Are Risk Index Models Useful for Firm Failure Prediction ?

Firm failure modeling again becomes relevant particularly since GFC (2007-2008) has led to a global recession (2008-2012) and increased number of failed firms in many countries. Good indicator of highly increased risk of Croatian firms’ failure during the recent recession is ratio of corporate nonperforming and total loans. According to the Croatian National Bank (2014), this ratio in Croatia has risen from 8% (2008) up to 28% (2014), i.e. it has increased 3.5 times due to prolonged recession. Increased ratios of corporate non-performing loans are also observable in other post-socialist SEE economies like Slovenia (25% in 2012), Serbia (20% in 2012) and Bosnia and Herzegovina (15.3% in 2012).


Introduction
Firm failure modeling again becomes relevant particularly since GFC (2007)(2008) has led to a global recession (2008)(2009)(2010)(2011)(2012) and increased number of failed firms in many countries.Good indicator of highly increased risk of Croatian firms' failure during the recent recession is ratio of corporate nonperforming and total loans.According to the Croatian National Bank (2014), this ratio in __________________________________________________________________________________________________________________ ______________ Ivica Pervan (2016), Journal of Financial Studies & Research, DOI: 10.5171/2016.751408goes into the bankruptcy procedure, many stakeholders can be negatively affected.Also, for every investor/creditor firm failure risk represents an important element in investment/credit decision and therefore as a relevant issue from the stakeholder, investor/creditor viewpoint emerges the issue of modeling the firm failure i.e. determining the financial profile that distinguishes successful firms from failed firms.
Academic papers provide data on firm failure prediction models for different countries but are not directly comparable due to different definitions of firm failure and used research methodology.According to the literature, firm failure can be defined as bankruptcy, insolvency, default of payment, financial restructuring, etc.For the modeling purposes, academics often use sophisticated techniques such as logistic regression, rough sets, neural networks, fuzzy logic in order to improve prediction accuracy.As a general conclusion from the previous studies, it may be stated that financial ratios can be effectively used for the firm failure prediction.
In countries like Croatia which are bank oriented, the problem of firm failure is particularly interesting from the perspective of the banks.Namely, every bank must control and manage credit risk (among other business risks) in order to survive and earn sufficient profit.The idea that lies behind academic models for firm failure prediction is effectively used and applied in banks.Commercial banks often employ risk index models ("credit scoring models") based on financial ratios and non-financial variables in order to evaluate firm failure risk.Therefore, it is very interesting to evaluate weather sophisticated modeling techniques (logistic regression and neural networks) significantly outperform risk index models.
The rest of the paper is organized as follows.Review of early and recent papers on firm failure prediction methodology and achieved level of prediction accuracy is presented in the second section of the paper.The third section deals with sample description, variables, methodology and empirical findings.The final section of the paper presents concluding remarks and issues for further research.

Modeling of firm failure in academic research
Although many papers point out that firm failure studies have started with seminal papers of Beaver (1966) and Altman (1968), there is even older evidence on academic research of this business phenomenon.Namely, in 1932 U.S. author Fitz Patrick analyzed firm failure of industrial firms by comparing financial ratios of failed and nonfailed firms.His findings indicated that the difference among ratios of failed and nonfailed firms exist at least for three years before firm failure.
In the 1966 paper of Beaver, he defines failure as inability of firm to pay its financial obligations as they mature.Such definition of failure incorporated bankruptcy, bond default, bank account overdrawn and non payment of preferred stock dividend.Beaver's research sample included 79 failed firms and paired 79 nonfailed firms.Nonfailed firms were matched to failed firms by asset size and industry.Dichotomous classification test has revealed that best ratio for prediction of firm failure was ratio cash flow/total debt.The use of data one year before failure resulted with prediction error of 13%, i.e. classification accuracy of 87%.With older data prediction accuracy declines, thus for example data four years before failure resulted with error of 24% and classification accuracy of 76%.
A significant improvement in research methodology was the application of multivariate analysis, i.e. multiple discriminant analysis-DA (Altman, 1968).For the purpose of research, firm failure definition was limited only to firm's bankruptcy.The sample of failed firms included 33 firms which declared bankruptcy in the period 1946-1965, while the sample of 33 nonfailed firms was matched by industry, size and year.In the initial set of independent variables, Altman used a large number of financial ratios, but in the final step in DA model identified only five ratios (working capital/assets, retained earnings/assets, EBIT/assets, market value of equity/book value of equity and sales/assets) as significant for firm bankruptcy prediction.Altman's model had classification accuracy of 95%, i.e. model error was 5% with data one year before bankruptcy.Classification error has increased up to 17% percent with data two years before bankruptcy, indicating classification accuracy of 83%.After Altman's initial usage of DA, many later papers (Edminister, 1972;Deakin, 1972;Pindado and Rodrigues, 2004;Vuran, 2009; etc.) used the same methodology.Deakin (1972) was the first one that questioned Altman's usage of DA.Namely, one of the basic DA assumptions is that observations in each group (failed and nonfailed) are randomly selected.Altman in his 1968 paper as well as many later authors did not use the random selection, but match pair sample approach.Deakin used randomly selected 11 failed and 23 non-failed firms and developed failure prediction model.Classification error of the model was relatively low up to three years before bankruptcy (3-4.5%),but for the fourth and fifth year before bankruptcy prediction error has sharply risen (21% and 17%).Such finding revealed that sample selection procedure directly affects model prediction accuracy, i.e. increased number of healthy firms in the sample leads to decreased classification accuracy in the segment of failed firms.
Another author who has raised questions about the usage of DA in failure studies was Ohlson (1980).He pointed out that DA has two very restrictive requirements, a requirement for normality of predictors and requirement for the same variancecovariance matrices for both groups (failed and nonfailed), which empirically can be rarely satisfied.In order to overcome DA mentioned problems, Ohlson decided to use logit model, statistical method which does not have any assumptions on a priori probabilities and distribution of predictors.In recent time, new more sophisticated methods are developed and used in order to reach higher predictive power of firm failure models.Here, we can point out very frequent usage of neural networks (Tsai and Wu, 2008; Kim and Kang, 2010; Lee, Choi, 2013), recursive partitioning (Marais et al, 1984;Muller, et. al, 2009), hazard models (Abdullah et. al, 2008, Bakhsnani, 2013;Fijorek and Grotowski, 2012;Foster and Zurada, 2013) and fuzzy models (Matviychuk, 2010;Karami et al, 2012).Some of the recent research indicated that modern techniques outperform classical methods (DA and logistic regression-LR), while some report similar level of failure prediction accuracy.

Risk index models of firm failure
Academic discussion of simple risk index models is pretty limited.One of the early authors dealing with this approach was Tamari (1966) Pervan and Peko in 2008 paper presented risk index models for two Croatian commercial banks (Banks "A" and "B").Both banks use qualitative (soft) variables and quantitative (financial ratios) variables.In the case of Bank "A", the following index was calculated on the basis of five financial ratios: On the basis of index which has included the previously described three variables, the author concluded that classification accuracy reaches 85% on average (86% for failed firms and 84% for healthy firms).

Research sample and variables
The research sample includes data for Croatian manufacturing firms retrieved from the Amadeus database (January 2015).For the purpose of research, we have decided that failed firm definition includes insolvent firms, i.e. in searching Amadeus database we have used criterion Status Active, Insolvency proceedings and Default of payment.After elimination of missing data firms, subsample of failed firms incorporated data for 323 observations.For each failed firm we have selected one healthy manufacturing firm similar in size.
An important element of the analysis was the selection of financial ratios that should explain probability of firm insolvency.In the selection of financial ratios, we have decided to use all major financial ratios groups: liquidity, activity, financial structure, profitability and cash flow.On the basis of financial statements from one year before insolvency (t-1), we have selected the following 15 financial ratios that were often used in bankruptcy literature: 1. ROE Although LR is a more robust method than DA, an important issue in its application can be the problem of multicollinearity among independent variables.Since some of the financial ratios that we selected use the same variables in the calculation, there is a real possibility of multicollinearity problem in the estimated model.The problem of multicollinearity in the estimated model causes inefficiently estimated parameters and high errors, which in turn results with many insignificant variables and high explanatory power of the estimated model (Hair et al, 2010).In order to control this problem, we have decided to use matrix of Pearson Correlation coefficients, where correlation higher than 0.8 indicates multicollinearity problem.Matrix of Pearson Correlation coefficients revealed that some of the initially selected financial ratios were highly correlated, since they had coefficients higher than 0.8.After the elimination of the highly correlated independent variables, only the following financial ratios were used in further analysis: ROE, ROA, EBITDA Margin, Net assets turnover, Credit period, Current ratio and Solvency ratio.
Before conducting multivariante analysis, it was useful to apply t-test (Table 3) for testing equality of financial ratio means in order to discover variables which should be the best discriminators among solvent and insolvent firms.3) has revealed that all variables (except Net asset turnover) have statistically significant mean differences and therefore represent potentially good discriminators among solvent and insolvent firms.

Failure prediction model based on logistic regression
In the final step of LR model, the calculated value of Chi-square was 18.3, with significance of 0.045% indicating that the overall fitting of the estimated model for insolvency prediction is good.Another approach of measuring the LR model fitting is Nagelkerke R Square.In this model, Nagelkerke R Square was 53.4% indicating a moderate relationship between the used financial ratios and insolvency prediction.The final LR model (Table 4) included five independent variables, while the constant was statistically insignificant.
ANN learns from the experience (input data), generalizes previous experience and makes decision using testing data.Significant difference of ANN usage for failure prediction in comparison with statistical methods like LR is the fact that ANN cannot evaluate model parameters.Instead of that, ANN calculates the importance of model variables on the scale from 0 to 1.For the purpose of empirical testing, we have decided to split the original data set 70%-30%.Namely, 70% of observations were used for ANN training, while 30% were used for testing.
In developing ANN, we have selected the same variables like in LR model.According to the ANN (Table 6), the highest importance had two liquidity ratios (Credit period and Current ratio).On the other side, Solvency ratio was evaluated as the least important variable.Such finding is logical since failed firms were insolvent firms, i.e. firms which cannot settle current liabilities, while Credit period and Current ratio describe firm ability to pay current liabilities.If firm failure was defined as bankruptcy than solvency ratio would be more important, while liquidity ratios might be relatively less important.

Failure prediction based on risk index
Risk index was developed on methodology similar to Bank "A" and Bank "B", as described in section 2.2.Index is calculated by using five financial ratios (ROA, EBITDA Margin, Credit period, Current ratio and Solvency ratio) that were confirmed as good discriminators in univariant analysis, LR and ANN models.Each financial ratio was grouped into six value groups (intervals) and evaluated on the scale from 0 to 5 (Table 5).
Therefore, for each company in the sample the minimal score according to the developed index was 0, while the maximum score was 25.Index is designed in way that higher score represents higher financial health of the firm consequently lower insolvency risk and vice versa.Since score is evaluated in range from 0 to 25, we have decided to use the median value of 12.5 as a cut off value for the predicted group (solvent vs. insolvent).
Namely, firms with score from 0 to 12.5 were classified as insolvent, while firms with score higher than 12.5 were classified as solvent.
(Balcaen and Ooeghe, 2006) body of firm failure literature, still there is no clear evidence which statistical method is the best for the modeling of firm failure.Studies reach heterogeneous conclusions and therefore there is no resulting consensus on statistical method choice(Balcaen and Ooeghe, 2006).
. He designed index with six elements and total score of 100.As index elements, he has decided to use the following: capital/liabilities, profit trend, current ratio, production/inventory, __________________________________________________________________________________________________________________ ______________ Ivica Pervan (2016), Journal of Financial Studies & Research, DOI: 10.5171/2016.751408

Table 9 : Classification results of risk index model Predicted group
, where it had classification accuracy of 86.4%, which was higher than the result of logistic regression model (82.8%) but lower than the result of artificial neural network (89.4%).However, in the segment of insolvent firms the risk index model has shown moderate result in comparison with sophisticated techniques.Namely, in the segment of insolvent firms, classification accuracy reached only 63.5%, which was much lower in comparison with logistic regression model (78.9%) and artificial neural network (84.4%).Some future research might improve the risk index performance by calculating the optimal cut of value (rather than median value), by using weighted financial ratios or additional ratios that were not available for Amadeus database.Empirical findings again confirm that financial ratios can be effectively used for insolvency risk prediction and management.Designed risk index model confirmed certain possibility of use, but also it was evident that logistic regression model and artificial neural network provide better classification results.That was especially obvious in the case of insolvent firms, which are very interesting to be identified by practical users in order to avoid losses.On the basis of such finding, we can conclude that it can be useful to invest time and money to develop a more sophisticated model in order to achieve more accurate failure predictions.________________________________________________________________________________________________________________________________IvicaPervan (2016), Journal of Financial Studies & Research, DOI: 10.5171/2016.751408 26.Pindado, J and Rodrigues, L F. (2004) 'Parsimonious models of financial insolvency in small companies,' Small Business Economics, 22, 51-66.