Refining Understanding of Corporate Failure through a Topological Data Analysis Mapping of Altman's Z-Score Model
RRefining Understanding of Corporate Failure through a TopologicalData Analysis Mapping of Altman’s Z-Score Model
Wanling Qiu ∗ , Simon Rudkin † , and Pawe(cid:32)l D(cid:32)lotko ‡ School of Management, University of Liverpool, United Kingdom Economics Department, Swansea University, United Kingdom Mathematics Department, Swansea University, United KingdomApril 23, 2020
Abstract
Corporate failure resonates widely leaving practitioners searching for understanding of default risk.Managers seek to steer away from trouble, credit providers to avoid risky loans and investors to mitigatelosses. Applying Topological Data Analysis tools this paper explores whether failing firms from theUnited States organise neatly along the five predictors of default proposed by the Z-score models. Firmsare represented as a point cloud in a five dimensional space, one axis for each predictor. Visualising thatcloud using Ball Mapper reveals failing firms are not often neighbours. As new modelling approachesvie to better predict firm failure, often using black boxes to deliver potentially over-fitting models, atimely reminder is sounded on the importance of evidencing the identification process. Value is addedto the understanding of where in the parameter space failure occurs, and how firms might act to moveaway from financial distress. Further, lenders may find opportunity amongst subsets of firms that aretraditionally considered to be in danger of bankruptcy but actually sit in characteristic spaces wherefailure has not occurred.
Keywords: Credit Scoring; Topological Data Analysis; Data Visualization; Bankruptcy Prediction
Credit default prediction models intuitively find direction from the financial fundamentals of the corporationand identify how such can be used to indicate likely future failures. In Beaver (1966) and Beaver (1968) indi-vidual financial ratios are tested for their ability to discriminate between firms that go bankrupt the followingyear and those who do not. Altman (1968) advances this to select the five ratios that best isolate potentialfailures and employs linear discriminate analysis to assign a coefficient to each. Subsequent developmentscharted for the 50th anniversary of the original Altman (1968) model in Altman et al. (2017a) have includedextensions of the considered ratio set, considerations of non-linearity, removal of the normal distributionassumption implicit in the original multiple discriminate analysis, and the introduction of machine learning(ML) models. Such works break down into those that extend the information set and those which seek toextract more from the information already provided. Each extension has merit but with both comes thedanger of over-fitting to particular values. More advanced techniques also bring questions of being a “blackbox” through which the link from input to output cannot be traced. It is then unsurprising that the originalapproaches continue to have resonance in the credit rating sector (Altman et al., 2017a). ∗ Corresponding Author . Full Address: Accounting and Finance Subject Group, School of Management, University of Liv-erpool, 20 Chatham Street, Liverpool, L69 7ZH, United Kingdom. Tel: +44 (0)7955 109334 Email:[email protected] † Full Address: Economics Department, School of Management, Swansea University, Bay Campus, Swansea, SA1 8EN,United Kingdom. Email:[email protected] ‡ Full Address: Mathematics Department, College of Science, Swansea University, Bay Campus, Swansea, SA1 8EN, UnitedKingdom. Email:[email protected]. a r X i v : . [ q -f i n . GN ] A p r his paper returns to the fundamental models of Altman (1968) and it’s predictions of default againstthe true cases of corporate failure. A Topological Data Analysis (TDA) Ball Mapper (BM) approach after(D(cid:32)lotko, 2019) is used to produce an abstract two-dimensional representation of the financial ratio spaceto visualise where failures occur amongst the combinations of firm characteristics. Mapping the space inthis way demonstrates how contemporary approaches in data science can break open the black box andilluminate how precisely the future of credit default prediction modelling should develop. Major advantagesof the approach include a robustness to high levels of correlation between variables, noise within the datasetand critically respect the underlying data rather than imposing distributional assumptions.Contributions of our work are thus threefold. Firstly, a demonstration of the application of a contempo-rary data science technique to financial analysis inspires a new understanding of the space upon which weconduct our evaluation of credit models. It is immediately seen that the areas of the characteristic spaceclassified as likely to contain failures cover a lot of volume in which there are no failures. Secondly, it isdemonstrated how non-linearity and interaction terms can all be accounted for in the evaluation of creditdefault risk; insights gained there from become invaluable for assessing firms. Finally a research agenda issignposted which can aid understanding of credit risk and open the “black box” of ML. TDA Ball Mapperthus offers much to the discussion of bankruptcy prediction.The remainder of the paper is organised as follows. Section 2 offers brief overview of the literaturedefining the problem of balancing fit against the risk of over-fitting. Section 3 details the data used for theempirical work, with Section 4 highlighting the TDA Ball Mapper approach and the theoretical expectationsformed prior to the analysis. Through Section 5 and Section 6 a detailed review is undertaken as to whetherthe expected patterns are emerging. Finally, Section 7 concludes. Corporate failure has obvious reprecussions for investors and those to whom the failed firm has liabilities.Consequently it is banks and financial institutions who are the greatest users of these models. For such usersdecisions on creditworthiness of potential borrowers must be traceable, clearly motivated by evidence andfree from any allegations of black boxes. Whilst machine learning methods may generate more accurate pre-dictions their data driven nature places too many unknowns in the process of getting from input informationto output decision. Hence whilst there is a growing literature extholing the virtues of a machine learningapproach this paper stands as a note of caution there against.
Altman (1968) and Altman (1983) models were constructed by multiple discriminate analysis (MDA) froma set of candidate factors. Each accounting ratio put forward as a potential explanatory variable is assessedfor its ability to explain firm failure, with only those making a significant contribution being consideredfor the final model. The five factors that made the final Altman (1968) Z-score were chosen from a set oftwenty-one. Later models from Altman (1983) were designed to reflect non-listed firms who did not have amarket value of equity to use in their evaluation. Such MDA approaches to identifying the drivers of firmfailure dominated the literature to the turn of the century (Altman et al., 2017a). As well as MDA there wasa growing thread after Ohlson (1980) application of logistic regression which employed probabilistic models.Extensions were made into other countries but the fundamental ideas fell into one of these two categories.Critically, as demonstrated by the removal of asset turnover in the second of the Altman (1983) frameworksbecause of differences between industries on this characteristic, the appraisal of the researcher was alwaysmaintained as a final check on the model produced.21st century work has been dominated by the growth of ML, with models seeking information from withinthe firm characteristic dataset through a variety of techniques. Many review studies chart this development,a good example being Barboza et al. (2017). In the early works two main families of model found favour.Support vector machines (SVMs) after Cortes and Vapenik (1995) have found credibility because they siton the bridge between the MDA and ML, generating functions for credit scoring without the restrictions onfunctional form that regression necessitates (Altman et al., 2017a). However, the empirical work on European See Altman et al. (2017b) for a review.
In its most intuitive form a topological analysis seeks to create a map of a dataset, helping the analyst viewwhat is going on in each part of the space. When constructing maps we intuitively turn to the longitudeand latitude as point co-ordinates and then overlay information as contours, as colour or as points. Thispaper demonstrates a contemporary approach that generates a “map” of the financial ratios of firms. Ratherthan simply the two dimensions of the page the algorithm is constructing a two-dimensional visualisationof the multiple dimensional data that can then be placed onto the page. Other characteristics of maps,like colour and labelling can readily be overlaid. Because of the loss of dimensionality required to createthe two dimensional map there is no longer any scale that allows measurement between points, but the fullinformation remains to allow the computer so to do. From a practice perspective there is obvious value tosuch an approach.TDA has the immediate advantage of being constructed solely from the data collected. Once the variablesto be collected are decided, measurements are taken and the data is plotted onto the space. By looking at theshape of the data TDA is constructing a map of what is there and not seeking to impose any relationshipsupon the data. Intuitively a linear model is assuming that there can be a straight line drawn through the3ata such that the outcome is a linear function of the input. Familiarity with adding regression lines toa scatter plot mean the effect of linear regression is understood. By using increasingly complex non-linearfunctions it is possible to fit the data better, reduce the residuals, but this comes at the cost of the modelestimated being particular to the data in the sample. This is also known as over-fitting. Machine learningis helping to make that non-linear relationship more accurate to the data, but the over-fitting criticism stillstands. Employing TDA enables understanding of the linearity of the relationship, and if there is no linearityallows the researcher to know more about the shape of the data. From a modelling perspective there arefurther obvious advantages created.Broadly TDA has been limited to the physical sciences where it is valued for its robustness to noise andability to capture relationships between data points irrespective of the way in which they are differentiallyperturbed. For example in looking for genetic mutations it is important to be able to distinguish betweensmall differentials between individuals and genuine changes in the gene that might signal a need for treat-ment(Nicolau et al., 2011). Work to bring TDA into the finance field has typically focused on time seriesand considers the possibility of financial crashes (Gidea and Katz, 2018). Therein it is the use of TDA tomonitor for potential crashes in dynamical systems, such as production lines, that provides the inspiration.This paper represents one of the first applications of the cross section approach outside the natural sciences .From the perspective of the institution seeking to understand the riskiness of a particular business thevalue of the model based approaches is obvious. However, to really understand where the risk is high,drawing from the data is more intuitive. Taking the known characteristics of a firm and placing it withinthe picture can guide on the risk for such a firm. BM does not offer regression coefficients, but by lookingat firms in the same space inference can be gained. Though analysts may have a feeling about which partsof space are risky, the BM algorithm may either confirm, OR sit at odds with, those initial thoughts. Aprocess of learning what is really going on in the data is then the first step to getting the best impression ofcredit default risk. Data is constructed from Compustat and covers the period from 1961 through to 2015. Although there ismore contemporary data available there are few recorded cases of failure since 2015 at the time of writing.This is due to the lag in cases entering the Compustat data. Formally a firm is regarded as failed if it eitherfiles for bankruptcy, or liquidates, in the financial year. For failed firms data from the most recent financialstatements is provided alongside a deletion reason .Explanatory variables are taken from the respective works of Altman (1968) and Altman (1983). Each isconstructed from Compustat data using the formulae defined in Table 1 and contains an allowance for sizethrough a denominator of either total assets or, in the case of X , the total liabilities of the firm. Theseratios capture the liquidity ( X ), profitability ( X ), productivity ( X ), leverage ( X ) and asset turnover( X ). After winzorising at the 1% level and removing any observations for which there is missing data, weare left with 110668 firm-years of which 3.7% are failed firms.Table 2 provides full sample correlations between the five explanatory ratios of Altman (1968) and thefirm failure dummy. Amongst these there are high correlations between profitability, X , and productivity, X , as well as between liquidity , X , and asset turnover, X . None of these values touch the 0.7 in absolutevalue that would be seen as a sign of multicollinearity, but the correlation is high and so regression analysesshould note that in their exposition.It is recognised that effects may vary from year to year and so Table 3 summarises the five ratios for eachyear considered in this paper. For brevity the number of years considered is just 6; being in 10 year intervalsto the most recent data and 2008 to capture what was happening at the start of the global financial crisis.Failure proportions were much higher during the early years, whilst most recently the failure proportion hasbeen very low. Even at the height of the financial crisis the percentage of firms that failed was just 1%. At the time of writing the only known example is Vejdemo-Johansson et al. (2012), which looks at voting behaviours in theUnited States of America House of Representatives. Specifically this paper considers either bankruptcy (code 02) or Liquidation (code 03) X Working Capital / Total Assets ( act − lct ) /at X Retained Earnings / Total Assets re/at -0.089 0.900 -7.699 0.694 X EBIT / Total Assets ( ni + xint + txt ) /at X Market Value of Equity / Total Liabilities ( csho × prcc f ) /tl X Sales / Total Assets sale/at delrsn = 1 , Notes: All data is sourced from Compustat. Description provides the formulae from Altman (1968) or Altman (1983) for theconstruction of the X variables. The column Compustat details the variable names used in the construction of theexplanatory factors ( X to X ). Compustat variable names are as follows act - current assets, lct - current liabilities, at - totalassets, re - retained earnings, ni - net income, xint - interest payments made, txt - taxation on earnings paid, chso - currentshares outstanding, prcc f - price of the share at the financial year end, sale - total sales of the firm and delrsn is the reasonfor deletion from the Compustat database. Firm failure is a dummy for deletion from the Compustat dataset in thesubsequent year owing to either bankruptcy or liquidation. Sample from 1961 to 2015, n = 110668 Table 2: Full Sample Correlations X X X X X Fail X X X X X Notes: All data is sourced from Compustat. Financial ratios are X (liquidity), X (profitability), X (productivity), X (leverage) and X (asset turnover). Fail is a dummy for deletion from the Compustat dataset in the subsequent year owing toeither bankruptcy or liquidation. Sample from 1961 to 2015, n = 110668 Table 3: Annual Summary StatisticsYear Financial Ratios Failure X X X X X (%)1975 0.273 0.259 0.111 1.229 1.169 6.91%(0.196) (0.250) (0.091) (1.792) (1.721)1985 0.234 0.073 0.045 2.492 2.519 6.26%(0.22) (0.544) (0.169) (3.497) (4.105)1995 0.214 -0.101 0.033 3.682 3.735 2.90%(0.225) (0.779) (0.174) (4.951) (5.541)2005 0.196 -0.361 0.021 4.081 4.076 2.01%(0.224) (1.222) (0.180) (4.992) (5.118)2008 0.188 -0.327 -0.013 2.446 2.278 1.01%(0.225) (1.181) (0.216) (3.796) (3.738)2015 0.17 -0.370 -0.013 3.072 3.163 0.03%(0.214) (1.228) (0.191) (4.225) (5.485) Notes:Financial ratios following Altman (1968) are X (liquidity), X (profitability), X (productivity), X (leverage) and X (asset turnover). Failure classified as de-listing from Compustat owing to either bankruptcy or liquidation. Data fromCompustat. Methodology
Analysis of the shape of the data begins with the construction of the point cloud. For credit default modellingusing Altman (1968) this is achieved by plotting each firm as a point in five dimensions. Each coordinatebeing one of the X j ’s used in the formation of the Z-score. In this paper different clouds are formed for eachof the years studied. Firms which are proximate in the five dimensional space must have similar values forall of the considered financial ratios. A theoretical introduction to the method follows, with considerationthen given to the representation that might be expected to emerge. Representation of the multi-dimensional point cloud is achieved using the BM algorithm of D(cid:32)lotko (2019)as implemented in the R package
TDABallMapper (Dlotko, 2019). There are a number of advantages of theD(cid:32)lotko (2019) approach over the original mapper algorithm developed by Singh et al. (2007) and implementedin the
TDAmapper package of R (Pearson et al., 2015). These advantages may be briefly summarised inthe consistency of representation of the point cloud created by the BM algorithm, and its use of a singleparameter rather than the more complex inputs required to the original approach.Formally, the BM algorithm of D(cid:32)lotko (2019) starts with the firm characteristic point cloud X and aconstant (cid:15) >
0. It select a subset C ⊂ X having the property that the set of balls B ( C ) = (cid:83) x ∈ C B ( x, (cid:15) )contains the whole set X . Such a subset C is referred to as an (cid:15) net. Algorithm 1 of the D(cid:32)lotko (2019) paperidentifies neatly how the (cid:15) -net C is formed. Algorithm 1:
Greedy (cid:15) -net(D(cid:32)lotko, 2019)
Input:
Point cloud X , (cid:15) > C = ∅ ;Mark all points of X as uncovered ; while There exist uncovered p ∈ X do C = C ∪ p ;;Mark every point x ∈ B ( p, (cid:15) ) as covered ; endOutput: CIn this construction the ball radius (cid:15) is the only exogenous input. Choosing (cid:15) recognises the competingforces of maintaining detail and producing a representation upon meaningful inference can be made. Thesequential process of Algorithm 1 can produce slightly different results based on the random selection of thenext uncovered point p , but because all possible (cid:15) nets are close to each other, the impact of this randomnessto the overall output is marginal. Owing to the way that the balls are formed the maximal distance of pointsfrom the ball’s centre is bounded by (cid:15) . This may be the entire (cid:15) on one axis and zero distance on others.The total quantity of the distance can also be shared out across all the axes. To think about this considerthe unit circle centered at the point (0 ,
0) being drawn on a two dimensional plane. In this case, every point( x, y ) that satisfy (cid:112) x + y ≤ (cid:15) and so it is left to the researcher to determine how big to set the radius. Algorithm 2:
Construction of a BM graph (D(cid:32)lotko, 2019)
Input:
C, X, (cid:15)V = abstract vertices, one per each element of C ; E = ∅ , ; for p , p ∈ C such that there exist x ∈ X ∩ B ( p , (cid:15) ) ∩ B ( p , (cid:15) ) do E = E ∪ { p , p } ; endResult: BM Graph, G=(V,E)Conversion of the output from Algorithm 1 into a TDA Ball Mapper graph requires a further stage of6igure 1: Interpreting Ball Mapper graphs
Notes: Example TDA Ball Mapper plot created using
BallMapper (Dlotko, 2019). Axes following Altman (1968) are X (liquidity), X (profitability), X (productivity), X (leverage) and X (asset turnover). Data is from Compustat andrepresents the value of these variables in 1975. Colouration is the Z-score as calculated using equation (1). All axes arenormalised to 0,1. (cid:15) = 0 . graph construction. Algorithm 2 provides such a step when an abstract graph to summarize the shape of X is constructed. As defined by the algorithm an edge is drawn between the centroids of every two balls whichhave data points in their intersection; such lines helping identify where in the cloud each ball sits relative tothe others. Because of the way that the graph is constructed it would be expected that more vertices wouldappear in the BM than when using conventional mapper. Consequently, there may be additional informationwhich is visualised in the BM graph.An important decision in the construction of graphs is whether the variables should be normalised. Inthis paper we do normalise all axes onto the range [0 ,
1] to recognise the variability in ranges identified inTables 1 and 3. In other applications normalisation might not be appropriate.
BM graphs have several key features that aid understanding the data they plot. Although necessarilyabstract, the BM graph does maintain topological faith to the underlying dimensions. As an illustration ofthe properties a plot using 1975 data is provided in Figure 1.Firstly, the colouration of the graph allows analysis of the distribution of an outcome of interest acrossthe space. This may simply be the average value for all the data points contained within the ball, as is donein this paper, but it is also possible to use counts, standard deviations, minima, maxima, etc. The choice offunction is methodologically left to the user to define. Because in most instances it is the average outcomethat is considered most representative of a ball, it is this which is the default function in the
BallMapper package. A scale to the right of the plot shows the values, here Z-scores from the Altman (1968) model. Inthis way it can be seen that the lowest scores sit to the right of the plot, with the only ball averaging over 2being ball 12.Secondly, the size of the balls gives indication of the number of data points located within that part ofthe plot. Bigger balls mean more points and a denser data concentration within that (cid:15) radius of the centralpoint of the ball. In Figure 1 it is ball 1 that has the most points contained within it, closely followed by2, 3 and 5. There are a number of less populated balls spreading out to the right of the figure. Some like 4and 12 are very close to these larger balls. This is Algorithm 3 in D(cid:32)lotko (2019). (cid:15) , will lead to more balls being needed tocover the set of data points. Precise determination of (cid:15) for any given application is a matter for the analystto determine, but we might conclude that the choice made in Figure 1 is too high as there is not much detailbeing gained at the centre of the plot. In what follows a smaller (cid:15) is used.BM thus has a number of useful features that can help interpret the link between firm characteristicsand firm failure. As with all methodologies the final choice of inputs will be the defining factor for the valueof the analysis performed.
Altman (1968) proposed the Z-score model for predicting firm failure as: Z = 0 . X + 0 . X + 0 . X + 0 . X + 0 . X (1)concluding that a Z-score of larger than 2.99 would place the firm in the safe zone and unlikely to sufferdistress. A Z-score between 1.8 and 2.99 places the firm in a “grey” zone where failure cannot be ruled out.Should the Z-score be below 1.8 then Altman (1968) assigns the firm to a “distress” zone.As an illustration of the benefit of TDA consider the space defined by the original Altman model. BMis used to construct an abstract representation of the firm characteristics space, here in the five dimensionsset out as X to X in Table 1. In this way the segments of the point cloud with low values for the Z-scorewill be clearly visible. It is then asked whether the firms that failed were indeed in this part of the space.If the model is effective then it would be expected that the proportion of firms within a given ball that failwould be highest in the balls identified as having low Z-scores. Further it should follow that the proportionof failure in balls with high Z-scores should be 0. Figure 2 examines this for 2015. To construct a BM plot aball radius must be selected. For this purpose (cid:15) = 0 . . / .
08. Moving to lower numbers produces a very large set of balls and makes inferencemore challenging .Figure 2 is divided into three key parts. Firstly the Z-scores predicted using equation (1) are plottedin panel (a). Lower values, associated with predictions of failure, are located to the bottom left of the plotand are denoted by reds and oranges in the shading. Higher values are found to the right and towards thetop represented by the blues and purples. In Altman (1968) a Z-score below 1.80 is considered as placing afirm in the “distress” zone. In the plot the “distress” zone will also include the big balls at the centre right.Panel (a) also reveals that no ball has an average Z-score above 2.99 meaning that no ball is consideredentirely within the “safe” zone. Panel (b) is coloured according to the proportion of firms within a givenball that suffer failure in the following year. The highest proportion is 1.5% and occurs to the lower left ofthe big mass. A comparison with panel (a) shows that these were indeed low Z-score balls. In this way the Results from other filtrations are available on request from the authors. (cid:15) = 0 . X (c) X (c) X (c) X (c) X Notes: TDA Ball Mapper (D(cid:32)lotko, 2019) plots of the five dimensions of the original Altman (1968) model generated usingDlotko (2019). Axes are X (liquidity), X (profitability), X (productivity), X (leverage) and X (asset turnover). All axisvariables are normalised to the range 0 to 1 for consistency. Panel (a) is coloured according to the z-score calculated byequation (1). Panel (b) is shaded according to the proportion of observations within the ball that did fail. Panels (c) to (g)are coloured based upon the variables used in the construction of the Z-score. Here we the abstract nature of the plots.Diversity in colour stems from the normalisation process as evidenced in comparison with the actual value plots. Figures areavailable in colour in the on-line version of the paper. X is much closer, both of the ballswith higher proportions of failure correspond to higher values of X . In (1) the coefficient on X is smalland so it has a much smaller effect on the Z score. There is some evidence in these plots that a highercoefficient would be beneficial to represent the 2015 data. X , shown in panel (d), is higher towards the top.There is some correspondence with the overall Z-score, but where the latter is low on the top left X is not.Explaining the low Z-Score in the top left is best done by looking at X . Indeed panel (e) confirms this.Panel (g), X also has a strong correlation with the Z-Score with higher values in the top right. Panel (f)shows a much greater diversity of spread for X , only in the top left of the plot is any consistency observed.Failure in 2015 appears to be most associated with X , but again there are a lot of high values of X wherefailure is not seen.An immediate observation from the 2015 data is that there are many stories behind the data which thelinear discriminate models are not bringing out. Indeed the use of the variables without interactions wouldstill not be able to make the discriminations suggested by Figure 2.Appraising the fit of the original model in older years, Figure 3 has the Z-score and failure proportionplots for 1975, 1985, 1995, 2005 and the height of the global financial crisis in 2008. All plots have a similarlattice format to the 2015 case. However, a narrower, longer, shape in 2005 suggests that the data was morecorrelated that year. It has already been noted that a strength of the TDA Ball Mapper approach is that itcan continue to be applied in cases like this.Panels (a) and (d) of Figure 3 show the Z-scores and firm failure proportions for 2015 and are includedfor reference. Interest begins with panels (b) and (e) which show these two outcomes for 2008. This was thestart of the global financial crisis so theoretically may have the most surprising exits from the Compustatdatabase. Compared to the 2015 plot only a few firms are obtaining the highest Z-scores. The regioncoloured green, yellow, orange and red, covers the majority of the space. The largest balls in the plot arealso in the distress region according to the Altman Z-score. Failures are indeed seen across the space, but thelarger proportions are concentrated in the lower centre of the graph. Again this informs that an interactionbetween the variables will be better to identify where exactly exiting the Compustat listing will occur.For 2005 panels (c) and (f) reveal a similar story of the “distress” zone covering a much larger proportionof the space than the other levels. There are failures in the lower part of the plot that correspond with thehigh Z-scores and low Z-scores. The most intense of the failure proportions appear in the top of the plot,far from the high Z-score end. Panels (g) and (j), plotting 1995, show that there is a smaller coverage ofballs with low Z-scores. Failure proportions in panel (j) here correspond more with the high Z-scores to theright of the plot; the Altman (1968) model does not perform well for 1995. By contrast panels (h) and (k)for 1985 have the failures primarily concentrated in the bottom right, an area with very low Z-scores. Thereare also low failure proportions within the biggest balls, here again the average Z-score is well below the 1.8cut off for the “distress” zone. Going back through time the same filtration produces fewer balls, the 1975plots in panels (i) and (l), are particularly simplified relative to the others. Here the failures sit to the leftof the plot in an area where there were some very high Z-scores noted.Overall the TDA Ball Mapper plots have usefully shown that the firms that failed have characteristics in,or around the boundary of, the “distress” zone. This will explain the high accuracy of prediction from theAltman (1968) model. However, there are also many cases where the failed firms sat in areas of the plot whereZ-scores were high and financial distress was not expected. Two important messages thus emerge. Firstlythere is a need to split the “distress” zone using the interactions between variables. Secondly, non-linearitybetween the financial ratios and outcomes apply across the space. This paper seeks to evaluate the effectiveness of credit modelling in predicting firm failure, the lens appliedmaking it possible to identify segments of the parameter space in which failures occur and, through thefunctionality of the
BallMapper package, to identify the intensity of the failure rate in any given part of10igure 3: Altman Z Scores and Firm Failures: Original Model(a) 2008 Z-score (b) 2005 Z-score (c) 1995 Z-Score(d) 2008 Failure proportion (e) 2005 Failure proportion (f) 1995 Failure proportion(g) 1995 Z-score (h) 1985 Z-score (i) 1975 Z-score(j) 1995 Failure proportion (k) 1985 Failure proportion (l) 1975 Failure proportion
Notes: TDA Ball Mapper plots generated using D(cid:32)lotko (2019) for the original Altman (1968) model. Axes are X (liquidity), X (profitability), X (productivity), X (leverage) and X (asset turnover). Panels (a) to (c) and (g) and (i) are colouredaccording to the z-score calculated by equation (1). Panels (d) to (d) and (j) to (l) are shaded according to the proportion ofobservations within the ball that did fail. Figures are available in colour in the on-line version of the paper. Introducing topological data analysis, and specifically the BallMapper algorithm of D(cid:32)lotko (2019), to simpledatasets on financial defaults offers much to understanding the contribution of individual financial ratiosto liquidation and bankruptcy. Mapping the financial ratio characteristic space it is shown that failureonly occurs within a subset of the space. It would therefore be beneficial to go further to identify that spaceusing interaction terms. This paper has shown that interactions between liquidity, profitability, productivity,leverage and asset turnover should be explored further. BM has the ability to signpost exactly where firmsare in the space, how close they are to areas where failure has been observed, and how we might thenunderstand the decision to give credit based thereupon. Placing firms on the “map” is an important step toevaluating creditworthiness. Contributions of a new approach and demonstrations of non-linearity are clear.Because everything that appears in the plot is driven by the data there is no “black-box” criticism to any ofthe results that emerge; this sets BM apart from the machine learning literature which has yet to be fullytrusted in practice.A number of potential extensions emerge, with applications to other datasets, addition of further axesand consideration of wider time-frames being obvious next steps. However, each of these is just a smallincrement relative to the demonstration of the power of the method. As done by other works in the fieldthese can all be left to future studies to explore. BM graphs are a map of the data cloud and hence havepotential to guide the segmentation thereof, this may be a fruitful line of enquiry for subsequent research ifit can be understood that there are no chances of failures occurring in any disregarded part of the space. Atthis stage shrinkage of the data cloud on the back of not having observed any bankruptcy outcomes wouldbe premature. Notwithstanding TDA BM represents a new system that offers a great amount to the studyof credit. This paper takes a critical first step on that representational journey.
References
Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of corporate bankruptcy.
The journal of finance , 23(4):589–609.Altman, E. I. (1983).
Corporate Financial Distress: A Complete Guide to Predicting, Avoiding, and Dealingwith Bankruptcy . New York: John Wiley & Sons.Altman, E. I., Iwanicz-Drozdowska, M., Laitinen, E. K., and Suvas, A. (2017a). Financial distress predictionin an international context: A review and empirical analysis of Altman’s Z-score model.
Journal ofInternational Financial Management & Accounting , 28(2):131–171.Altman, E. I., Iwanicz-Drozdowska, M., Laitinen, E. K., and Suvas, A. (2017b). Financial distress predictionin an international context: A review and empirical analysis of Altman’s Z-score model.
Journal ofInternational Financial Management & Accounting , 28(2):131–171.Barboza, F., Kimura, H., and Altman, E. (2017). Machine learning models and bankruptcy prediction.
Expert Systems with Applications , 83:405–417.Beaver, W. (1966). Financial ratios as predictors of failure.
Journal of Accounting Research , 4:71–111.Beaver, W. (1968). Alternative accounting measures as predictors of failure.
The Accounting Review , 43:113–122.Choi, H., Son, H., and Kim, C. (2018). Predicting financial distress of contractors in the constructionindustry using ensemble learning.
Expert Systems with Applications , 110:1–10.Cortes, C. and Vapenik, V. (1995). Support-vector networks.
Machine Learning , 20:273–297.13e Bock, K. (2017). The best of two worlds: Balancing model strength and comprehensibility in businessfailure prediction using spline-rule ensembles.
Expert Systems with Applications , 90:23–30.D(cid:32)lotko, P. (2019). Ball mapper: a shape summary for topological data analysis. arXiv preprintarXiv:1901.07410 .Dlotko, P. (2019).
BallMapper: Create a Ball Mapper graph of the input data . R package version 0.1.0.Gidea, M. and Katz, Y. (2018). Topological data analysis of financial time series: Landscapes of crashes.
Physica A: Statistical Mechanics and its Applications , 491:820–834.Li, H., Sun, J., Li, J. C., and Yan, X. Y. (2012). Forecasting business failure using two-stage ensemble ofmultivariate discriminant analysis and logistic regression.
Expert Systems , 30(5):385–397.Liu, C., Xie, J., Zhao, Q., Xie, Q., and Liu, C. ( in press 2019). Novel evolutionary multi-objective softsubspace clustering algorithm for credit risk assessment.
Expert Systems with Applications , 138.Nicolau, M., Levine, A., and Carlsson, G. (2011). Topology based data analysis identifies a group of breastcancers with a unique mutational profile and excellent survival.
Proceedings of the National Academy ofSciences , 107:7265–7270.Ohlson, J. (1980). Financial ratios and the probabilistic prediction of bankruptcy.
Journal of AccountingResearch , 18:109–131.Pearson, P., Muellner, D., and Singh, G. (2015).
TDAmapper: Analyze High-Dimensional Data UsingDiscrete Morse Theory . R package version 1.0.Petropoulos, A., Chatzis, S., and Xanthopoulos, S. (2016). A novel corporate credit rating system based onstudent’s-t hidden Markov models.
Expert Systems with Applications , 53:87–105.Singh, G., M´emoli, F., and Carlsson, G. E. (2007). Topological methods for the analysis of high dimensionaldata sets and 3d object recognition. In
SPBG , pages 91–100.Son, H., Hyun, C., Phan, D., and Hwang, H. ( in press 2019). Data analytic approach for bankruptcyprediction.
Expert Systems with Applications , 138.Vejdemo-Johansson, M., Carlsson, G., Lum, P. Y., Lehman, A., Singh, G., and Ishkhanov, T. (2012). Thetopology of politics: voting connectivity in the us house of representatives. In
NIPS 2012 Workshop onAlgebraic Topology and Machine Learning .Ziba, M., Tomczak, S. K., and Tomczak, J. M. ( in press 2019). Ensemble boosted trees with syntheticfeatures generation in application to bankruptcy prediction.