- Количество слайдов: 26
PREDICTING ROMANIAN FINANCIAL DISTRESSED COMPANIES Supervisor: Prof. Ph. D. Moisa ALTAR MSc Student: Madalina Ecaterina ANDREICA
Summary q Motivation q Literature review q Research design ¨ Data description ¨ Financial ratios ¨ Models and methodologies q Results of the analysis ¨ Principal component analysis ¨ CHAID decision tree model ¨ The logistic and the hazard model ¨ Artificial Neural Network q Conclusions
Motivation The financial crisis has already thrown many companies out of business all over the world. In Romania, for example, a study made by Coface Romania and based on the data provided by the National Trade Register Office, stated that around 14. 483 companies became insolvent by the end of 2008 when they were not able to pay their financial obligations due to inadequate cash flows. Looking at the above situation, we realise that only when a company can build up an efficient early warning system for financial distress and take effective actions before happening, will the company manage to keep on-going in the fierce competition. That is way, the study will focus on identifying a group of distressed and non-distressed Romanian listed companies for which financial ratios for several years will be calculated and then used to predict financial distress based on several models, such as: the Logistic and the Hazard model, the CHAID decision tree model and the Artificial Neural Network model. The study also includes a Principal Component Analysis, in order to better estimate the importance of each financial ratio included in the study.
Literature review Beaver (1966) developed a dichotomous classification test based on a simple t-test in a univariate framework and identified Cash flow/Total Debt as best predictor of bankruptcy. n Altman (1968) suggested the Multivariate Discriminant Analysis (MDA) and identified five predictors: Working Capital to Total Assets, Retained Earnings to Total Assets, Earnings before Interest and Taxes to Total Assets, Market Value of Equity to Book Value of Total Debt and Sales to Total Assets. n Ohlson (1980) used the Logit model and showed that size, financial structure(Total Liabilities to Total Assets), performance and current liquidity were best determinants of bankruptcy. n n Zmijewski’s (1984) first applied the probit model to the firm failure prediction problem. Shumway (2001) propused the hazard model for predicting bankruptcy and found that it was superior to the logit and the MDA models. n Nam, Kim, Park and Lee (2008) developed a duration model with time varying covariates and a baseline hazard function incorporating macroeconomic variables. n In recent years heuristic algorithms such as neural networks, hybrid neural networks and decision trees have also been applied to the distress prediction problem and several improvements were noticed for distress prediction: Zheng and Yanhui (2007) with decision tree models, Yim and Mitchell (2005) with hybrid ANN and others. n
Research design 1. Data description n For this study, public financial information for the period 2005– 2008 was collected from the Bucharest Stock Exchange’s web site. The sample consisted in 100 Romanian listed companies on RASDAQ, equally divided into 50 “distressed” and 50 “non-distressed” companies, that were matched by assets size and activity field. n Since there is no standard definition for a “distressed” company, I followed the same main classification criteria used in other similar studies (Zheng and Yanhui (2007), Psillaki, Tsolas and Margaritis (2008)). That is why, a company was considered “distressed” in case it had losses and outstanding payments for at least 2 consecutive years. n The selection of the main set of financial ratios for each company was conditioned by those variables that appeared in most empirical work, but also restricted to the availability of the financial data.
Research design 2. Financial ratios: Category Return on Assets Net Profit or Loss / Total Assets *100 I 3 Return on Equity Net Profit or Loss / Equity Profit per employee Net Profit or Loss / number of employees Operating Revenue per employee Operating revenue / number of employees I 6 Current ratio Current assets / Current liabilities I 7 Debts on Equity Total Debts / Equity *100 I 8 Debts on Total Assets Total Debts / Total Assets *100 I 9 Working capital per employee Working capital / number of employees I 10 Total Assets per employee Total Assets / number employees I 11 Growth rate on net profit (Net P/ L 1 - Net P/L 0) / Net P/L 0 I 12 Growth rate on total assets (Total Assets 1 – Total Assets 0) / Total Assets 0 I 13 Size Net Profit or Loss / Turnover *100 I 5 Growth ability Profit Margin I 4 Asset utilization Definition I 2 Solvency Financial ratios I 1 Profitability Code Turnover growth (Turnover 1 - Turnover 0) / Turnover 0 I 14 Company size ln (Total Assets) *100
Research design 3. Models and methodologies PCA involves a mathematical procedure that reduces the dimensionality of the initial data space by transforming a number of possibly correlated variables into a smaller number of uncorrelated variables called principal components. These components are synthetic variables of maximum variance, computed as a linear combination of the original variables. CHAID decision tree model finds for each predictor the pair of values that is least significantly different with respect to the dependent variable, based on the p-value obtained from a Pearson Chi-squared test. For each selected pair, CHAID checks if p-value obtained is greater than a certain merge threshold. If the answer is positive, it merges the values and searches for an additional potential. The logistic model is a single-period classification model which uses maximum likelihood estimation to provide the conditional probability of a firm belonging to a certain category given the values of the independent variables for that firm, having the following form: where logit(pi) is the log odds of distress for the given values xi, 1, xi, 2, . . , xi, k of the explanatory variables and β is the coefficient vector
Research design 3. Models and methodologies The hazard model is a multi-period logit model, which includes a baseline hazard function, which can be time-invariant or time varying, depending on its specification. It has the following form: where distress, is the hazard function, xi, t represents the vector of explanatory variables used to forecast is the baseline hazard function and β is the coefficient vector. ANN models have the ability to construct nonlinear models by scanning the data for patterns. The multilayer structure of the feed forward neural network used in this study is the following: an input layer, one hidden layer (following Jain and Nag’s study (2004)) and one output layer. The network was trained in order to learn how to classify companies as distressed and non-distressed. The hybrid ANN method includes as predictors only those variables that were highlighted as being relevant by the previous CHAID, LOGIT and HAZARD models and are marked as ANN – Ii, . . Ik, where Ii, . , Ik are the predictors from the previous models.
Results of the analysis Several distress prediction models were built in search for the model that has best out of sample performances and identifies the financial ratios that are most relevant in distress prediction problem. The following cases of initial data sets were tested: n first-year data, when using the financial ratios of the year 2008 to predict financial distress one year ahead n second-year data, when using the financial ratios of the year 2007 to predict financial distress two years ahead n third-year data, when using the financial ratios of the year 2006 to predict financial distress three years ahead n cumulative three-year data, when using all the financial ratios of the years 2006 -2008 to predict financial distress one year ahead by letting the variables vary in time For each of the four data sets, a descriptive analysis was first conducted in order to be proper informed of any missing data, of the nature of the correlation between all 14 variables, of the differences in mean for each of the two types of companies.
Results of the analysis Data Description First step consisted in identifying the financial ratios that have the highest ability to differentiate between distressed and non-distressed companies based on a mean difference t-test for each of the four data sets.
Results of the analysis Data Description To conclude, here are the significant mean differences in each of the 4 sets of data: q q first-year data set: second-year data set: third-year data set: cumulative three-year data set: I 1, I 2, I 3, I 4, I 5, I 8, I 13 and I 7 I 1, I 2, I 3, I 4, I 5 and I 8 I 1, I 2, I 4, I 5, I 8, I 9 and I 11 I 1, I 2, I 3, I 4, I 5, I 8, I 13 and I 7
Results of the analysis PRINCIPAL COMPONENT ANALYSIS The starting point for the PCA consisted in keeping only those variables that passed the mean differences test, while the purpose was to reduce its dimensions to a space that can allow visual interpretation of the data. The results of the PCA are presented in the following table:
Results of the analysis After applying the PCA for each of the 4 data sets the initial space was reduced to a 3 -dimensional one, without loosing too much information. Now, it can be easily seen how the distressed companies form a separate group from the rest of the non-distressed companies, indicating that the financial information that is used in this study can be significant to classify and to predict the Romanian financial distressed companies.
Results of the analysis Training decision tree for PANEL 2 CHAID CLASSIFICATION TREE: The initial sample of 100 companies was divided into a 70% training sample and a 30% test sample for each of the 4 data sets. In order to measure the decision tree model efficiency, the out-of-sample performances were calculated. SPSS 16. 0 software was used and for each data set two decision trees resulted (one for the training sample and one for the test sample). CHAID was not only used to define the variables that can be used in the measurement of financial distress, but also to determine consistent classification rules, since a decision tree generates a rule for each of its leaves. Training decision tree for PANEL 1
Results of the analysis Training decision tree for PANEL 3 Training decision tree for PANEL 4
Results of the analysis CHAID CLASSIFICATION TREE: The results are summarized in the table below: DATA SETS PANEL 1: first-year data set PANEL 2: second-year data set PANEL 3: third-year data set PANEL 4: cumulative three-year data set principal components selected % in % out of sample 1 88, 6% 93, 3% 1, 2 91, 4% 96, 7% 1, 2 87, 1% 70, 0% 1, 2 84, 3% 84, 4%
Results of the analysis THE LOGISTIC and the HAZARD MODELS: n The study was once again divided into 4 parts, by distinctly analyzing each set of data. In the first three panels, since considering only one year financial data for each company, a single-period logit model was estimated, while when using panel 4 two hazard models were estimated: first a hazard model with time invariant baseline hazard function followed by a hazard model with time varying baseline hazard function incorporating macroeconomic variables. n Once again, the initial sample was divided into a 70% training sample and a 30% forecasting sample n The following steps were taken in order to find the best logistic model for distress prediction: ü First a backward looking procedure ü Then a forward looking procedure ü Then, for each resulting model, each coefficient sign was checked to see if it corresponds to the economic theory and in case of a different sign, the corresponding value was dropped. ü Lastly, the remaining models (in case of more than just one model) were compared based on the following criteria: out-of-sample performance, Mc. Fadden value, LR value, AIC value, the goodness of fit Test (H-L Statistics) and total gain in comparison to the simple constant model.
Results of the analysis PANEL 1: first- year data set PANEL 2: second- year data set Dependent Variable: TIP Method: ML - Binary Logit (Quadratic hill climbing) Date: 06/22/09 Time: 12: 41 Sample: 1 70 Included observations: 70 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Dependent Variable: TIP Method: ML - Binary Logit (Quadratic hill climbing) Date: 06/19/09 Time: 07: 59 Sample: 1 70 Included observations: 70 Convergence achieved after 10 iterations Covariance matrix computed using second derivatives Variable Coefficient Std. Error z-Statistic Prob. C I 1 -1. 777167 -0. 666528 0. 753385 0. 224153 -2. 358909 -2. 973543 0. 0183 0. 0029 Mean dependent var S. E. of regression Sum squared resid Log likelihood Restr. log likelihood LR statistic (1 df) Probability(LR stat) Obs with Dep=0 Obs with Dep=1 0. 500000 0. 173235 2. 040701 -7. 469767 -48. 52030 82. 10107 0. 000000 S. D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Avg. log likelihood Mc. Fadden R-squared 35 Total obs 35 0. 503610 0. 270565 0. 334807 0. 296083 -0. 106711 0. 846049 70 Coefficient Std. Error z-Statistic Prob. C I 3 I 5 I 8 22. 57301 -0. 020148 -2. 138905 0. 033396 6. 459198 0. 009676 0. 592367 0. 012510 3. 494708 -2. 082219 -3. 610778 2. 669627 0. 0005 0. 0373 0. 0003 0. 0076 Mean dependent var S. E. of regression Sum squared resid Log likelihood Restr. log likelihood LR statistic (3 df) Probability(LR stat) Obs with Dep=0 Obs with Dep=1 0. 500000 0. 394996 10. 29744 -30. 59488 -48. 52030 35. 85084 8. 05 E-08 S. D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Avg. log likelihood Mc. Fadden R-squared 35 Total obs 35 0. 503610 0. 988425 1. 116911 1. 039461 -0. 437070 0. 369442 70
Results of the analysis PANEL 4: cumulative three-year data set PANEL 3: third- year data set Hazard model with time-invariant baseline function Dependent Variable: TIP Method: ML - Binary Logit (Quadratic hill climbing) Date: 06/23/09 Time: 02: 42 Sample: 1 70 Included observations: 70 Convergence achieved after 6 iterations Covariance matrix computed using second derivatives Dependent Variable: TIP Method: ML - Binary Logit (Quadratic hill climbing) Date: 06/23/09 Time: 04: 12 Sample: 1 210 Included observations: 210 Convergence achieved after 7 iterations Covariance matrix computed using second derivatives Variable Coefficient Std. Error z-Statistic Prob. C I 5 I 2 15. 20659 -1. 391093 -0. 141076 6. 621508 0. 602248 0. 052150 2. 296545 -2. 309833 -2. 705218 0. 0216 0. 0209 0. 0068 C I 2 I 4 -1. 945630 -0. 157566 -0. 000303 0. 358775 0. 056946 8. 15 E-05 -5. 422981 -2. 766925 -3. 714388 0. 0000 0. 0057 0. 0002 Mean dependent var S. E. of regression Sum squared resid Log likelihood Restr. log likelihood LR statistic (2 df) Probability(LR stat) Obs with Dep=0 Obs with Dep=1 0. 500000 0. 362683 8. 813093 -28. 75849 -48. 52030 39. 52362 2. 62 E-09 S. D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Avg. log likelihood Mc. Fadden R-squared 35 Total obs 35 0. 503610 0. 907385 1. 003750 0. 945662 -0. 410836 0. 407290 70 Mean dependent var S. E. of regression Sum squared resid Log likelihood Restr. log likelihood LR statistic (2 df) Probability(LR stat) Obs with Dep=0 Obs with Dep=1 0. 400000 0. 253761 13. 32969 -43. 41455 -141. 3325 195. 8358 0. 000000 S. D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Avg. log likelihood Mc. Fadden R-squared 126 Total obs 84 0. 491069 0. 442043 0. 489859 0. 461373 -0. 206736 0. 692820 210
Results of the analysis PANEL 4: cumulative three-year data set Hazard model with time-varying baseline function Dependent Variable: TIP Method: ML - Binary Logit (Quadratic hill climbing) Date: 06/23/09 Time: 05: 07 Sample: 1 210 Included observations: 210 Convergence achieved after 8 iterations Covariance matrix computed using second derivatives Variable Coefficient Std. Error z-Statistic Prob. CHANGE_EUR C I 2 I 4 0. 129721 -2. 254988 -0. 195007 -0. 000329 0. 047878 0. 431613 0. 063609 8. 67 E-05 2. 709395 -5. 224558 -3. 065688 -3. 790368 0. 0067 0. 0000 0. 0022 0. 0002 Mean dependent var S. E. of regression Sum squared resid Log likelihood Restr. log likelihood LR statistic (3 df) Probability(LR stat) Obs with Dep=0 Obs with Dep=1 0. 400000 0. 241901 12. 05428 -39. 17753 -141. 3325 204. 3098 0. 000000 S. D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Avg. log likelihood Mc. Fadden R-squared 126 Total obs 84 0. 491069 0. 411215 0. 474969 0. 436988 -0. 186560 0. 722799 210
Results of the analysis DATA SETS THE LOGISTIC and the HAZARD MODELS : PANEL 1: first-year data set PANEL 2: secondyear data set PANEL 3: third-year data set PANEL 4: cumulative three-year data set principal components % in selected sample % out of sample 1, 3 90. 0% 1, 2 94, 3% 96, 7% no valid model 1 87, 1% 86, 7%
Results of the analysis First, the four data sets were transformed as follows: all the positive values of each predictor were scaled to the interval [0, 1], while all the negative values of each predictor were scaled to the interval [-1, 0]. A program using a feed forward backpropagation network was then implemented in MATLAB. THE ANN: first-year all 14 1 % in sample % out of sample 100, 0% 90, 0% PANEL 2: secondyear data set all 14 PANEL 3: third-year data set all 14 1 100, 0% 66, 7% all 14 1 98, 6% 88, 9% PANEL 4: cumulative three-year data set 1 100, 0% type of hybrid ANN PANEL 1: data set first-year PANEL 2: second-year data set PANEL 3: third-year data set PANEL 4: cumulative three-year data set % in sample % out of sample ANN - I 1 1 98, 6% 100, 0% ANN - I 3, I 5 1 91, 4% 100, 0% 1 87, 1% 73, 3% ANN - I 2, I 5 1 85, 7% 76, 7% ANN - I 2, I 4 DATA SETS no. neurons ANN - I 1, I 11 Initial set of variables for no. ANN neurons DATA SETS PANEL 1: data set THE HYBRID ANN: 1 93, 3% 91, 1% ANN - I 2, I 3 1 90, 5% 90, 0%
Conclusions Panel 1: Best financial distress predictor: I 1 (profitability ratio) Best prediction models: single-period logit model and ANN – I 1 Panel 2: Best financial distress predictors: all 14 (profitability, solvency, asset utilization, growth and size ratios) Best prediction model: ANN Panel 3: Best financial distress predictors: (I 1, I 11), (I 2, I 5) (profitability and growth ) Best prediction models: single-period logistic model, CHAID model, ANN –I 1, I 11 and ANN – I 2, I 5 Panel 4: Best financial distress predictors: (I 2, I 4, exchange rate) (profitability ratios and macroeconomic variable) Best prediction model: hazard model with time varying baseline hazard function incorporating macroeconomic variables
References n n n n Abdullah, N. A. H, A. Halim, H. Ahmad and R. M. Rus (2008), “Predicting Corporate Failure of Malaysia’s Listed Companies: Comparing Multiple Discriminant Analysis, Logistic regression and the Hazard Model. ”, International Research Journal of Finance and Economics, 15, Euro. Journals Publishing Inc. Altman, E I. (1968 a), "Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy", Journal of Finance, 23, pp. 589 -609. (1993 b), “Corporate financial distress and bankruptcy”, 2 nd. Ed. , Wiley, New York. Appetiti, S. (1984), “Identifying unsound firms in Italy”, Journal of Banking and Finance, 8, pp 269 -279. Barnes, P. (1987), “The Analysis and Use of Financial Ratios: A Review Article”, Journal of Business Finance and Accounting, 14(4), pp. 449 -461. Beaver, W. (1966), "Financial ratios as predictors of failure", Journal of Accounting Research (Supplement), 4, pp. 71 -102. Beck N, J. N. Katz and R. Tucker (1998), „Taking time seriously: time-series-cross-section analysis with a binary dependent variable”, American Journal of Political Science 42: 1260– 1288. Eisenbeis, R. (1977), “Pitfalls in the application of discriminant analysis in business, finance and economics”, Journal of Finance, 32, pp. 875 -900. Greene, W. H. (1990), Econometric Analysis, Macmillan Publishing Company, New York. Gujarati, D. N. (2004), Basic Econometrics. Mc. Graw-Hill International Editions, 4 th Edition. Hillegeist, S. A. , E. K. Keating, D. P. Cram and K. G. Lundstedt (2004), “Assessing the probability of Bankruptcy”, Review of Accounting Studies, 9, pp. 5 -34. Jain, B. A. and B. N. Nag (1998), „A neural network model to predict long-run operating performance of new ventures”, Annals of Operations Research 78, pp. 83 – 110 Jones, F. L. (1987), "Current techniques in bankruptcy prediction", Journal of Accounting Literature, 6, pp. 131 -164. Koyuncugil, A. S. and N. Ozgulbas (2007), “Detecting financial early warning signs in Istanbul Stock Exchange by data mining”, International Journal of Business Research Lennox, C. (1999), "Identifying failing companies: A re-evaluation of the logit, probit and MDA approaches", Journal of Economics and Business, 51(4), pp. 347 -364. Low, S. , M. N. Fauzias and A. Z. Ariffin, (2001), “Predicting corporate distress using logit model: The case of Malaysia”, Asian Academy of Management Journal, 6(1), pp. 49 -62.
References n n n n Menard, S. (1995), “Applied Logistic Regression Analysis”, Sage University Paper series on Quantitative Applications in the Social Sciences, 07 -106. Thousand Oaks, CA: Sage. Mohamed, S. , A. J. Li, and A. U. Sanda (2001), “Predicting corporate failure in Malaysia: An application of the Logit Model to financial ratio analysis” Asian Academy of Management Journal, 6(1), pp. 99 -118. Nam, J and T. Jinn (2000), “Bankruptcy prediction: Evidence from Korean listed companies during the IMF crisis”, Journal of International Financial Management and Accounting, 11(3), pp. 178 -197. Nam, C. W. , T. S. Kim, N. J. Park and H. K. Lee (2008), “Bankruptcy prediction using a Discrete-Time Duration Model Incoprorating Temporal and Macroeconomic dependencies”, Journal of Forecasting, 27, pp. 493 -506. Ohlson, J. A. (1980), "Financial ratios and the probabilistic prediction of bankruptcy", Journal of Accounting Research, 18, pp. 109 -131. Psillaki, M. , I. E. Tsolas and D. Margaritis (2008), “Evaluation of credit risk based on firm performance”, (EFMA) European Financial Management Association 2008 Annual Meetings June 25 - 28, 2008 Athens, Greece Salchenberger, L. M. , E. M. Cinar and N. A. Lash (1992), „Neural networks: A new tool for predicting thrift failures”, Decision Sciences 23, pp. 899 – 916. Shirata, C. Y. (1998), “Financial ratios as Predictors of Bankruptcy in Japan: An Empirical Research”, Tsukuba College of Technology Japan: 1 -17. Shumway, T. (2001), "Forecasting bankruptcy more accurately: A simple hazard model", Journal of Business, 74 (1), pp. 101 -124. Smith, K. and J. Gupta (2002), ” Neural Networks in Business: Techniques and applications”, Idea Group Publishing Tam, K. Y. and M. Y. Kiang (1992), „Managerial applications of neural networks: The case of bank failure predictions”, Management Science 38926 – 947. Yim J. and H. Mitchell (2005), “ A comparison of corporate distress prediction models in Brazil: hybrid neural networks, logit models and discriminant analysis”, Nova Economia Belo Horizonte 15, 73 -93 Zheng, Q. and J. Yanhui (2007), “Financial Distress Prediction on Decision Tree Models”, IEEE Zmijewski, M. E. (1984), "Methodological issues related to the estimation of financial distress prediction models", Journal of Accounting Research, 22, pp. 59 -86. Zulkarnain, M. S. , A. H. Mohamad Ali, M. N. Annuar and Z. M. Abidin (2001), “Forecasting corporate failure in Malaysian industrial sector firms” Asian Academy of Management Journal, 6(1), pp. 15 -30