Grammatical Evolution and Corporate Failure Prediction

Anthony Brabazon Michael O’Neill

Dept. Of Accountancy, Dept. Of Computer Science And Information Systems,
University College Dublin, Ireland. University of Limerick, Ireland.

Robin Matthews Conor Ryan

Center for International Business Policy Dept. Of Computer Science And Information Systems,
Kingston Business School, London University of Limerick, Ireland.

Abstract defects, leading to poor decisions, leading to financial dete-

rioration and finally resulting in corporate collapse [3] [20].
This study examines the potential of Grammatical Most attempts to predict corporate failure implicitly assume
Evolution to uncover a series of useful rules which that management decisions critically impact on firm perfor-
can assist in predicting corporate failure using in- mance [5] [20]. The premise of this paper is that a series of
formation drawn from financial statements. A sam- poor decisions lead to a deterioration in the financial health of
ple of 178 publically quoted, failed and non-failed the firm and finally to its demise. Although the decisions are
US firms, drawn from the period 1991 to 2000 are not directly observable, their consequent affect on the finan-
used to train and test the model. The preliminary cial health of the firm can be observed.
findings indicate that the methodology has much Previous studies have utilised a wide variety of explanatory
potential. variables in the construction of corporate distress models 1 ,
including variables drawn from the financial statements of
firms, from financial markets, general macro-economic vari-
1 Introduction ables [29], and non-financial, firm-specific information, in-
cluding director turnover [27]. In this study, we limit our at-
The objective of this study is to determine whether an evolu- tention to information drawn from the financial statements of
tionary automatic programming methodology, Grammatical firms.
Evolution, is capable of uncovering useful structure in finan-
cial ratio information which can be used to predict corporate 1.1 Potential for application of evolutionary automatic
failure. programming
Corporate failure is an essential component of an efficient
market economy, allowing the recycling of financial, human There are a number of reasons to suppose that the use of an
and physical resources into more productive organisations evolutionary automatic programming (EAP) approach such as
[10] [32]. However, many parties including shareholders, Genetic Programming (GP) or GE can prove fruitful in the
providers of debt finance, employees, suppliers, customers, prediction of corporate failure. The problem domain is char-
managers and auditors have an interest in the financial health acterised by a lack of a strong theoretical framework, with
of organisations as corporate failure can impose significant many plausible, competing explanatory variables. The selec-
private costs on all these groups. Even where total failure can tion of quality explanatory variables and model form repre-
be averted by firm reorganization, the costs of major restruc- sents a high-dimensional combinatorial problem, giving rise
turing can be as high as 12% to 19% of firm value [41]. If a to potential for an EAP methodology. Use of EAP also facil-
trajectory leading to corporate failure can be identified suffi- itates the development of complex fitness functions including
ciently early to allow successful intervention, these costs can discontinuous, non-differentiable functions. This is of partic-
be reduced. [14] suggest that indicators of corporate failure ular importance in a prediction domain as fitness criteria may
can be present up to ten years prior to final failure, provid- be complex. Generally, the cost of misclassifications of fail-
ing an opportunity for construction of models which predict ing / non-failing firms will be asymmetric. Another useful
corporate failure. feature of an EAP approach is that it can produce human-
readable rules that have the potential to enhance understand-
Corporate failure can arise for many reasons. It may occur ing of the problem domain.
due to a single catastrophic event or it may be the end result
of a lengthy process of decline. Under the second perspective, 1
[3] and [22] provide good reviews of the development of empir-
corporate failure is a process which starts with management ical research in bankruptcy prediction.
1.2 Motivation for study have included neural networks [34] [33] [40], genetic algo-
rithms [39] [19] and hybrid neural network / genetic algorithm
This study was motivated by a number of factors. Although a models [6].
substantial volume of research utilising traditional statistical
modelling techniques has been undertaken in the corporate 2.1 Definition of Corporate Failure
failure domain, to date only a limited number of studies have
applied GA / GP methodologies [39] [19] [6]. This study No unique definition of corporate failure exists [3]. Possi-
builds on these initial studies and adopts a novel evolutionary ble definitions range from failure to earn an economic rate of
automatic programming approach. return on invested capital given the risk of the business, to
legal bankruptcy followed by liquidation of the firm’s assets.
1.3 Structure of paper Any attempt to uniquely define corporate failure is likely to
prove problematic. While few publicly quoted companies fail
This contribution is organised as follows. Section 2 provides a in any given year2 , poorer performers are liable to acquisition
short discussion of prior literature in the corporate failure do- by more successful firms. Thus, two firms may show a sim-
main and outlines the definition of corporate failure employed ilar financial trajectory towards failure, but one firm may be
in this study. Section 3 provides an introduction to Grammat- acquired and ‘turned-around’ whilst the other may fail.
ical Evolution. Section 4 describes both the data utilised, and The definition of corporate failure adopted in this study is
the model development process adopted in this paper. Section the entry of a firm into Chapter 7 or Chapter 11 of the US
5 provides the results of the constructed model. Finally, con- Bankruptcy code. The selection of this definition provides
clusions and a discussion of the limitations of the contribution an objective benchmark as the occurrence, and date of oc-
are provided in Section 6. currence, of either of these events can be determined through
examination of regulatory filings. Chapter 7 covers corpo-
2 Background rate liquidations and Chapter 11 covers corporate reorgani-
zations, which usually follow a period of financial distress.
Formal research into the prediction of corporate failure has Under Chapter 11, management is required to file a reorgan-
a long history [12] [35] [16]. Early statistical studies such isation plan in bankruptcy court and seek approval for this
plan. When the court grants approval for the plan the firm is
as [7], adopted a univariate methodology, identifying which
accounting ratios had greatest classification accuracy when released from Chapter 11 bankruptcy and continues to trade.
In most cases, Chapter 11 reorganisations involve significant
identifying failing and non-failing firms. Although this ap-
financial losses for both shareholders [30] and creditors [11]
proach did demonstrate classification power, it suffers from
the shortcoming that a single weak financial ratio may be off- of the distressed firm. [23], in a study of the outcomes of
Chapter 11 filings, found that ‘there were few sucessful re-
set (or exacerbated) by the strength (or weakness) of other
financial ratios. [1] addressed this issue by employing a lin- organisations’ (p. 125), despite a perception that some man-
ear discriminant analysis (LDA) model, which utilised both agement teams were using Chapter 11 filings as a deliberate
strategy for dealing with certain firm specific events such as
financial and market data concerning a firm, and this was
found to improve the classification accuracy of the developed onerous labor contracts or produce liability claims 3 .
models. The discriminant function which produced the best
classification performance in Altman’s 1968 study was: 2.2 Explanatory variables utilised in prior literature

Z = 012 1 + 014 2 + 033 3 + 006 4 + 999

: X : X : X : X : X5
Five groupings of explanatory variables, drawn from financial
statements, are given prominence in prior literature [4]:

where: i. Liquidity
X1 = working capital to total assets
X2 = retained earnings to total assets ii. Debt
X3 = earnings before interest and taxes to total assets
X4 = market value of equity to book value of total debt iii. Profitability
X5 = sales to total assets iv. Activity / Efficiency
v. Size
LDA assumes both multi-variate normality and the equality
2
of the covariance matrices of each classification group. Gen- [42] reports that this rate is less than 0.75% in the US and [22]
suggests that the rate is below 2% in the UK.
erally, these assumptions do not hold for financial ratio data. 3
[23] report that out of a sample of 73 firms entering Chapter 11
Other statistical methodologies which have been applied in- between 1980 and 1986, only 44 were sucessfully reorganized with
clude logit and probit regression models [13] [42] [24]. In only 15 of these firms emerging from Chapter 11 with more than
recent times, methodologies applied to this problem domain 50% of their prebankruptcy assets.
Liquidity refers to the availability of cash resources to meet 3 Grammatical Evolution
short-term cash requirements. Debt measures focus on the
relative mix of funding provided by shareholders and lenders. Grammatical Evolution (GE) is an evolutionary algorithm
Profitability considers the rate of return generated by a firm, in that can evolve computer programs in any language. Rather
relation to its size, as measured by sales revenue and/or asset than representing the programs as syntax trees, as in tradi-
base. Activity measures consider the operational efficiency of tional GP [18], a linear genome representation is adopted. A
the firm in collecting cash, managing stocks and controlling genotype-phenotype mapping process is used to generate the
its production or service process. Firm size provides infor- output program for each individual in the population. Each in-
mation on both the sales revenue and asset scale of the firm dividual, a variable length binary string, contains in its codons
and also provides a proxy metric on firm history. The group- (groups of n-bits, where n equals 8 here) the information to se-
ings of potential explanatory variables can be represented by lect production rules from a Backus Naur Form (BNF) gram-
a wide range of individual financial ratios, each with slightly mar. The BNF is a plug-in component to the mapping process,
differing information content. The groupings themselves are that represents the output language in the form of production
interconnected, as weak (or strong) financial performance in rules. It is comprised of a set of non-terminals that can be
one area will impact on another. For example, a firm with a mapped to elements of the set of terminals, according to the
high level of debt, may have lower profitability due to high in- production rules.
terest costs. Whatever modelling methodology is applied, the
An example excerpt from a BNF grammar is given below.
initial problem is to select a quality set of model inputs from
These productions state that S can be replaced with either one
a wide array of possible financial ratios, and then to combine
of the non-terminals expr, if-stmt, or loop.
these ratios using suitable weightings in order to construct a
high quality classifier. Given the large search space, an evo- S ::= expr (0)
lutionary automatic programming methodology has promise. | if-stmt (1)
| loop (2)

The grammar is used in a generative process to construct

a program by applying production rules, selected by the
2.3 Results of Prior Literature genome, beginning from the start symbol of the grammar.
In order to select a rule in GE, the next codon value on the
genome is read and placed in the following formula:
Earlier studies [1] [2] [3] [4] [9] [6] have suggested that the
classification accuracy of failure models increases rapidly as Rule = C odon V alue M OD # Rules

the date of final failure approaches. Generally, results indi-

cate that the most significant deterioration in financial ratios If the next codon integer value was 4, given that we have
3 rules to select from as in the above example, we get
occurs in the third year prior to eventual failure. Although
sample sizes, dates and methodologies differ between studies, 4 M OD 3 = 1 . S will therefore be replaced with the
these findings have been replicated in a broad series of stud- non-terminal if-stmt.
ies. In Altman’s 1968 study [1], the developed LDA model Beginning from the left hand side of the genome codon in-
correctly identified (in-sample) 95% of failing firms one year teger values are generated and used to select rules from the
prior to failure. The classification accuracy fell to 72% and BNF grammar, until one of the following situations arise:
48% in the second and third year prior to failure. [2] demon-
strated a classification accuracy (in-sample) of approximately i. A complete program is generated. This occurs when all
93% in the year prior to failure declining to 68% four years the non-terminals in the expression being mapped, are
prior to failure. [40] report in-sample classifications of 98.7% transformed into elements from the terminal set of the
for a neural network model, one-year prior to failure, and 95% BNF grammar.
for a logit model on the same data. [39] reports in-sample
classification accuracy for a GA based model of approxi- ii. The end of the genome is reached, in which case the
mately 97%, one year prior to failure. In-sample classification wrapping operator is invoked. This results in the re-
accuracies provide a limited assessment of model generalis- turn of the genome reading frame to the left hand side
ability. Enhanced in-sample classification accuracies could of the genome once again. The reading of codons will
result from data-mining. Hence, in this study, developed mod- then continue unless an upper threshold representing the
els are solely assessed based on classification performance on maximum number of wrapping events has occurred dur-
out-of-sample data. ing this individual’s mapping process. This threshold is
currently set to ten events.
A description of the evolutionary automatic programming
system used to evolve rules for prediction of corporate fail- iii. In the event that a threshold on the number of wrap-
ure is provided in the next section. ping events is exceeded and the individual is still incom-
 pletely mapped, the mapping process is halted, and the The choice of explanatory variables is hindered by the lack of
individual assigned the lowest possible fitness value. a clear theoretical framework which explains corporate fail-
ure [5] [38] [40]. Most empirical work on corporate failure
GE uses a steady state replacement mechanism, such that, adopts an ad-hoc approach to variable selection. Prior to the
two parents produce two children the best child replacing the selection of the potential explanatory variables for inclusion
worst individual in the current population if the child has a in this study, a total of ten previous studies were examined
greater fitness. In the case where both children have the same [7] [1] [2] [9] [24] [33] [17] [6] [37] [21]. These studies em-
fitness and are better than the current population worst, a child ployed a total of 58 distinct ratios divided amongst the five
is chosen at random. The standard genetic operators of point classifications noted by [4]. A subset of 22 of the most com-
mutation, and crossover (one point) are adopted. It also em- monly used financial ratios was selected for this study. The
ploys a duplication operator that duplicates a random number selected ratios were:
of codons and inserts these into the penultimate codon posi-
tion on the genome. A full description of GE can be found in i. EBIT / Sales

[26] [25] [31]. ii. EBITDA / Sales

iii. EBIT / Total Assets

4 Problem Domain & Experimental Approach iv. Gross Profit / Sales

v. Net Income / Total Assets

This section describes both the data utilised by, and the model
development process adopted in, this study. vi. Net Income / Sales

vii. Return on Assets

4.1 Sample Definition and Model Data
viii. Return on Equity

A total of 178 firms were selected judgementally (89 failed, ix. Return on Investment
89 non-failed), from the Compustat Database [8] 4 . The crite- x. Cash / Sales
ria for selection of the failed firms were:
xi. Sales / Total Assets

i. Inclusion in the Compustat database in the period 1991- xii. Inventory / Cost of Goods Sold

2000 xiii. Inventory / Working Capital

ii. Existence of required data for a period of three years xiv. Fixed Assets / Total Assets

prior to entry into Chapter 7 or Chapter 11 xv. Retained Earnings / Total Assets

iii. Sales revenues must exceed $1M xvi. Cash from Operators / Sales

xvii. Cash from Operations / Total Liabilities

The first criterion limits the study to publicly quoted, US cor- xviii. Working Capital / Total Assets
porations. The second criterion injects an element of bias into
xix. Quick Assets / Total Assets
the sample in that companies without a three year financial
history prior to entering Chapter 7 or Chapter 11 are omitted. xx. Total Liabilities / Total Assets
Twenty-two potential explanatory variables, are collected for xxi. Leverage
each firm for the three years prior to entry into Chapter 7 or
xxii. EBIT / Interest
Chapter 115 . For every failing firm, a matched non-failing
firm is selected. They are matched both by industry sector
and size (sales revenue three years prior to failure) 6 . The set 5 Results
of 178 matched firms are randomly divided into model build-
ing (128 firms) and out-of-sample (50 firms) datasets. The
Accuracy of the developed models is assessed based on the
dependant variable is binary (0,1), representing either a non-
overall classification accuracy arising in both the model-
failed or a failed firm.
building and out-of-sample datasets. For simplicity, the cost
4
Firms from the financial sector were excluded on grounds of of each type of classification error is assumed to be symmet-
lack of comparability of their financial ratios with other firms in the ric in this study. The fitness function could be easily altered
sample.
5 to bias the model development process to minimise a specific
The date of entry into Chapter 7 or Chapter 11 was determined
by examining regulatory filings for each firm. type of classification error if required, and later studies will
6
It is recognised that the use of an equalised, matched sample address this issue.
entails sampling bias and eliminates firm size and industry nature as
potential explanatory variables (see [22] for a detailed discussion of The classification problem which plays an important role in
these points). It is noted that utilising an unmatched sample imposes decision-making, consists of assigning observations to dis-
its own bias. joint groups [28]. The decision scenario faced in this study
comprises a binary classification. In general, the construc- models generated falls off gracefully as we move out each
tion of classifier systems such as linear discriminant analysis, year. It is interesting to note that out-of-sample there is no
logit or ANN models consists of two components, the deter- performance difference between the evolved models in peri-
mination of a valuation rule which is applied to each observa- ods T1 and T2, both giving 80% correct classifications.
tion, and the determination of a ‘cut-off’ value. The grammar
The best classifiers evolved for each period are given in Table
adopted in this study is given below and its output is inter-
2.
preted using a fixed 0.5 cut-off value to produce a classifica-
tion.
5.1 Discussion
lc : output = expr ;
expr : ( expr ) + ( expr )
| coeff * var
The classification results of the evolved models show
var : var1[index] | var2[index] | var3[index] promise. Despite drawing a sample from a wide variety of
| var4[index] | var5[index] | var6[index]
| var7[index] | var8[index] | var9[index]
industrial sectors, the models demonstrate a high classifica-
| var10[index] | var11[index] | var12[index] tion accuracy in and out-of-sample, which degrades grace-
| var13[index] | var14[index] | var15[index]
| var16[index] | var17[index] | var18[index]
fully rather than suddenly in the third year prior to failure.
| var19[index] | var20[index] | var21[index] Although the evolved models were free to select from twenty-
| var22[index]
coeff : ( coeff ) op ( coeff )
two potential explanatory variables, it is notable that each
| float model only employed a small subset of these. This lends sup-
op : + | - | *
float : 20 | -20 | 10 | -10 | 5 | -5 | 4 | -4
port to the proposition that many financial ratios have similar
| 3 | -3 | 2 | -2 | 1 | -1 | .1 | -.1 information content and that classification accuracy is not en-
hanced through the construction of models with a large num-
ber of these ratios. It is also notable that each model has (ap-
The above grammar generates classifiers of the form: proximately) included one variable drawn from the four main
categories of explanatory variables suggested in the corpo-
output =( < some expression > varX )+ :: rate failure literature (Liquidity, Debt, Profitability, and Ac-
+( < some expression > varY ) tivity/Efficiency), lending empirical support to earlier work 8 .
The risk factors suggested by each model differ somewhat and
Any combination of the twenty two explanatory variables can contain some less-intuitive but nonetheless plausible findings.
be exploited by an evolved classifier, including zero or more
Examining the best classifer evolved for T1 suggests that risk
occurrences of any one variable. This is in contrast to an LDA
factors include low return on assets, low retained earnings
approach where classifiers would generally utilise all the ex-
and a high ratio of total liabilities to total assets, which con-
planatory variables within the expression. In the LDA case,
cords with financial intuition. Less obviously, a high ratio
of course some of those variables could be switched off by
of inventory to net liquid assets (inventory+receivables+cash-
multiplying their value by zero.
payables) is also a risk factor, possibly resulting from deple-
Three series of models were constructed using explanatory tion of cash or build-up of inventories as failure approaches.
variables drawn from one, two and three years (T1, T2 and
Risk factors for firms at T2 include low return on assets and
T3) prior to failure. For each set of models, 30 runs were
a low ratio of earnings to interest costs. Less intuitive risk
conducted using population sizes of 500, running for 100 gen-
factors indicated are a low ratio of fixed assets to total assets
erations, adopting one-point crossover at a probability of 0.9,
and a high ratio of sales to total assets. The former could in-
and bit mutation at 0.01, along with a steady state replacement
dicate firms with a lower safety cushion of saleable resources
strategy.
which could be sold to stave-off collapse, the latter could be
A plot of the mean average and mean best fitness values over serving as a proxy variable for firms with rapid sales growth.
the 30 runs for each time period can be seen in Figure 1. Over-rapid sales growth can be a danger signal, indicating that
management resources are being spread too-thinly.
Years Prior to Failure In Sample Out Of Sample
1 85.9% 80% Finally, risk factors indicated for firms at T3 include low re-
2 82.8% 80% turn on assets, a low ratio of profit to interest charge, a low
3 75.8% 70% level of cash generated from operations and as for T2, a high
ratio of sales to total assets.
Table 1: The best classification accuracies reported for each Although each model is evolved seperately, the general form
of the three years prior to failure. of each model appears consistent with the hypothesis that
The best individuals evolved for each period are reported in els rejects a null hypothesis, at the 5% level, that the out-of-sample
classification accuracies are not significantly better than chance.
Table7 1. In-sample it can be seen that the performance of the 8
Size, the fifth category, is not considered in this study due to the
7
Calculation of Press’s Q statistic [15] for each of these mod- matching process utilized.
 Grammatical Evolution - Corporate Failure Grammatical Evolution - Corporate Failure
0.9 0.9

0.8 0.8

0.7 0.7

0.6 T1 0.6 T1

Average Fitness
Best Fitness
T2 T2
T3 T3

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Generation Generation

Figure 1: A comparison of the mean best fitness between T1, T2, and T3 (left), and of the mean average fitness values (right)
for the same time periods.

Years Prior to Failure Best Classifier

1 output = -3*var7-5*var8+3*var17-20*var19+4*var24
2 output = -2*var8+10*var15-10*var18-2*var25
3 output = -4*var8+20*var15-72.9*var20-10*var25

Table 2: The best classifiers evolved for each of the years analysed.

there is a financial trajectory towards failure. Low profits formance could further improve classification accuracy. An-
and high interest payments as a percentage of profits in pe- other limitation of all models of corporate distress is that the
riods T3 and T2 indicate a firm in financial difficulties, with underlying relationships may not be stationary [4] [17]. Ac-
an erosion of the safety cushion provided by high levels of counting standards and economic environment faced by firms
(saleable) fixed assets indicated in the risk factors at T2. The will vary over time. Finally, the firms sampled in this study
final year prior to failure sees additional risk factors indicated are relatively large and are publically quoted. Thus, the find-
by high levels of debt and reducing cash balances / inventory ings of this study may not extend to small businesses.
build-up.
Despite these limitations, the high economic and social costs
of corporate failure imply that models which can indicate de-
6 Conclusions & Future Work clining financial health will have utility. Given the lack of a
clear theory underlying corporate failure, empirical modelling
GE was shown to successfully evolve useful rules for pre- usually adopts a combinatorial approach, a task for which GE
diction of corporate failure with a performance equivalent to is well suited. The results of this preliminary study indicate
that reported in prior studies. In assessing the performance of that GE has useful potential for the construction of corporate
the developed models, a number of caveats must be borne in failure models.
mind. The premise underlying this paper (and all empirical
work on corporate failure prediction) is that corporate failure
is a process, commencing with poor management decisions,
References
and that the trajectory of this process can be tracked using ac-
[1] Altman, E. (1968). ‘Financial Ratios, Discriminant
counting ratios. This approach does have inherent limitations.
Analysis and the Prediction of Corporate Bankruptcy’,
It will not forecast corporate failure which results from a sud-
Journal of Finance, 23:589-609.
den environmental event. It is also likely that the explanatory
variables utilised contain noise. Commentators [5] [36] have [2] Altman, E., Haldeman, R. and Narayanan, P. (1977).
noted that managers may attempt to utilise creative account- ‘ZETA Analysis: A new model to identify bankruptcy
ing practices to manage earnings and / or disguise signs of risk of corporations’, Journal of Banking and Finance,
distress. Additionally, financial data is produced on a time- 29-54.
lagged basis. Although not undertaken in this preliminary
study, the incorporation of non-financial qualitative explana- [3] Altman, E. (1993). Corporate Financial Distress and
tory variables or variables related to the firm’s share price per- Bankruptcy, New York: John Wiley and Sons Inc.
 [4] Altman, E. (2000). ‘Predicting Financial Distress of [19] Kumar, N., Krovi, R. and Rajagopalan, B. (1997). ‘Fi-
Companies: Revisiting the Z-score and Zeta models’, nancial decision support with hybrid genetic and neu-
http://www.stern.nyu.edu/ ealtman/Zscores.pdf, Octo- ral based modelling tools’, European Journal of Oper-
ber 2001. ational Research, 103:339-349.

[5] Argenti, J. (1976). Corporate Collapse: The Causes [20] McRobert, A. and Hoffman, R. (1997). Corporate
and Symptoms, London: McGraw-Hill. Collapse: An early warning system for lenders, in-
vestors and suppliers, Roseville:NSW, McGraw-Hill
[6] Back, B., Laitinen, T., Sere, K. and van Wezel, M. (Australia).
(1996). ‘Chosing Bankruptcy Predictors using Dis-
criminant Analysis, Logit Analysis and Genetic Al- [21] Moody’s (2000). ‘RiskCalc For Private
gorithms’, Technical Report no. 40, Turku Centre for Companies: Moody’s Default Model’,
Computer Science, Turku School of Economics and http://riskcalc.moodysrms.com/us/research/crm/56402.pdf,
Business Administration.. July 2001.

[7] Beaver, W. (1968). ‘Financial Ratios as Predictors of [22] Morris, R. (1997). Early Warning Indicators of Corpo-
Failure’, Journal of Accounting Research - Supplement: rate Failure: A critical review of previous research and
Empirical Research in Accounting, 71-102. further empirical evidence, London: Ashgate Publish-
ing Limited.
[8] Compustat Database. http://www.compustat.com
[23] Moulton, W. and Thomas, H. (1993). ‘Bankruptcy As
[9] Dambolena, I. and Khoury, S. (1980). ‘Ratio Stability a Deliberate Strategy: Theoretical Considerations and
and Corporate Failure’, Journal of Finance, 35(4):1017- Empirical Evidence’, Strategic Management Journal,
1026. 14:125-135.
[10] Easterbrook, F. (1990). ‘Is Corporate Bankruptcy Effi- [24] Ohlson, J. (1980). ‘Financial Ratios and the Probabilis-
cient?’, Journal of Financial Economics, 27:411-417. tic Prediction of Bankruptcy’, Journal of Accounting
Research, 18:109-131.
[11] Ferris, S., Jayaraman, N. and Makhija, A. (1996). ‘The
Impact of Chapter 11 filings on the Risk and Return of [25] O’Neill, M. (2001). Automatic Programming in an Ar-
Security Holders, 1979-1989’, Advances in Financial bitrary Language: Evolving Programs in Grammatical
Economics, 2:93-118. Evolution. PhD thesis, University of Limerick, 2001.

[12] Fitzpatrick, P. (1932). A Comparison of the Ratios of [26] O’Neill M., Ryan C. (2001) Grammatical Evolution,
Successful Industrial Enterprises with Those of Failed IEEE Trans. Evolutionary Computation. 2001.
Companies, Washington: The Accountants’ Publishing
Company. [27] Peel, M., Peel, D. and Pope, P. (1986). ‘Predicting Cor-
porate Failure: Some Results for the UK Corporate Sec-
[13] Gentry, J., Newbold, P. and Whitford, D. (1985). ‘Clas- tor’, Omega International Journal of Management Sci-
sifying Bankrupt Firms with Funds Flow Components’, ence, 14:5-12.
Journal of Accounting Research, 23(1):146-160.
[28] Pendharkar, P. (2001). An empirical study of design and
[14] Hambrick, D. and D’Aveni, R. (1988). ‘Large Cor- testing of hybrid evolutionary-neural approach for clas-
porate Failures As Downward Spirals’, Administrative sification, Omega, 29:361-374.
Science Quarterly, 33:1-23.
[29] Rose, P., Andrews, W.and Giroux, G. (1982). ‘Predict-
[15] Hair, J., Anderson, R., Tatham, R. and Black, W. ing Business Failure: A Macroeconomic Perspective’,
(1998). Multivariate Data Analysis, Upper Saddle Journal of Accounting, Auditing and Finance, Fall:20-
River, New Jersey: Prentice Hall. 31.

[16] Horrigan, J. (1965). ‘Some empirical bases of financial [30] Russel, P., Branch, B. and Torbey, V. (1999). ‘Market
ratio analysis’, The Accounting Review, July 1965, 558- Valuation of Bankrupt Firms: is there an anomaly?’,
568. Quarterly Journal of Business and Economics, 38:55-
76.
[17] Kahya, E. and Theodossiou, P. (1996). ‘Predicting
Corporate Financial Distress: A Time-Series CUSUM [31] Ryan C., Collins J.J., O’Neill M. (1998). Grammatical
Methodology’, Review of Quantitative Finance and Ac- Evolution: Evolving Programs for an Arbitrary Lan-
counting, 13:71-93. guage. Lecture Notes in Computer Science 1391, Pro-
ceedings of the First European Workshop on Genetic
[18] Koza, J. (1992). Genetic Programming. MIT Press. Programming, 83-95, Springer-Verlag.
[32] Schumpeter, J. (1934). The Theory of Economic Devel-
opment, Cambridge, MA: Harvard Business Press.

[33] Serrano-Cina, C. (1996). ‘Self organizing neural net-

works for financial diagnosis’, Decision Support Sys-
tems, 17:227-238.
[34] Shah, J. and Murtaza, M. (2000). ‘A Neural Network
Based Clustering Procedure for Bankruptcy Prediction’,
American Business Review, 18(2):80-86.
[35] Smith, R. and Winakor, A. (1935). ‘Changes in the Fi-
nancial Structure of Unsuccessful Corporations’, Uni-
versity of Illinois, Bureau of Business Research, Bul-
letin No. 51.
[36] Smith, T. (1992). Accounting for Growth. London: Cen-
tury Business.
[37] Sung, T., Chang, N. and Lee, G. (1999). ‘Dynamics of
Modelling in Data Mining: Interpretative Approach to
Bankruptcy Prediction’, Journal of Management Infor-
mation Systems, 16(1):63-85.
[38] Trigueiros, D. and Taffler R. J. (1996). ‘Neural Net-
works and Empirical Research in Accounting’, Ac-
counting and Business Research, 26(4):347 - 355.
[39] Varetto, F. (1998). ‘Genetic algorithms in the analysis
of insolvency risk’, Journal of Banking and Finance,
22(10):1421 - 1439.
[40] Wilson, N., Chong, K. and Peel, M. (1995). ‘Neural
Network Simulation and the Prediction of Corporate
Outcomes: Some Empirical Findings’, International
Journal of the Economics of Business, 2(1):31-50.
[41] Wruck, K. (1990). ‘Financial Distress, Reorganization
and Organizational Efficiency’, Journal of Financial
Economics, 27(2):419-444.
[42] Zmijewski, M. (1984). ‘Methodological Issues Related
to the Estimation of Financial Distress Prediction Mod-
els’, Journal of Accounting Research - Supplement, 59-
82.