=Paper=
{{Paper
|id=Vol-1152/paper46
|storemode=property
|title=Classification of Financial Conditions of the Enterprises in Different Industries of Ukrainian Economy Using Bayesian Networks
|pdfUrl=https://ceur-ws.org/Vol-1152/paper46.pdf
|volume=Vol-1152
|dblpUrl=https://dblp.org/rec/conf/haicta/ChernyakC11
}}
==Classification of Financial Conditions of the Enterprises in Different Industries of Ukrainian Economy Using Bayesian Networks==
https://ceur-ws.org/Vol-1152/paper46.pdf
Classification of Financial Conditions of the Enterprises
in Different Industries of Ukrainian Economy Using
Bayesian Networks
Oleksandr Chernyak1, Yevgen Chernyak2
1
Department of Economic Cybernetics, Taras Shevchenko National University of Kyiv,
Kyiv, Ukraine, e-mail: chernyak@univ.kiev.ua
2
Department of International Economics, Taras Shevchenko National University of Kyiv ,
Kyiv, Ukraine, e-mail: evgenius206@mail.ru
Abstract. In this work the analysis of branches of Ukrainian economy was
done, particularly average financial parameters were found. For each parameter
the boundaries were determined which divide enterprises into 5 parts and allow
making more detailed ratings. The ratings were made by each parameter and
then the aggregate rating was found. The analysis of indices interrelation was
made using Bayesian network (BN). The coefficient of partial correlation in
BN was used to analyze the interrelation of indices. This subject-matter was
developed for Ministry of Industrial Policy of Ukraine. We recommend to use
cascade naive Bayes model in financial planning.
Keywords: financial indices, bankruptcy, Bayesian networks, naïve Bayes,
partial correlation.
1 Introduction
Each industry of economy is characterized by numerous features which distinguish
one particular industry from a variety of others, for instance such features are length
of operating cycle, requirement in available funds, tax policy of the state etc. The
peculiarity of every industry causes the difference in major financial indices. That is
why defining indices standards, their average values within the industry is an
important issue, which helps to describe the place of each enterprise in the industry
and also to compare industries with each other.
Setting the problem of standardizing of the financial indices estimation in frames
of industries at once raises a question about the necessity of calculating the
bankruptcy probability for each industry separately. To be mentioned, defining
bankruptcy probability following problems are faced: a) the fact of bankruptcy is
influenced not only by quantitative but also by qualitative indices like the possibility
of getting preferential crediting, support of the state, uninteresting of creditors to
confess a debtor to be a bankrupt; b) inadequate statistics of bankruptcies (procedure
________________________________
Copyright ©by the paper’s authors. Copying permitted only for private and academic purposes.
In: M. Salampasis, A. Matopoulos (eds.): Proceedings of the International Conference on Information
and Communication Technologies
for Sustainable Agri-production and Environment (HAICTA 2011), Skiathos, 8-11 September, 2011.
519
of bankruptcy stretches on a few years and fact of confession a bankrupt becomes
separate from the beginning of problems what could have been foreseen before by
the changes of financial indices); c) absence of adequate, representative base of
bankruptcies, which would allow estimating probability of bankruptcy within
industries.
An estimation of arising of overdue payments probability from the side of
enterprise would be more precisely, as problems which are described above level
with the estimation of non-fulfillment of creditor liabilities.
1 2
Financial Problems with liabilities
Bankruptcy
indices fulfillment
Fig. 1. Interaction financial indices and bankruptcy.
Qualitative indices influence the stage 2 in a greater measure than stage 1 (Fig. 1).
A fact and a term of payment delay is accurately fixed by credit organizations.
Statistics of overdue debt is collected by credit organizations, delays in payments
happen more frequently, so that estimation of probability for every industry is more
exact.
Therefore we stress on adequacy and possibility of estimation on the stage 1 and
mention that during the transition from the first stage to the second one accuracy is
being lost, and that is why estimation of the link “financial indices – bankruptcy” is
considered to be purposeless.
2 Criterion of Choosing the Standards
The only assumption we will use to make the analysis is that we know the direction
of index influence, in other words, increasing of the index influences positively the
enterprise state or contrariwise. The standard values of index can be based on
following considerations:
a) Through the influence of index on a resulting index (investigation of different
variants both negative and positive: fact of overdue debt, bankruptcy debt, increase
of net income). Recommended value of index would be the one which guarantees
fulfillment of obligations with certain probability.
b) Finding average value within the industry, medians or division into several groups
of sorted index values (more than 2) and finding average value for every group. This
approach is similar to rating; some part receives the highest rate and the other lowest.
Moreover, it is convenient to follow the indices moving from one group to other and
afterwards to check stability of a model.
The disadvantage of the first variant is difficulty to work with the correlated
indices because we have to define which of them exactly influences the result. The
exclusion of the strongly correlated indices from a model will not deprive us of
possibility to estimate standards for them. For example, we will have to use one of
the indices of liquidity only. The disadvantage of the second variant is a risk that
industry is in the phase of recession/growth and we will not get the standard values,
520
but correspondingly decreased (increased). The best way would be to compare the
results which were found by two methods and exactly to estimate in what parts the
whole set is divided by probability found by first method and what probabilities we
will get for the indices were found by dividing the set into equal parts.
3 Breaking on the Branch
Companies were divided into the industries according to The Classifier of Kinds of
Economic Activities (CKEA). But the way of fragmentation of CKEA was different
from the standard approach. We tried to pick out specific industries. For example,
insurance was picked out of financial sector, pharmacy – out of chemical industry.
Such method turned out to be appropriate, that was proved by the difference between
indicators.
We tried to provide the fragmentation as accurately as possible to be sure that the
company’s activity is the same that is in the industry. For example, how production
of metal should be divided from production of metal products, wholesale trade and
subsidiary services? Trade and subsidiary services may differ much one from
another. But at the same time it is inappropriate to combine them in one industry.
Therefore, companies were divided into the next classes: extraction, production,
engineering industry, wholesale trade, retail trade, rent and services.
Finally, we have got the following distribution of all the enterprises (376151) into
the industries: Auto – 9 384, Building – 41 831, Building materials – 12 271, Power
engineering – 4 427, Cafe and hotels – 10 400, Municipal service – 6 208, Culture
and education – 10 602, Wooding – 11 656, Medicine – 5 446, Metallurgy -5 469,
Real estate – 30 671, Fuel – 8 134, Polygraphy – 6 537, Cattle breeding – 6 473,
Textile – 6 396, Telecommunications – 14 568, Transport – 12 684, Tourism – 4 978,
Pharmacy – 5 481, Media – 3 529, Food industry – 28 058, Chemical – 7 061,
Wholesale trade – 50 019, Retail – 27 121, Machinery construction – 9 685,
Financial services – 15 685, Insurance – 726, Non-financial services – 16 892, Law –
3 759.
4 Dividing into Groups with the Further Purpose to Make Ratings
Now we will determine the average indices (see Table 1).
521
Table 1. Financial indices for the enterprises
Name Definition
1. Moment liquidity ratio: ML AHL / Lc
2. Current ratio: CR Al / Lc
3. General liquidity ratio: GL Aw / Lc
4. Current assets to equity ratio CA ((Eq
Eq Anon
An _ current) / Eq
5. Independence coefficient: IC OF / Eq
6. Return on assets: R(a) (NP 12) /( AA N)
7. Return on sales: R(s) (NI 12) /(NP N)
8. Inventory turn(days): IT N 30 ICavg
IC g/ NP
9. Debtors accounts turn(days): DT N 30 ARavg
AR / NP
10. Creditors accounts turn (days): CT N 30 AP
APavg / NP
11. Capital assets depreciation: D(ca) D / OC
12. The proportion of capital assets and
goods in process in total assets: CAinA (CA G) / А
AHL – high-liquidity assets, which consist of cash, their equivalents and current
financia investmens; Lc – current liabilitis which consist of short-term credits and
accounts with creditors; Al – liquid assets which consist of high-liquidity assets,
accounts receivable and billss of exchange received; Aw – working assets; Eq -
equity; Anon_current – non-current assets; OF- obtained funds; Eq- equity; NP – net
profit; N – number of monthes in period; NI – net income; AA – average value of
assets is calculated as (assets at the beginning of period + assets at the end of
period)/2; ICavg – average value of inventoryis calculated as (inventory a the
beginning of a period+inventory at the end of a period)/2; ARavg – the average sum
of the accounts receivable is calculated as (accounts receivable at the beginning of a
period + accounts receivable at the end of a period)/2; APavg – the average sum of
accounts payable is calculated as (accounts payable at the beginning of a
period+accounts payable at the end of a period)/2; OC – original cost of capital
assets; D – depreciation; CA-capital assets; G-goods-in-process; А-assets ( see
definitions in Van Horne and Wachowicz, (2008) or Stickney et al., 2010).
The period for NI, NP, AA , IT, DT , CT is quarter.
The differences between the branch indices showed the necessity of the work
which was done. The short-term indices them selves don’t allow to estimate the
enterprises adequately, their place in the whole field. The values of each index were
divided by quantity into 5 equal groups (see Table 2).
522
Table 2. Fragment division (value average of diapasons).
Food-industry ML CR GL CA IC R(a) R(s) IT
100% 5 10 15 10 50 1 1 500
80% 0,115 0,996 2,058 1,000 2,545 0,045 0,025 203,774
60% 0,022 0,508 1,154 0,967 0,557 0,002 0,005 108,812
40% 0,003 0,221 0,883 0,397 0,074 0,000 0,000 52,493
20% 0,000 0,045 0,426 -0,009 -1,155 -0,025 -0,024 18,246
0% 0,000 0,000 0,000 -10,000 -50,000 -1,000 -1,000 1,000
Insurance ML CR GL CA IC R(a) R(s) IT
100% 5 10 15 10 50 1 1 500
80% 2,627 5,479 7,569 0,981 0,281 0,188 0,300 28,630
60% 1,224 2,451 3,429 0,689 0,058 0,041 0,091 8,203
40% 0,408 1,333 1,554 0,184 0,008 0,004 0,031 4,418
20% 0,049 0,515 0,757 0,000 0,001 0,000 0,000 2,323
0% 0,000 0,000 0,000 -10,000 -50,000 -1,000 -1,000 1,000
It gives the possibility to determine the position of an enterprise by each of the
parameters more precisely. In this table we can see that 20% (after filtered of
information) enterprises of food industry have high value of ML in range [0,115; 5],
also 20% enterprises of insurance industry have high value of ML in range [2,627; 5].
40% enterprises of food industry have low value of ML in range [0; 0,003], also 20%
of insurance industry have low value of ML in range [0; 0,049]. We recommend use
this information in comparative analysis and determination position in industry.
After making the division for each enterprise by all the parameters the ratings
were made (0 means error, 2-6 according to the value of parameter: the less
parameter is the bigger the rating is, 1 was used for errors testing and isn’t applied as
a rating estimation). In this work there were considered both those coefficients which
increase is positive for an enterprise (return on assets, absolute liquidity) and those,
which increase is negative (depreciation, stock turn). For making the general rating
it’s necessary to make transformation so that the increase of the rating estimation by
all the parameters will cause increase of the general rating. Let’s convert the rating
estimation of the parameters, which increase is positive by the following
formula: R 8 R . This transformation leads to 2 6 , 3 5 , 4 4 , 5 3 ,
6 2.
Below is given the rating of three branches enterprises (Fig. 2):
523
Rate distributin of enterprises rating
0,14
0,12
0,1
Frequency
0,08
0,06
0,04
0,02
0
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Value of rating
Food industry Building Financial services
Fig. 2. Rate distribution of enterprises rating.
As a result we have a distribution close to even (it was expected because the
coefficients with the least correlation values were chosen for this rating). The
similarity for different branches is the evidence of the proposed method adequacy
and gives the possibility to compare enterprises from different fields by means of this
rating. For making rating 5 parameters were used: GL, IC, R(a), CT, D(ca). While
forming the rating the following indices were transformed: GL, IC, R(a); so the
higher the value of R is, the more risk there is for solvency in the future. Visual
similarity of distributions causes a question about the similar connection between the
values notwithstanding the branch. The more detailed research of the parameters
influence using Bayesian networks will be given further.
5 Construction of Bayesian Network
Bayesian networks are used for modelling subject domains which are
characterized by uncertainty. BNs are often used for the classification problem
(Friedman et al., 1997). There are the direction of using Bayesian networks in
economics: bankruptcy prediction (Sun and Shenoy, 2007), early warning of bank
failures (Sarkar and Sriram, 2001), credit risk modeling ( Pavlenko, Chernyak, 2010),
portfolio risk analysis and others.
Now we calculate the coefficients of correlation among the variables. In the
Table 3 represented values of the coefficients of correlation among the variables.
Colored cells represent coefficients of correlation which 0,1 .
524
Table 3. Value of the coefficients of correlation.
ML CR GL CA IC R(a) R(s) IT DT CT D(ca) CAinA
ML 1 0,49 0,43 -0,09 0,11 0,17 0,14 -0,09 -0,14 -0,24 0,01 -0,02
CR 1,00 0,70 -0,05 0,19 0,19 0,21 -0,14 0,22 -0,23 0,02 -0,19
GL 1,00 -0,09 0,26 0,26 0,28 0,10 0,06 -0,28 0,06 -0,21
CA 1,00 -0,52 -0,08 -0,09 0,08 0,08 0,07 0,24 -0,44
IC 1,00 0,23 0,19 0,01 0,03 0,04 -0,08 -0,09
R(a) 1,00 0,89 -0,01 -0,03 -0,11 -0,04 -0,09
R(s) 1,00 0,03 0,06 -0,09 -0,06 -0,06
IT 1,00 0,16 0,25 0,03 -0,14
DT 1,00 0,29 -0,05 -0,19
CT 1,00 -0,02 -0,11
D(ca) 1,00 -0,31
CAinA 1,00
According to the Table 3 results the connection graph was built ( Fig. 3). On this
graph R-rating is the value of the 0-level. ML, CR, GL, IC, R(a), R(s), CT, IT, D(ca)
are the first-level values (on the graph ML, GL are imaged not on the same level with
the other values of the first-level for the better visual perception and for showing the
influence of this value on the other, their interdependency).The second-level indices
(DT, CAinA, CA) have the biggest influence on the turnover indices (CT, IT) and
liquidity (ML, CR, GL). We chose 0,1 to be the level of link value.
In case if the influence of some index (eliminating the other indicators influence)
on rating is inessential (absolute value of partial correlation is less then 0,1) then this
index will be moved from the first- level to the second and then its influence on the
first-level indices will be estimated. If some index of the second-level will influence
all the first-level linked indices inessential then it will be moved into the third-level.
While moving into the lower level we “break” only the links with the indices of the
upper level (while moving the index into the second-level only the link with the
rating is broken).The following are the values of partial correlations for indices,
which are linked on the graph (Table 4):
525
Fig. 3. Dependences among the variables ( 0,1 ).
Table 4. Partial correlations (first-level indices).
corr(R;ML|CR) -0,38 corr(R;CR|ML) -0,05 corr(R;GL|ML) -0,09 corr(R;ML|GL) -0,39
corr(R;ML|IC) -0,49 corr(R;IC|ML) -0,15 corr(R;GL|R(s)) -0,21 corr(R;R(s)|GL) -0,20
corr(R;ML|R(a)) -0,47 corr(R;R(a)|ML) -0,36 corr(R;GL|IC) -0,23 corr(R;IC|GL) -0,12
corr(R;ML|R(s)) -0,48 corr(R;R(s)|ML) -0,19 corr(R;GL|R(s)) -0,21 corr(R;R(s)|GL) -0,20
corr(R;ML|CT) -0,48 corr(R;CT|ML) 0,01 corr(R;GL|IT) -0,36 corr(R;IT|GL) 0,52
corr(R;CR|GL) -0,11 corr(R;GL|CR) -0,12 corr(R;GL|CT) -0,25 corr(R;CT|GL) 0,04
corr(R;CR|R(s)) -0,22 corr(R;R(s)|CR) -0,22 corr(R;R(a)|R(s)) -0,16 corr(R;R(s)|R(a)) 0,08
corr(R;CR|IC) -0,24 corr(R;IC|CR) -0,14 corr(R;R(a)|GL) -0,34 corr(R;GL|R(a)) -0,33
corr(R;CR|R(a)) -0,22 corr(R;R(a)|CR) -0,36 corr(R;R(a)|CT) -0,39 corr(R;CT|R(a)) 0,08
corr(R;CR|R(s)) -0,22 corr(R;R(s)|CR) -0,22 corr(R;IC|R(a)) -0,10 corr(R;R(a)|IC) -0,36
corr(R;CR|IT) -0,23 corr(R;IT|CR) 0,45 corr(R;IC|R(s)) -0,14 corr(R;R(s)|IC) -0,24
corr(R;CR|CT) -0,25 corr(R;CT|CR) 0,05 corr(R;IT|CT) 0,45 corr(R;CT|IT) 0,00
According to the given calculations we come to the conclusion that the influence
CT on R is inessential so this index should be moved into the second-level. Colored
cells show insignificant correlations (absolute value of partial correlation is less then
0,1).
526
Now we only have to calculate the partial correlations between the first and
second-level taking into account translation of CT into the second-level. Before
moving CT we have following result (Table 5).
Table 5. Partial correlations (second-level indices).
corr(CR;DT|CT) 0,30 corr(CR;CT|DT) -0,31
corr(CR;DT|CAinA) 0,19 corr(CR;CAinA|DT) -0,16
corr(D(ca);CAinA|CA) -0,22 corr(D(ca);CA|CAinA) 0,23
corr(D(ca);CA|CAinA) 0,12 corr(D(ca);CAinA|CA) -0,22
corr(IT;DT|CAinA) 0,13 corr(IT;CAinA|DT) -0,12
corr(CT;DT|CAinA) 0,28 corr(CT;CAinA|DT) -0,05
Here we may conclude that the influence of CAinA on CT is inessential. After
moving CT we have following result (Table 6).
Table 6. Partial correlations (second-level indices, after moving CT).
corr(CR;DT|CT) 0,30 corr(CR;CT|DT) -0,31
corr(CR;CAinA|CT) -0,23 corr(CR;CT|CAinA) -0,27
corr(IT;DT|CT) 0,09 corr(IT;CT|DT) 0,22
corr(IT;CAinA|CT) -0,12 corr(IT;CT|CAinA) 0,25
corr(GL;CAinA|CT) -0,25 corr(GL;CT|CAinA) -0,32
corr(ML;DT|CT) -0,07 corr(ML;CT|DT) -0,21
We come to the conclusion that the link between IT and DT, ML and DT is
absent. As a result we get the following links (Fig. 4 – cascaded naïve Bayes model):
527
Fig.4. The structure for the cascade naïve Bayes model.
In the article (Sun and Shenoy, 2007) it was proposed to set the value level 0,1
analogically. Finding bigger threshold value of , the influence of the second-level
indices on first-level indices was confirmed, it didn’t lead to any changes in the graph
structure.
We recommend using cascade naive Bayes model while making financial
planning. For example, an enterprise seeks to minimize the risk of insolvency – it
should seek to decrease/increase the correspondent index (depending on the
correlation sign), taking into consideration that the first-level indices are influenced
by the second-level indices. Measure and character of the influence have to be
compared using the following tables of conditional probabilities (Tables 7, 8):
Table 7. Conditional probabilities of insolvency depending of moment liquidity ratio.
ML Rating
0 High Medium Low Sum
Error 10,61% 10,49% 4,29% 0,18% 25,58%
High 0,00% 0,47% 7,29% 7,08% 14,84%
Medium-High 0,00% 0,87% 8,42% 5,61% 14,91%
Medium 0,00% 1,15% 9,83% 3,91% 14,89%
Low-Medium 0,00% 2,59% 13,24% 2,64% 18,47%
Low 0,00% 3,62% 7,16% 0,55% 11,32%
Sum 10,61% 19,18% 50,24% 19,97% 100,00%
528
Table 8. Conditional probabilities of insolvency depending of return on assets.
ML Rating
0 High Medium Low Sum
Error 10.61% 10.49% 4.29% 0.18% 25.58%
High 0.00% 0.47% 7.29% 7.08% 14.84%
Medium-High 0.00% 0.87% 8.42% 5.61% 14.91%
Medium 0.00% 1.15% 9.83% 3.91% 14.89%
Low-Medium 0.00% 2.59% 13.24% 2.64% 18.47%
Low 0.00% 3.62% 7.16% 0.55% 11.32%
Sum 10.61% 19.18% 50.24% 19.97% 100.00%
Tables of conditional probabilities are very useful when we have incomplete
information. For example, value of ML – is known (high – level) and other
information – absent.
0,0047 (1)
P(R high/ ML high) 0,032 ,
0,0047 0,0729 0,0708
0,0729 (2)
P( R medium/ ML high) 0,49 ,
0,0047 0,0729 0,0708
0,0708 (3)
P( R low/ ML high) 0,478 ,
0,0047 0,0729 0,0708
P(R(a) high/ ML high) 0,032 0,1918 0,49 0,5024
5 0,478 0,1997 0,3476 . (4)
6 Conclusions
The main idea of this research is to demonstrate the differences between the
financial indices for different industries. The analysis of indices interrelation was
made using Bayesian network. The coefficient of partial correlation in BN was used
to analyze the interrelation of indices. While making ratings there was made an
assumption about the independence of the distribution form in which the rating
frequency is described for all enterprises from branch.
The explanation of the inadequacy of the bankruptcy probability estimation is
given (especially in terms of Ukrainian economy). The bigger accuracy of the
solvency estimation is pointed out. The assumption is made about keeping the
coefficients proportions in discriminatory models of solvency estimation
notwithstanding the branch.
This subject-matter is being developed for Ministry of Industrial Policy with the
purpose of temporary revelation of the enterprises subordinate to these Ministry
financial problems.
529
References
1. Friedman, N., Geiger, D., Goldszmidt, M. (1997) Bayesian network classifiers.
Machine Learning, 29, p.131-163.
2. Pavlenko, T., Chernyak, O. (2010) Credit risk modeling using bayesian networks.
International Journal of Intelligent Systems, 25, N4, p.326 – 344.
3. Sarkar, S., Sriram, R.S. (2001) Bayesian models for early warning of bank
failures. Management science, 47, N 11, p.1457-1475.
4. Stickney, C.P., Weil, R.L., Francis, J. (2010) Financial Accounting: An
Introduction to Concepts, Methods and Uses. 13th Edition. South-Western:
Cengage Learning.
5. Sun, L., Shenoy, P.P. (2007) Using Bayesian networks for bankruptcy prediction:
Some methodological issues. European Journal of Operation Research, 180,
p.738-753.
6. Van Horne, J.C., Wachowicz, J.M. (2008) Fundamentals of Financial
Management. 12th Edition. Lebanon, Indiana, USA: FT Prentice Hall.
530