استفاده از رگرسیون لجستیک نیروهای خشن برای مدل های امتیاز دهی اعتباری شرکت های بزرگ: شواهدی از صورتهای مالی صربستان
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|24991||2013||13 صفحه PDF||سفارش دهید||11839 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 40, Issue 15, 1 November 2013, Pages 5932–5944
In this paper a brute force logistic regression (LR) modeling approach is proposed and used to develop predictive credit scoring model for corporate entities. The modeling is based on 5 years of data from end-of-year financial statements of Serbian corporate entities, as well as, default event data. To the best of our knowledge, so far no relevant research about predictive power of financial ratios derived from Serbian financial statements has been published. This is also the first paper that generated 350 financial ratios to represent independent variables for 7590 corporate entities default predictions’. Many of derived financial ratios are new and were not discussed in literature before. Weight of evidence (WOE) method has been applied to transform and prepare financial ratios for brute force LR fitting simulations. Clustering method has been utilized to reduce long list of variables and to remove highly correlated financial ratios from partitioned training and validation datasets. The clustering results have revealed that number of variables can be reduced to short list of 24 financial ratios which are then analyzed in terms of default event predictive power. In this paper we propose the most predictive financial ratios from financial statements of Serbian corporate entities. The obtained short list of financial ratios has been used as a main input for brute force LR model simulations. According to literature, common practice to select variables in final model is to run stepwise, forward or backward LR. However, this research has been conducted in a way that the brute force LR simulations have to obtain all possible combinations of models that comprise of 5–14 independent variables from the short list of 24 financial ratios. The total number of simulated resulting LR models is around 14 million. Each model has been fitted through extensive and time consuming brute force LR simulations using SAS® code written by the authors. The total number of 342,016 simulated models (“well-founded” models) has satisfied the established credit scoring model validity conditions. The well-founded models have been ranked according to GINI performance on validation dataset. After all well-founded models have been ranked, the model with highest predictive power and consisting of 8 financial ratios has been selected and analyzed in terms of receiver-operating characteristic curve (ROC), GINI, AIC, SC, LR fitting statistics and correlation coefficients. The financial ratio constituents of that model have been discussed and benchmarked with several models from relevant literature.
Credit scoring models play an important role in contemporary banking risk management practice. They contribute to the key requirement in loan approval process, which is to accurately and efficiently quantify the level of credit risk associated with a customer. The credit scoring models objective is to predict future behavior in terms of credit risk by relying on past experience of customers with similar characteristics. The level of credit risk of a borrower is associated with probability that it will default on approved loan over given time horizon, usually 1 year. The main task of credit scoring model is to provide discrimination between the ones who do default and the ones who do not, i.e. between good and bad corporate entities in terms of their creditworthiness. Discrimination ability is the key indicator of model successfulness. The higher the discrimination power the more precise the credit scoring model will be. The models can be established on judgmental basis or with support of statistical tools. Judgmental or expert-based models are established through set of formal ‘rule-of-thumb’ quantitative criteria. It is an easiest way to incorporate the best practices and the knowledge of credit managers into formal automated decision rules. On the contrast, statistical scoring models are built upon optimization algorithm which is applied on historical data of credit performance of both good and bad customers. For an extensive review of statistical methods and their credit scoring application we refer to study of Crook, Edelman, and Thomas (2007). Contemporary risk management practice emphasizes and promotes the use of credit scoring models for various asset classes of bank’s credit portfolio (BCBS., 2006). Retail banking practice uses application and behavioral credit scoring models for automation of loan approval process for individuals (Kennedy et al., 2013 and Sustersic et al., 2009). By employing process automation, the bank’s staff costs are reduced, loan approval process is simplified, speeded up and more control on approval decision making process is attained (Blochlinger & Leippold, n.d.). In retail banking decision to grant a loan based on fundamental analysis and credit analyst assessment is left to be applied only for high amount or non-standard loans. Latest Serbian credit bureau data report states that about 68% of banking loan exposures belongs to corporate entities (ASB., 2013). In process of financial statement analysis, in order to evaluate the financial health of a corporate entity the financial ratios are commonly used as a part of fundamental analysis. Granting loans to corporate entities based solely on credit scoring models is generally performed only for smaller loan amounts and for particular standardized loan products. More often, credit scoring models are used as an additional tool or decision making criteria. Credit scoring models based on logistic regression (LR) modeling technique provide the results in form of probability of default (PD) level for particular corporate entity. The PD level represents a quantitative estimate of credit risk inherited into corporate entity and it plays a valuable role in credit risk assessment within bank. The PD estimate has been referred to as one of the main and most widely used risk factor in Basel II era (Pluto & Tasche, 2010). The main business utilization of estimated PD serves for calculating: expected and unexpected losses, statistical rating of corporate entities, loan loss provisions and cost of risk component of interest rate (Altman and Sabato, 2007 and Ruthenberg and Landskroner, 2008). Applying a credit scoring model with a higher discrimination power could result in lower capital requirements and more accurate PD estimates (Altman & Sabato, 2007). The Basel II standards recommend portfolio segmentation1 of corporate entities based on sales (BCBS, 2006). Many banks already follow this segmentation practice when modeling credit risk, but in academic literature according to Altman and Sabato (2007) a study that reveals significant benefit of such choice is lacking. The segmentation study of Bijak and Thomas (2012) has shown that segmentation does not always improve credit scoring model performance. Governed with this idea, the underlying principle of our study was to treat all corporate entities as one segment. The goal was to try to build “one-size-fits-all” credit scoring model for whole population of corporate entities. According to relevant literature, two mostly used statistical methods for building credit scoring models for corporate entities are discriminant analysis (DA) and logistic regression (LR). The first attempt to link financial ratios of corporate entities to risk in terms of probability of bankruptcy was done by Beaver (1967). The well-known Z-score model, developed on small corporate entities sample by Altman (1968) and afterwards enhanced by Altman, Haldeman, and Narayanan (1977), was the first credit risk model to predict default probabilities of corporate entities using DA technique. For the first time LR was applied in default prediction study of ( Ohlson (1980). The main benefits of LR over DA were emphasized in terms of less restrictive modeling assumptions. The linearity, normality conditions, as well as, independence among independent variables is not assumed in LR approach which leaves more flexibility in working with real-life data. The first reported LR prediction results were of less predictive power than the ones reported in DA studies. Later on, studies have shown that LR is a sound and powerful statistical approach for modeling credit risk. Further researches of models for predicting business failures using LR are discussed and implemented by Johnsen and Melicher, 1994, Dimitras et al., 1996, Laitinen and Laitinen, 2000, Becchetti and Sierra, 2003, Westgaard and Wijst, 2001, Altman and Sabato, 2007, Kumar and Ravi, 2007 and Chen, 2011. In the last decade the extensive development of credit scoring models has been done. Default prediction as a classification problem entails forecast of corporate entity failure likelihood given a number of independent variables in terms of financial ratios (Altman and Sabato, 2007, Fantazzini and Figini, 2009 and Westgaard and Wijst, 2001). Credit scoring models were first built on data from developed world economies and only later they started to utilize data from different emerging markets. The study of Zekic-Susac, Sarlija, and Bensic (2004) compared LR results with other different estimation methodologies on Croatian bank dataset. The paper of Hermanto and Gunawidjaja (2010) tested the performance of LR model on Indonesian SME data over the period of 2005–2007. The LR study performed on 700 SME loans in Slovakia between 2000 and 2005 pointed out that liquidity and profitability factors are important determinants of SME defaults (Fidrmuc & Hainz, 2010). The recent research of Louzada, Ferreira-Silva, and Diniz (2012) tried to reveal the LR models performance on state-dependent sample extracted from a portfolio of a Brazilian bank. Furthermore, the research of Jain, Gupta, and Sanjiv (2011) examined the behavior of default risk measures and explored the most significant financial variables for SMEs using LR technique. For the purpose of mentioned research, the Indian database of about 3000 SMEs has been used, covering years from 2007 to 2009. Another research, based on Korean dataset (Sohn & Kim, 2012) tried to reveal the best behavioral credit scoring model for technology-based SMEs. The behavioral scoring results have been revealed and compared to its application credit scoring counterpart. Finally, in the most recent study of Blanco, Pino-Mejias, Lara, and Rayo (2013) compared LR results with other non-parametric techniques, based on a sample of almost 5500 microfinance borrowers from Peru. Recent studies for corporate entities show that beside financial ratios there is potential value added, in prediction power terms, when economic, environmental and non-financial information are included in the model as a default predictors (Blanco et al., 2013 and Moon and Sohn, 2010). Even with the existence of more sophisticated classification models for credit scoring, such as neural networks (Derelioglu and Gurgen, 2011, Lee et al., 1996 and Leshno and Spector, 1996), support vector machines (Kim & Ahn, 2012) and case based reasoning (Vukovic, Delibasic, Uzelac, & Suknovic, 2012) the popularity and usage of LR has continued mostly due to its practicality and theoretical soundness. To the best of our knowledge, this is the first study that has examined all possible combinations of the models given the short list of financial ratios as input variables. In comparison to other studies we have generated long list of 350 financial variables and then tailored the principal component clustering technique in order to reduce this long list to a short list of 24 variables. We examined in details the predictive power of financial ratios as standalone variables, as well as, the all possible combinations of models that include 5–14 financial ratio variables.2 The total number of estimated credit scoring models in this study, using logistic regression is around 14 million. As a final result of our research we proposed eight variable LR model with highest predictive power among all developed models. Predictive power of our proposed LR model, based on weight of evidence transformation of financial ratios, has been in line with results of Altman and Sabato (2007) and better when contrasted to results of (Sohn and Kim (2012). The rest of this paper is organized as follows. A description of the Serbian dataset is given in Section 2. The details about research methodology are revealed in Section 3. Firstly, we propose approach for financial ratio construction and transformation based on official balance sheet and income statement dataset of corporate entities. Secondly, the attribute constellation of variables is discussed and performance measures together with LR estimates are presented. This section also emphasizes and describes the variable reduction technique using cluster analysis. In Section 4, we present and discuss LR simulation results and compare the prediction performance of the best models. Finally, we bring our conclusions in Section 5.
نتیجه گیری انگلیسی
In our research we propose corporate entity credit scoring model capable of predicting probability of bankruptcy in 1 year period. Dataset in this research comprised of 5 years of financial statements and performance data (default event data) dating from 2006 to 2011. The long list of 350 financial ratios has been initially constructed based on corporate financial statements and default event data. The research has taken into account 7590 corporate entities. The study used brute force LR to come up with the most predictive credit scoring model. The WoE data transformation technique has been applied in order to divide financial ratios into corresponding attributes and to eliminate problems connected to special values and outliers in financial ratios. A specialized grouping algorithm has been utilized in order to achieve the highest predictive power per financial ratio measured in information value (IV) terms. Data partitioning of the whole population on train and validation sample has been performed using stratification by default rate over the years. We found that partitioning algorithm achieved good partitioning quality over variables and that default rate per year between train and validation sample was balanced. The variable clustering results have been used to select the final short list of 24 variables. The association of each cluster with its containing variables has been measured with R squared ratio. We found that the clustering results were satisfying since 62.31% of total variability has been explained by 24 clusters. We have proposed the main representatives of each cluster by finding the financial ratio variable with highest prediction power within each cluster. The most predictive variable in our research turned out to be EBTADJ/LIAB which achieved information value of 0.923. It comes from the group of ratios explaining corporate entity’s repayment potential. In order to check partitioning quality of train and validation sample in statistical terms we applied Kolmogorov–Smirnov (KS) test. No statistically significant differences were found in cumulative distribution function (cdf) between train and validation samples over 24 analyzed financial ratios. According to such result we concluded that the partitioning results were satisfying. Using the same KS statistics we have tested the short list variables ability to discriminate between good and bad corporate entities. The results have shown statistically significant ability to discriminate. Only one variable (TBDBT/CEQTY) has shown low and non-significant discrimination power, but we found that this variable comes from a low prediction power cluster. In this paper a brute force logistic regression (LR) modeling approach has been utilized to find the best model. The transformed WoE values of financial ratios have been used as inputs for LR simulation fitting. The brute force LR simulation goal has been to obtain all possible combinations of models that comprise of 5–14 financial ratio variables from the short list of 24 financial ratios. The total number of resulting models that have been fitted through time consuming brute force LR simulations process is around 14 million. By predefined validity rules we rejected models that have shown significant multicollinearity and other unstable signs. Finally, we ended up with 342,016 correct and stable models, referred to as well-founded models. The validation sample GINI has been used to benchmark and find the best models comprising of 5–14 variables. By analyzing the “well-founded” models performance we concluded that validation GINI is lower than GINI on train dataset for each model. Finally, we proposed the credit scoring model that comprises of eight variables that has shown the best prediction power performance among all well-founded models. In the last step of research we presented the LR estimates and found a modest correlation persistence structure within the final model variables. We have compared the ROC performance analysis of the proposed model on both train (AUC = 0.835) and validation (AUC = 0.811) sample. Our LR results based on WoE transformation of financial ratios according to AUC results has been in line with results of Altman and Sabato (2007) which compared original values LR results and LR logarithm transformed predictors results. Our final model AUC values are better when contrasted to results of Sohn and Kim (2012). According to elaborated findings, quality of the obtained results and the fact that the credit scoring model is based on actual financial statements and default event data we conclude that the final model can be implemented as a deployable credit scoring model within a bank that operates in Serbia or in the region of South Eastern Europe. However, the model developed in this study has its limitations. The default probability prediction of corporate entities in other economies using this model might be questionable because of different accounting treatments, corporate default rates and economic drivers of other countries. The model may be additionally improved in future research by supplementing it with macroeconomic indicators and non-financial information in order to higher its predictive power.