Bureau (India) Limited, 2009. htt;p://www .cibil.com/ 1.