The stability of bankruptcy predictors in the construction and manufacturing industries at various times before bankruptcy

116 2017, XX, 2 Ekonomika a management DOI: 10.15240/tul/001/2017-2-009 Introduction According to Wu (2010), the internal causes of fi rm bankruptcy may be seen in insuffi cient management skills, marketing and an inability to compete. They are refl ected in company performance. For this reason, accounting data, or rather fi nancial ratios, are a frequent source of information for assessing the stability and viability of an enterprise. The literature (Chen & Hsiao, 2008) categorizes companies undergoing business crises as follows:  Companies lacking the capital to manage the business and starting to have problems paying their short-term debts (current liabilities) – see Deakin (1972) and Gilson (1989). Financially, this condition is detectable in the values of current liquidity, quick liquidity, accounts receivable, cash fl ow, total asset turnover, and other factors.  Companies with a negative value of retained earnings for two consecutive periods or negative growth for at least 1 year. The signs of fi nancial problems appear in the following indicators (Altman, 1983): asset profi tability, sales receipts, earnings before and after taxes, and operating profi t margin.  Companies whose shares on a public stock market show an overall drop, are excluded from trading, or withdrawn from the market. Timely recognition of signs pointing to potential bankruptcy provides a chance to avert it. This is why the economic research has long been on a quest for indicators that could signal the threat of bankruptcy at the earliest possible time. In devising a model, it is rather diffi cult to collect suffi cient data on bankrupt companies, as bankruptcy is relatively rare in business. The fi rst models (Altman (1968), Ohlson (1980), Zmijewski (1984) and others) were designed on the basis of fi nancial ratios calculated using company data one year prior to bankruptcy (the period t+1). One of the methods of increasing the accuracy of a model is to use indicators covering several years before bankruptcy (e.g. Perry et al., 1984). Deakin (1972) found that the ranking of predictor signifi cance changes with receding time. Deakin’s conclusion was confi rmed by the work of Grice and Dugan (2001). Shumway (2001) criticizes the earlier bankruptcy models (of Altman, Zmijewski and Ohlson) as static since the time factor is ignored. These issue were also considered by Henerby (1996) who, aided by Cox’s model (see Cox, 1972), analyzed the appropriateness of cash-fl ow-based indicators for predicting bankruptcy, and concluded that these indicators are statistically most signifi cant 3 years before the event and can therefore serve as early indicators. Lin, Liang, and Chen (2011) summarize this problem in the following way: “Early studies tend to treat fi nancial ratios measuring profi tability, liquidity and solvency as signifi cant indicators for the detection of fi nancial diffi culties. However, reliance on these fi nancial ratios can be problematic. The order of their importance, for example, remains unclear as different studies suggest different ratios as the major indicators of potential fi nancial problems.” Another question debated by the academic community is whether the models are transferrable, i.e. whether they can be applied in any environment other than that in which they were created. From a different point of view, authors such as Platt and Platt (1990), Grice and Dugan (2001), Delina, Pácková (2013), Niemann et al. (2008) and Wu, Gaunt and Gray (2010) have pointed out this problem and indicated that the predication accuracy of THE STABILITY OF BANKRUPTCY PREDICTORS IN THE CONSTRUCTION AND MANUFACTURING INDUSTRIES AT VARIOUS TIMES BEFORE BANKRUPTCY Michal Karas, Mária Režňáková EM_2_2017.indd 116EM_2_2017.indd 116 14.6.2017 9:29:3514.6.2017 9:29:35 117 2, XX, 2017 Business Administration and Management bankruptcy models (their ability to differentiate correctly between a company threatened by bankruptcy and a prospering company) falls markedly when they are applied to a different branch, period or economic environment than the original environment. Thomas Ng, Wong and Zhang (2011) also concur with this view and point out the need of creating models for branches such as construction, as the existing models are inappropriate for this branch. Kaplinski (2008) claims that bankruptcy prediction models should be adjusted to the economic conditions of the given country or even branch. The possible explanation could be that the signifi cance of bankruptcy predictors is not stable over time or these predictors are specifi c for a given time, place and branch. As a result, it is recommended that consideration is given to external environment factors in the creation of prediction models (see Carling et al., 2007; Gertler, 2015). The research referenced above inspired the authors to verify the conclusions presented using data on companies operating in the Czech Republic. The purpose of this article is to verify whether bankruptcy predictors are specifi c in terms of industry or time. Another objective is to establish which predictors can signal an imminent bankruptcy more than one period before bankruptcy occurs. 1. Sample and Methods The sample under investigation is comprised of 34,533 companies in two branches in the Czech Republic, namely Manufacturing and Construction. This number represents the total number of observations available in the AMADEUS (Analysis Major Database for European Sources) database, from which the data under investigation was obtained. As signifi cantly fewer companies do business in construction than in manufacturing, we did not create a select group (for instance, through the stratifi cation method), so as not to compromise the comparability of the select groups in terms of their share in the total number of companies operating in the individual industries. There are 34,229 active (or fi nancially healthy companies) and 304 bankrupt companies (a year before bankruptcy), i.e. the share of bankrupt companies is only 0.88%, which is one more reason why we did not create a select group. For more details, see the following table (Tab. 1). The bankrupt companies in our sample declared bankruptcy during years 20082013, although data on these companies was monitored over 5 years. In the case of the bankrupt companies, the fi rst interval studied is the year before bankruptcy which is referred to as period t+1. The period studied is then the period between the years 2003 and 2013. We test a set of 16 fi nancial ratios covering several aspects of a company’s fi nancial health. These ratios are often used in studies on bankruptcy prediction problems (see Tian et al., 2015; Gordini, 2014; Laitinen et al., 2014; Kwak et al., 2014; Bányiová et al., 2014; Faltus, 2014; Carling et al., 2007; Karas & Režňáková, 2013; Cút, 2014). As certain authors point out the instability of bankruptcy predictors over time (e.g. Beaver, 1966; Deakin, 1972; Henerby, 1996; Niemann et al., 2008), we have used fi nancial ratios found in the relevant publications over the last decade for our tests. One of the frequently mentioned risk factors affecting companies, and thus relevant to bankruptcy, is company size (see for example Ohlson, 1980; Peel & Peel, 1987; Karas & Režňáková, 2013; Homolka, Knápková, & Pavelková, 2015). This factor was not, however, examined in this research because the group of going concerns and bankrupt companies was heterogeneous in terms of size. Branch (according to NACE rev 2. Main section) Active % Bankrupt % Total % C – Manufacturing 27,748 81.07% 161 52.96% 27,909 80.82% F – Construction 6,481 18.93% 143 47.04% 6,624 19.18% Total 34,229 100.00% 304 100.00% 34,533 100.00% Source: own analysis of data from the Amadeus database Tab. 1: Numbers of Investigated Companies EM_2_2017.indd 117EM_2_2017.indd 117 14.6.2017 9:29:3514.6.2017 9:29:35 118 2017, XX, 2 Ekonomika a management Signifi cant predictors were fi rst identifi ed on a univariate basis using the parametric t-test or F-test, and the multivariate non-parametric Boosted Trees method was also applied for the sake of comparison. 1.1 T-test and F-test For identifying potential predictors we used a two sample t-test with equal or rather unequal variances. The test procedure can be described in the following way. To test the equality of variances, the F-test was applied. Let there be two independent random samples (X1, …, Xn) from distribution N(μ1;σ2) respectively (Y1, …, Ym) from distribution N(μ2;σ2). We assume that 0> ; 2 ; 22   mn . The t-test tests the null hypothesis that the difference between the means of both groups (μ1, μ2) is equal to some constant (Δ), in most cases to zero (Δ=0), i.e.:  210 :  H (1) Against the alternative hypothesis  211 :  H (2) The test criterion, under the assumption of equal variances, can be written in the following form:    22 11     yx SmSn YX T    2 2      mn t mn mnnm (3) where 22 ,,, y xSSYX are characteristics of the two random samples. The test criterion, under the assumption of unequal variances, can be written in the following form: No. Ratio Abbreviation 1 2 3 4 5 6 7 8 9 1. Current ratio CR x x x x x 2. Working capital/total assets WC/TA x x x x x 3. Working capital/sales WC/S x 4. EBIT/total assets EBIT/TA x x 5. EBITDA/total assets EBITDA/TA x x 6. EAT/equity ROE x 7. Current liabilities/total assets CL/TA x 8. Long-term liabilities/total assets LTL/TA x 9. Debt-equity ratio DER x 10. Sales/total assets S/TA x x x x 11. Sales/stocks S/St. inv. inv. 12. Sales/debtors S/Deb. inv. inv. 13. EBIT/interest EBIT/Int. x 14. EBITDA/interest EBITDA/Int. inv. x 15. Fixed assets/total assets FA/TA x 16. Sales/operating revenue S/OR inv. Source: Tian et al. (2015), Gordini (2014), Laitinen et al. (2014), Kwak et al. (2014), Bányiová et al. (2014), Faltus (2014), Carling et al. (2007), Karas and Režňáková (2013), Cút (2014) Note: Inv. – the ratio was used in the mentioned literature in an inverse form. Source: 1 – Tian et al. (2015), 2 – Gordini (2014), 3 – Laitinen et al. (2014), 4 – Kwak et al. (2014), 5 – Bányiová et al. (2014), 6 – Faltus (2014), 7 – Carling et al. (2007), 8 – Karas, Režňáková (2013), 9 – Cút (2014). Tab. 2: The list of investigated ratios nm EM_2_2017.indd 118EM_2_2017.indd 118 14.6.2017 9:29:3514.6.2017 9:29:35 119 2, XX, 2017 Business Administration and Management   yx mynx vv tvtv S YX      11 (4) where          2 1 2, 1 1XnX n S n i ix          2 1 2 1 1YmY n S m i iy (5), (6) m S v n S v m S n S Sy y x x y x 2 2 2 2 ,,  (7), (8), (9) To test the equality of variances, the traditional F-test can be applied, inter alia, in order to verify the zero hypothesis yx H  : 0 (10) based on the presumption that both samples are independent and come from a normal distribution. The testing criterion is as follows (see Meloun & Militký, 1998, p. 196):         2 2 2 2 ;max x y y x S S S S F (11) if the H0 hypothesis is true and 22 > y xSS the test statistics have an F-distribution with v1 = (n-1) and v2 = (m-1) degrees of freedom. 1.2 Boosted Trees Method The method of Boosted Trees is a combination of the classifi cation and regression trees method (CART) (see Breiman et al., 1983), with a boosting algorithm introduced by J. Friedman (see Friedman, 2001). Using the boosting algorithm increases the accuracy of the classifi cation algorithm to which it is applied by progressively reducing the error term (Braun & Mues, 2012; Friedman, 2001). The resultant classifi cation rule represents a set of many „weak“ learners. The boosting algorithm is most often applied to CART, but an Artifi cial Neural Network (ANN) application may be encountered as well (Kim & Kang, 2010). Classifi cation and Regression Trees (CART) The basic idea behind the trees is the division of a complex problem of feature space in a set of smaller parts known as regions (R) which can be described through simpler models (for example, constants). The central problem of the method of using trees is establishing the optimal divisional boundaries t between these regions R. The boundaries are established in such a way that the demarcated regions, or trees, fulfi ll specifi c defi ned properties. This property of the regions, or trees, is defi ned as a node impurity and the aim of the method is its minimization. For classifi cation purposes, where the output can take the value 1, 2, …, K, it is possible to describe node impurity in the following way, see (Hastie et al., 2009, p. 306). In the m-th node, representing the m-th region Rm with Nm, the number observed is a proportion of the group k in the node m, given by the relation:  1 ˆ im mk i xR m pIyk N   (9) It is then necessary to defi ne the majority of observed elements of the k-th group in the node m as:  ˆ arg max kmk km xp (10) Node impurity of the tree T or Qm(T) can be defi ned using several standards, for example cross-entropy or deviance 1ˆˆ log K mk mk kpp   (11) Deviance as a level of node impurity was used here as part of the presented research. Boosting Boosting is a general approach for making the fi nal deciding rules as a set of several “weak” rules or classifi ers. Amongst the boosting algorithms AdaBoost.M1 is one most frequently applied, see (Freund & Schapire, 1997), the principle of which will be described further. The basis of boosting is the gradual application of the classifi er G(X) to the repeatedly modifi ed version of data and thus to gradually produce other M “weak” classifi ers Gm(X), m = 1, 2, …, M. The resulting classifi er Gfi nal(X) is then made up of the individual partial rules Gm(X) which are given the weights αm. The output is standardized to attain a value of only -1 or 1, see (Hastie et al., 2009, p. 338). EM_2_2017.indd 119EM_2_2017.indd 119 14.6.2017 9:29:3514.6.2017 9:29:35 120 2017, XX, 2 Ekonomika a management The weights α1, α2, …, αM are calculated using a boosting algorithm representing the partial contribution of each classifi er Gm(X). The modifi cation of data in each step of the boosting algorithm is the application of the weights w1, w2, …, wN for each pair of training data (xi, yi), where i = 1, 2, …, N. At the start of the algorithm the weights are set at the value wi = 1/N. In every other iteration m = 2, 3, …, M the weights of individual observations are adjusted. In the m-th iteration the weights of those observations which had been wrongly classifi ed in the previous step are increased by the classifi er Gm-1(X), while the weights of those which were successful are lowered. By this method, the wrongly classifi ed observation is given more attention in order to increase the accuracy of the whole rule. The algorithm Adaboost.M1 is well described in Hastie et al. (2009, p. 338-339). A useful feature of this method is that it allows the sorting out of the variables xj according to their relative infl uence Ij on the variability of the approximation function  xG ˆ across the entire division of input predictors, this measurement can be described as follows, see (Friedman, 2001):   21 var ˆ                   jx j xj x x xG EI (12) Among the advantages of the Boosted Trees method, aside from its nonparametric nature (the data need not be normally distributed), is its tolerance for outliers in the input variable space (Twala, 2010). In addition, the method can even capture non-linear relationships between the variables (Guelman, 2012). Since the lack of normality and the presence of outliers tend to be commonplace in fi nancial data (Barnes, 1982; Schumway, 2001; Wu et al., 2010) it can be expected that a method which is immune to these aspects will deliver higher classifi cation accuracy. Financial ratios were defi ned for each analyzed company in the sample of active companies and in the sample of bankrupt companies. The ratios were calculated according to the status of the company (i.e. active or bankrupt), the last reported year (2008, 2009, …, 2013), i.e. in case of bankrupt companies the year of bankruptcy, and fi nally according to the number of years prior to bankruptcy (t+1, t+2, …, t+5). The F-test and t-test were applied to test the potential differences between the samples of active and bankrupt companies’ ratios and the p-values of these tests were analyzed. 2. Results First, the results of univariate testing by means of the t-test and the F-test will be presented, with results of the application of the Boosted Trees method to follow. 2.1 Results of Univariate Testing The ratios were tested separately for every year prior to bankruptcy (t+1, t+2, …t+5). For example, CR C1 is the current ratio calculated using the data on construction companies one year prior to bankruptcy. The Tab. 3 below shows the results of the t-test and the F-test for the above-defi ned indicators for companies for 2008 through 2012, i.e. it contains results for companies that went bankrupt in 2013, plus results for fi nancially healthy companies. The equality of the variance of values of fi nancial indicators was assessed at the level of 5% signifi cance of the F-test. Due to the large volume of results, indicators that were not signifi cant even at the 10% level during this period (bankruptcy in 2013) were not included in the table. The statistical signifi cance of the t-test results was highlighted by the use of the following designation: *statistically signifi cant at the 10% level, **statistically signifi cant at the 5% level, *** statistically signifi cant at the 1% level. Similarly, company data for 2007-2011 was analyzed, i.e. the year in which bankruptcy occurred was decisive – 2012 in this particular case, and further, for years 2006 through 2010 (bankruptcy year 2011), for years 2005 through 2009 (i.e. bankruptcy year 2010) and years 2004 through 2008 (i.e. bankruptcy year 2009). Since a description of the results for a period of fi ve years before bankruptcy in the abovementioned manner would be excessive (960 test results), it was necessary to aggregate the results in a certain way. For this reason, we indicate moments preceding bankruptcy (t+1, …, t+5), when the relevant indicator was signifi cant without specifying the relevant bankruptcy year (the „last year“). The number of periods before bankruptcy is recorded in an abbreviated manner by a fi gure, where “1” represents the period preceding bankruptcy by 1 year (i.e. the t+1 period), etc. The number of EM_2_2017.indd 120EM_2_2017.indd 120 14.6.2017 9:29:3614.6.2017 9:29:36 121 2, XX, 2017 Business Administration and Management t-stat. df p-val. t-stat.* df* p-val.* F-stat. p-val.** CR 2 C** 2.5282 4263 0.011501 2.43117 53.215 0.018451 1.08 0.634200 CR 4 C** 2.9240 3389 0.003478 2.16025 45.666 0.036038 1.86 0.000872 CR 2 M* -0.30490 2032 0.760475 -0.74771 71.725 0.457078 7.11 0.000000 CR 4 M* -2.11242 1746 0.034792 -0.93430 44.397 0.355202 5.86 0.000000 EBIT/Int. 1 C*** -0.2786 2203 0.780596 -2.69028 434.532 0.007414 491.19 0.000000 EBIT/Int. 2 C** -0.4548 2069 0.649333 -2.05933 25.251 0.049915 24.98 0.000000 EBIT/Int. 3 C*** -0.3261 1827 0.744376 -2.87739 447.563 0.004202 424.59 0.000000 EBIT/Int. 4 C* -0.4594 1666 0.646012 -1.78321 37.730 0.082599 19.28 0.000000 EBIT/TA 4 M* 0.04754 3498 0.962086 0.30246 3492.443 0.762318 1388.00 0.000000 EBIT/TA 5 M* 0.12531 3105 0.900289 0.77600 3077.361 0.437809 990.53 0.000000 EBITDA/Int. 1 C*** -0.3757 1872 0.707147 -3.72636 510.969 0.000216 640.29 0.000000 EBITDA/Int. 3 C*** -0.3798 1618 0.704154 -3.21931 599.612 0.001354 485.79 0.000000 EBITDA/Int. 4 C** -0.5304 1494 0.595927 -2.47817 38.196 0.017739 30.96 0.000000 EBITDA/TA 2 C** -7.3612 3133 0.000000 -2.10017 30.043 0.044214 13.93 0.000000 EBITDA/TA 3 C* -1.9268 2886 0.054104 -1.77173 28.481 0.087140 1.19 0.457900 EBITDA/TA 5 C*** -0.9026 2336 0.366818 -2.81518 25.485 0.009272 10.70 0.000000 FA/TA 1 M*** 0.02000 1504 0.984044 0.15798 1468.815 0.874494 1877.10 0.000000 FA/TA 2 M*** -0.12646 1420 0.899388 -0.96590 1069.734 0.334311 695.21 0.000000 FA/TA 3 M** -0.08528 1325 0.932054 -0.65588 1312.237 0.512019 16734.33 0.000000 FA/TA 4 M*** -0.02011 1226 0.983960 -0.15145 1225.567 0.879647 4068.06 0.000000 FA/TA 5 M** -0.19532 1137 0.845173 -1.48635 1136.478 0.137463 3126.21 0.000000 FA/TA 1 C*** -2.8363 4962 0.004583 -3.54023 62.376 0.000762 1.58 0.024866 FA/TA 3 C** -1.8206 4050 0.068740 -2.05191 53.777 0.045061 1.28 0.258011 S/Deb. 1 M*** 0.24018 846 0.810247 2.45836 843.581 0.014157 30561.81 0.000000 S/Deb. 2 M*** 0.18902 818 0.850127 2.03449 816.507 0.042225 31395.27 0.000000 S/Deb. 4 M** 0.35327 790 0.723977 3.40328 640.173 0.000707 1710.21 0.000000 S/OR 1 M*** 0.45755 1163 0.647362 3.16173 94.965 0.002106 126.28 0.000000 S/TA 1 M*** 0.47726 3157 0.633214 2.92033 3143.366 0.003521 2496.99 0.000000 S/TA 2 M*** 2.42592 2826 0.015332 10.49020 298.188 0.000000 38.03 0.000000 S/TA 3 M*** 2.40274 2491 0.016345 7.92118 153.663 0.000000 16.01 0.000000 S/TA 4 M** 0.73328 2161 0.463467 3.94037 2011.465 0.000084 372.60 0.000000 S/TA 5 M** 0.26586 2914 0.790366 1.79391 2900.744 0.072931 4043.73 0.000000 S/TA 3 C*** -1.9226 4050 0.054605 -3.48700 56.827 0.000950 3.43 0.000000 S/TA 4 C*** -1.3817 3545 0.167137 -4.32837 66.242 0.000052 11.39 0.000000 S/TA 5 C** -1.0647 3050 0.287092 -2.15736 51.207 0.035693 4.36 0.000000 Source: own Tab. 3: The t-test and F-test results for the analyzed ratios in the period 2008-2012 EM_2_2017.indd 121EM_2_2017.indd 121 14.6.2017 9:29:3614.6.2017 9:29:36 122 2017, XX, 2 Ekonomika a management valid observations for certain indicators was insuffi cient, and such results are therefore given as „n.a.“ in the Tab. 4. One year before bankruptcy (t+1), the signifi cant indicators of return on assets (ROA) in the manufacturing industry are calculated as the share of earnings before interest, taxes, depreciation and amortization (EBITDA/TA). Indicators evaluating indebtedness and debt service capacity are then calculated as the debt/ equity ratio (DER 1) and interest coverage, i.e. the ratio earnings before interest and taxes and interest (EBIT/Int. 1). Other important indicators evaluate asset management ability: asset turnover, inventory turnover and receivables turnover (S/TA 1, S/St. 1 and S/Deb. 1). The last group of indicators relevant for the period t+1 consists of indicators showing the composition of revenues, or assets, namely, sales and operating revenue ratio (S/OR 1) and the fi xed assets/ total assets ratio (FA/TA 1). These indicators represent 50% of the indicators analyzed (8 out of 16). Two years before bankruptcy (period t+2), ROA indicators (EBIT/TA 2 or EBITDA/ TA 2, with EBIT/TA being signifi cant only in this particular period) are once again signifi cant in this industry. Then indicators evaluating liquidity and ability to pay interest, such as current ratio (CR 2) and interest coverage (EBIT/Int. 2). Other signifi cant indicators evaluate asset management capacity (S/TA 2, S/St. 2 and S/ Deb. 2). The last group of indicators signifi cant for this period are asset composition indicators (FA/TA 2). These indicators represent 50 % of the indicators analyzed (8 out of 16). In the third year before bankruptcy (period t+3), the same signifi cant indicators as in the previous period apply: EBITDA/TA 3, CR 3, EBIT/Int. 3, S/TA 3, S/St. 3, FA/TA 3. Another signifi cant indicator is DER 3, the signifi cance of which was not confi rmed a year earlier. The ratio of net working capital to assets (WC/TA 3) turned out to be newly signifi cant. This form of liquidity indicator is only signifi cant in this period. The same applies to the short-term indebtedness indicator (CL/TA). These indicators represent 56 % of the indicators analyzed (9 out of 16). In the fourth year before bankruptcy (t+4), 50% of the indicators analyzed continue to be signifi cant – specifi cally the EBITDA/TA 4, CR 3, DER 3, EBIT/Int. 3, S/TA 3, S/St. 3 or S/ Deb. 3, FA/TA 3 ratios. In the fi fth year before bankruptcy (t+5), no indicator is signifi cant with regard to the industry. Ratio/Branch Manufacturing Construction CR 2, 3, 4 2, 3, 4 WC/TA 3 WC/S 1, 3 EBIT/TA 2 EBITDA/TA 1, 2, 3, 4 1, 2, 3, 4 ROE CL/TA 3 LTL/TA n.a. n.a. DER 1, 3, 4 1 S/TA 1, 2, 3, 4 1, 3, 4 S/St. 1, 2, 3, 4 1, 2, 3, 4 S/Deb. 1, 2, 4 3, 4 EBIT/Int. 1, 2, 3, 4 n.a. EBITDA/Int. n.a. 1, 2, 3, 4 FA/TA 1, 2, 3, 4 1, 2, 3, 4 S/OR 1 1, 2, 3, 4 Source: own analysis of data from the Amadeus database Tab. 4: Summary of t-test results EM_2_2017.indd 122EM_2_2017.indd 122 14.6.2017 9:29:3614.6.2017 9:29:36 123 2, XX, 2017 Business Administration and Management Now on to the construction industry. Similarly, no indicator is signifi cant with regard to the industry in the fi fth year before bankruptcy (t+5). In all four periods before bankruptcy (t+1, t+2, t+3, t+4), ROA indicators are signifi cant in the EBITDA/TA form, as is the indicator of asset structure (FA/TA), revenue composition indicators (S/OR), interest coverage indicators evaluating company indebtedness (EBIT/Int.) and asset management indicators in the form of stock turnover (S/St). The debt-equity ratio indicator (DER 1) was only signifi cant one year before bankruptcy (t+1). Other important indicators are indicators evaluating the asset management capacity, specifi cally the total assets turnover for 3 periods (namely S/TA 1, S/TA 3, S/TA 4). The current ratio indicator (CR 2, CR 3, CR 4) proved to be signifi cant in three periods, similarly to the assets turnover indicator (S/TA 1, S/TA 3, S/TA 4). The indicator of management of business credit (receivables) in the form of sales to debtors is an important predictor in the third and fourth years before bankruptcy (S/Deb. 3, S/Deb. 4). Another indicator boasting a high degree of distinction is a modifi ed liquidity indicator calculated as the net working capital to sales ratio (WC/S 1, WC/S 3). It needs to be pointed out in connection with this indicator that it did not turn out to be signifi cant with regard to the manufacturing industry. The number of statistically signifi cant predictors changed over the years. Their number was the lowest two years before bankruptcy (t+2), specifi cally 6; in the other years, their number was 8 or 9 (in t+3). 2.2 Results of Multivariate Testing The results were further verifi ed by means of an alternative method, specifi cally the Boosted Trees method which makes it possible to obtain a different perspective on the signifi cance of the indicators examined. For the purpose of application of the method, the sample was divided into a part serving for model derivation (70%) and a part serving for model testing (30%). In accordance with the literature (see Hastie et al., 2009, p. 363), the overall number of terminal nodes was limited to 6. The parameter of the number of terminal nodes determines the maximum number of iterations between variables. The model is derived by means of an iterative calculation aimed at obtaining the optimum number of trees where the total error Fig. 1: Process of iterative calculation of Boosted Trees model – sample of manufacturing companies Source: own analysis of data from the Amadeus database EM_2_2017.indd 123EM_2_2017.indd 123 14.6.2017 9:29:3614.6.2017 9:29:36 124 2017, XX, 2 Ekonomika a management (or deviance in this case) is minimal. The process of the calculation for manufacturing companies is shown in the graph above (Fig. 1). According to the graph, the optimum number of trees per manufacturing industry sample is 9, while the maximum number of trees was 200. The specifi c error obtained in the training and test sample for both industries is shown in the Tab. 5. Average Multinomial Deviance (risk estimate) values are lower in the model for the manufacturing industry than in the model for Sample Risk (Estimate) Standard (error) Manufacturing Construction Manufacturing Construction Train 0.083051 0.056039 0.002840 0.001264 Test 0.000000 0.020290 0.000000 0.007591 Source: own analysis of data from the Amadeus database Tab. 5: Summary of results of Boosted Trees method application Ratio RI Ratio RI Ratio RI S/St. 1 1.0000 DER 2 0.4561 S/Deb. 5 0.3118 EBITDA/TA 2 0.8966 CR 5 0.4558 EBIT/TA 5 0.3118 WC/S 1 0.8406 WC/TA 2 0.4431 EBITDA/TA 5 0.2981 S/TA 1 0.8213 S/St. 4 0.4423 ROE 1 0.2931 DER 4 0.7836 WC/TA 1 0.4313 CL/TA 4 0.2903 CR 2 0.7310 CL/TA 2 0.4028 EBIT/Int. 5 0.2758 S/Liab. 2 0.7168 WC/S 2 0.3943 DER 1 0.2616 S/Liab. 1 0.7132 S/Deb. 1 0.3823 S/Liab. 5 0.2604 WC/S 5 0.7018 S/TA 3 0.3790 FA/TA 5 0.2377 CR 4 0.6921 FA/TA 2 0.3720 FA/TA 3 0.2354 EBIT/TA 2 0.6827 S/TA 4 0.3689 EBIT/TA 4 0.2211 EBITDA/TA 3 0.6410 S/Deb. 2 0.3653 FA/TA 4 0.1950 ROE 4 0.6409 S/St. 3 0.3530 DER 3 0.1941 ROE 2 0.6155 CR 3 0.3525 S/St. 5 0.1908 S/Deb. 3 0.6029 WC/TA 3 0.3462 EBITDA/TA 4 0.1855 CR 1 0.5777 FA/TA 1 0.3434 ROE 3 0.1765 S/Liab. 4 0.5671 WC/TA 4 0.3421 EBIT/Int. 1 0.1637 EBITDA/TA 1 0.5342 WC/S 4 0.3411 DER 5 0.1216 EBIT/TA 3 0.5254 S/Liab. 3 0.3364 ROE 5 0.1202 S/Deb. 4 0.5254 S/St. 2 0.3351 EBIT/Int. 4 0.1186 S/TA 2 0.5212 CL/TA 5 0.3296 EBIT/Int. 3 0.1044 EBIT/TA 1 0.5202 CL/TA 3 0.3243 EBIT/Int. 2 0.0776 WC/S 3 0.4778 S/TA 5 0.3199 CL/TA 1 0.4707 WC/TA 5 0.3136 Source: own analysis of data from the Amadeus database Tab. 6: Results of application of the Boosted Trees method – sample of manufacturing companies EM_2_2017.indd 124EM_2_2017.indd 124 14.6.2017 9:29:3714.6.2017 9:29:37 131 2, XX, 2017 Business Administration and Management Analysis. In O. Deev, V. Kajurová, & J. Krajíček (Eds.), Proceedings of the 11th International Scientifi c Conference European Financial Systems 2014 (pp. 173-177). Brno: Masaryk University. Freund, Y., & Schapire, R. (1997). A decision-theoretic generalization of online learning and an application to boosting. Journal of Computer and System Sciences, 55(1), 119-139. doi:10.1006/jcss.1997.1504. Friedman, J. H. (2001). Greedy Function Approximation: A Gradient Boosting Machine. Annals of Statistics, 29(5), 1189-1232. doi:10.1214/aos/1013203451. Gertler, Ľ. (2015). Interactions of Unconventional Monetary Policy Measures with the Euro Area Yield Curve. Czech Journal of Economics and Finance, 65(2), 106-126. Gordini, N. (2014). A genetic algorithm approach for SMEs bankruptcy prediction: Empirical evidence from Italy. Expert Systems with Applications, 41(14), 6433-6445. doi:10.1016/j.eswa.2014.04.026. Gilson, S. C. (1989). Management turnover and fi nancial distress. Journal of Financial Economics, 25(2), 241-262. doi:10.1016/0304405X(89)90083-4. Grice, J. S., & Dugan, M. T. (2001). The limitations of bankruptcy prediction models: Some cautions for the researchers. Review of Quantitative Finance and Accounting, 17(2), 151-166. doi:10.1023/A:1017973604789. Guelman, L. (2012). Gradient boosting trees for auto insurance loss cost modelling and prediction. Expert Systems with Applications, 39(3), 3659-3667. doi:10.1016/j.eswa.2011.09.058. Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. New York: Springer-Verlag. Henerby, K. L. (1996). Do Cash Flows Variables Improve the Prediction Accuracy of a Cox Proportional Hazards Model for Bank Failure? The Quarterly Review of Economics and Finance, 36(3), 395-409. doi:10.1016/S1062-9769(96)90023-X. Homolka, L., Knápková, A., & Pavelková, D. (2015). Plastics Cluster Members and Their competitors – DEA benchmarking study. In E. Pastuszková, Z. Crhová, J. Vychytilová, B. Vytrhlíková, & A. Knápková (Eds.), Proceedings of the 7th International Scientifi c Conference Finance and Performance of Firms in Science, Education and Practice (pp. 409-415). Zlín: Tomas Bata University in Zlín. Chen, L. H., & Hsiao, H. D. (2008). Feature selection to diagnose a business crisis by using a real GA-based support vector machine: An empirical study. Expert Systems with Applications, 35(3). doi:10.1016/j. eswa.2007.08.010. Karas, M., & Režňáková, M. (2013). Bankruptcy Prediction Model of Industrial Enterprises in the Czech Republic. International Journal of Mathematical Models and Methods in Applied Sciences, 7(5), 519-531. Kapliński, O. (2008). Usefulness and credibility of scoring methods in construction industry. Journal of civil engineering and management, 14(1), 21-28. doi:10.3846/1392-3730.2008.14.21-28. Kim, M. J., & Kang, D. K. (2010). Ensemble with neural networks for bankruptcy prediction. Expert Systems with Applications, 37(4), 3373-3379. doi:10.1016/j.eswa.2007.05.019. Kwak, W., Cheng, Y., Ni, J., Shi, Y., Gong, G., & Yan, N. (2014). Bankruptcy Prediction for Chinese Firms: Comparing Data Mining Tools with Logit Analysis. Journal of Modern Accounting and Auditing, 10(10), 1548-6583. Laitinen, E. K., Lukason, O., & Suvas, A. (2014). Behaviour of Financial Ratios in Firm Failure Process: An International Comparison. International Journal of Finance and Accounting, 3(2), 122-131. doi:10.5923/j.ijfa.20140302.09. Lin, F., Liang, D., & Chen, E. (2011). Financial ratio selection for business crisis prediction. Expert Systems with Applications, 38(12), 15094-15102. doi:10.1016/j.eswa.2011.05.035. Meloun, M., & Militký, J. (1994). Statistické zpracování experimentálních dat. Praha: Plus. Niemann, M., Schmidt, J. H., & Neukirchen, M. (2008). Improving performance of corporate rating prediction models by reducing fi nancial ratio heterogeneity. Journal of Banking & Finance, 32(3), 434-446. doi:10.1016/j. jbankfi n.2007.05.015. Ng, S. T., Wong, J. M. W., & Zhang, J. (2011). Applying Z-score model to distinguish insolvent construction companies in China. Habitat International, 35(1), 599-607. doi:10.1016/j. habitatint.2011.03.008. Ohlson, J. A. (1980). Financial Ratios and the Probabilistic Prediction of Bankruptcy. Journal of Accounting Research, 18(1), 109-131. doi:10.2307/2490395. Perry, L., Henderson, G., & Cronan, T. (1984). Multivariate analysis of corporate bond ratings and industry classifi cation. EM_2_2017.indd 131EM_2_2017.indd 131 14.6.2017 9:29:3814.6.2017 9:29:38 132 2017, XX, 2 Ekonomika a management Journal of Financial Research, 7(1), 27-36. doi:10.1111/j.1475-6803.1984.tb00351.x. Platt, D. H., & Platt, M. B. (1990). Development of a Class of Stable Predictive Variables: The Case of Bankruptcy Prediction. Journal of Business Finance & Accounting, 17(1), 31-51. doi:10.1111/j.1468-5957.1990. tb00548.x. Peel, M. J., & Peel, D. A. (1987). Some further empirical evidence on predicting private company failure. Accounting and Business Research, 18(69), 57-66. doi:10.1080/0001478 8.1987.9729348. Shumway, T. (2001). Forecasting Bankruptcy More Accurately: A Simple Hazard Model. Journal of Business, 74(1), 101-124. doi:10.2139/ssrn.171436. Tian, S., Yu, Y., & Guo, H. (2015). Variable selection and corporate bankruptcy forecasts. Journal of Banking & Finance, 52(1), 90-100. doi:10.1016/j.jbankfi n.2014.12.003. Twala, B. (2010). Multiple classifi er application to credit risk assessment. Expert Systems with Applications, 37(4), 3326-3336. doi:10.1016/j.eswa.2009.10.018. Wu, W. (2010). Beyond business failure prediction. Expert Systems with Applications, 37(3), 2371-2376. doi:10.1016/j. eswa.2009.07.056. Wu, Y., Gaunt, C., & Gray, S. (2010). A comparison of alternative bankruptcy prediction models. Journal of Contemporary Accounting & Economics, 6(1), 34-45. doi:10.1016/j.jcae.2010.04.002. Zmijewski, M. E. (1984). Methodological issues related to the estimation of fi nancial distress prediction models. Journal of Accounting Research, 22(1), 59-82. doi:10.2307/2490859. Ing. Michal Karas, Ph.D. Brno University of Technology Faculty of Business and Management Institute of Finances [email protected] prof. Ing. Mária Režňáková, CSc. Brno University of Technology Faculty of Business and Management Institute of Finances [email protected].cz EM_2_2017.indd 132EM_2_2017.indd 132 14.6.2017 9:29:3814.6.2017 9:29:38 133 2, XX, 2017 Business Administration and Management Abstract THE STABILITY OF BANKRUPTCY PREDICTORS IN THE CONSTRUCTION AND MANUFACTURING INDUSTRIES AT VARIOUS TIMES BEFORE BANKRUPTCY Michal Karas, Mária Režňáková This article focuses on the design of bankruptcy models, specifi cally the selection of suitable predictors. Previous research has drawn mainly on data concerning manufacturing companies one year before bankruptcy. Our research examines fi nancial ratios that are suitable bankruptcy indicators in two different industries (the construction and manufacturing industries) over a period of fi ve years prior to bankruptcy. Our main objective is to verify whether bankruptcy predictors are industry-specifi c. Another objective was to determine which indicators can detect signs of bankruptcy earlier than one period before bankruptcy. We presume that the application of industryspecifi c indicators can help increase the predictive accuracy of bankruptcy models when applied to a particular industry. Per analogiam, we assume that the inclusion of indicators capable of detecting signs of bankruptcy more than a year before its occurrence will increase their predictive capacity. Signifi cant predictors were fi rst identifi ed on a linear basis using the parametric t-test or F-test; for the sake of comparison, a non-linear non-parametric Boosted Trees method was also applied. Data for a total of 34,229 active companies and 304 companies that went bankrupt during the relevant period was analyzed. The research confi rmed our presumption that bankruptcy predictors are both industry and time specifi c. Four years before bankruptcy, the indicators return on assets, inventory turnover and asset structure are important predictors in both the manufacturing and construction industries. The net working capital to total assets ratio is a specifi c predictor for manufacturing companies in the third year before bankruptcy, as is the short-term indebtedness indicator. In the construction industry, specifi c predictors are the net working capital to sales ratio in the third and fi rst years before bankruptcy, and the interest coverage indicator in all four years preceding bankruptcy. Were these indicators to be included in a model for an alternative industry, they would be likely to reduce its accuracy. Key Words: Financial ratios, bankruptcy prediction models, time-specifi c predictors, branchspecifi c predictors, manufacturing, construction, boosted trees. JEL Classifi cation: G33, C51. DOI: 10.15240/tul/001/2017-2-009 EM_2_2017.indd 133EM_2_2017.indd 133 14.6.2017 9:29:3814.6.2017 9:29:38