=Paper=
{{Paper
|id=Vol-2927/paper10
|storemode=property
|title=Clustering Approach to Analysis of the Credit Risk and Profitability for Nonbank Lenders
|pdfUrl=https://ceur-ws.org/Vol-2927/paper10.pdf
|volume=Vol-2927
|authors=Andrii Kaminskyi,Maryna Nehrey
}}
==Clustering Approach to Analysis of the Credit Risk and Profitability for Nonbank Lenders==
https://ceur-ws.org/Vol-2927/paper10.pdf
125
CLUSTERING APPROACH TO ANALYSIS OF THE
CREDIT RISK AND PROFITABILITY FOR
NONBANK LENDERS
Andrii Kaminskyi1, Maryna Nehrey2
1Department of Economic Cybernetics, Taras Shevchenko National University of Kyiv,
e-mail: kaminskyi.andrey@gmail.com
2Department of Economic Cybernetics, National University of Life and Environmental
Sciences of Ukraine, e-mail: marina.nehrey@gmail.com
Abstract. The paper is devoted to application customer profitability analysis to
nonbanking lenders which predominantly focus on payday loans. The “Whale
curve” has been constructed and special clusters were singled out. The approach
based on joint together customer profitability management and credit risk
management are considered. One significant effect was marked and grounding
that higher risk interconnects with overpayments. The approach of fuzzy
clustering was applied as the second approach to clustering. Such approaches
may be considered as the basis of loan granting strategies elaborating.
Keywords: Customer Profitability Analysis, Clustering, Risk-Management,
Nonbank Lending, Payday Loans, Machine Learning.
1. Introduction
The modern financial system is transforming. The intensive growth of fintech
organizations reshapes the landscape of classical financial services. This turns critically
on banking. Harvard Business Review Analytic Services surveyed more than 300
executives in classical financial institutions. Sixty-five percent identify as an essential
threat by 2022 [1]. Fintech organizations focus on software, algorithms, and technology
to propose services similar to banking and other financial services. Very often costs of
their services lower cost than traditional financial institutions.
Online lending one of the crucial directions of developing fintech. Such companies
have been largely successful last 10 years. They actively implement new technologies
in credit risk-management and account for 15-20% of volume banking loans.
Customers of online crediting tend to the younger generation (fig. 1).
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons
License Attribution 4.0 International (CC BY 4.0).
126
Fig. 1. Loans age structure
Online lenders use different technologies for loan granting processes: verifications,
advanced scoring, and others. Most of them specialize in short-term loans (pay-day-
loan segment). Credit portfolio becomes large. It is logical to apply customer
profitability analysis (CPA) for optimizing strategies of loan granting. Our practical
analysis of applying CPA indicated some economic effects. The first effect concerns
the necessity to consider CPA jointly with risk management. The second effect
concerns the specificity of profit generation at considered credit portfolios. The
specificity is a (large) overpayment of some categories of borrowers. The third effect
shows itself through the positive correlation between risk and profits.
The paper devotes to presentation customer profitability analysis for nonbanking
lenders. Detailed consideration of joined together CPA and risk-management provided
economic logic for creation optimal strategy.
2. Materials and methods
2.1. Literature review
The problem of customer profitability analysis has been studied by many scientists. In
particular, Pobrić [2] investigates methods of measuring customer profitability under
different views. It considers customer profitability as customer group profitability.
Storbacka [3] focused on existing client segmentation as a valuable marketing
approach. Anandanatarajan [4], considers CRM as the process of acquiring, satisfying,
retaining, and growing profitable customers.
A lot of papers are devoted to lending in the aspect of financial market development.
Patalano & Roulet substantiate the increase in the level of public and corporate debt
and the scale of global credit markets, which is caused by non-bank financial
institutions [5]. Chernenko et al. analyzed a random sample of the credit market during
127
2010-2015 and showed that 32% of all loans were provided by nonbanking institutions
[6]. Credit cyclicality for banks and non-banking institutions is studied by Fleckenstein
et al. [7]. Kondova & Bandyopadhyay discussed a nonbank lending impact on bank
efficiency [8]. The asset pricing model with both bank and non-bank financial
institutions was simulated by d'Avernas et al. [8]. Distinctions of dealing with
information scarcity between the bank and nonbank financial institutions were analyzed
by Han [10]. Eichholtz et al. investigated the influence of local information on pricing
for banks and non-banks [11].
Experience of nonbanking lending from different countries presents by the next
authors: Bédard-Pagé [12] – Canada, Lee [13] – Korea, Rateiwa & Aziakpono [14] –
Egypt, Nigeria, and South Africa, Vasileva [15] – Bulgaria, Soukal et al. [16] – Czech
Republic, Sanfilippo-Azofra et al. [17] – Asia and Latin America.
Various theoretical aspects of the use of machine learning in the financial field are
studied in the scientific literature. Mathur [18] presented the overview of machine
learning in finance. Different aspects of modeling in finance discussed by Damodaran
et al. [19], Guryanova et al. [20], Derbentsev et al. [21], Kuzmenko et al. [22], Kiv et
al. [23], Sova & Lukianenko [24].
Machine learning approach for credit market modeling and its risk assessment
presented in Manasov & Ivanovska [25], Pokorná & Sponer [26], Liberman et al. [27],
Babenko et al. [28], Agarwal et al. [29], Venegas [30], Nyangena [31], Papouskova &
Hajek [32].
Machine learning algorithms are effectively used to work with the customer base; in
particular, clustering methods that use different types of data and make it possible to
divide customers into groups/segments and develop an individual proposal for each
group. Machine learning application for customer segmentation used in Monil et al.
[33], Costea & Bleotu [34], Cuadros-Solas & Rodríguez-Fernández [35].
2.2. Research methodology
The study used two methodologies as a basis. The first in our study is the customer
profitability analysis methodology. The methodological approaches of this analysis are
widely used in the corporate segment. It is especially effective in cases where the
corporation has many products in various segments. At the same time, the main
approach of the analysis is focused on considering the income from sales/services
provided in the context of the product line and/or segments. In assessing revenues, such
parameters as marketing costs, revenues in absolute terms, expenses for subsequent
servicing of products sold, the frequency of repeated sales to customers, and others are
usually considered. The result of this approach is clustering by the profitability of
products and segments. One submission of such clustering is ABC / XYZ structuration
by Noche [36]. More detailed of this that ABC focuses on generating income by clusters
and XYZ focuses on stability. ABC analysis is based on the classic Pareto rule (20% -
80%) and deals with the share of income in the general ledger. XYZ can be assessed
on the basis of risk measures related to the variability of income stream (Table 1).
The key methodological principle of analysis is to create matrix 3X3. The analysis
involves assessing profit in each cell of the matrix and leads to elaboration customer
relations management for each cell (it can be range from disposal strategy to active
growth).
128
Table 1. ABC/XYZ methodologies comparison
Focus on generating income Focus on variation income`s stream
A – cluster, which generate 80% value (or X – stable income stream
close to it)
B – cluster which provide 15% value Y – income stream with middle variation
including seasonal fluctuations
C – cluster with only 5% value Z – completely unstable income
One of the shortcomings of the abovementioned approach related to operations with
“blocks” such as products, segments, or others. It may be lost differences in generating
income by concrete clients at each cell of the matrix. This particular feature we have
considered in our research because, as we illustrate below, borrowers essentially
different by profitability. To implement this approach, all clients can be ordered in order
of increasing profitability. Moving "left-to-right" along this ordering, you can calculate
in cumulative terms the percentage of income that customers bring in relation to total
income. The result will be the Whale curve. Typical examples of such curves are given
in fig.2.
Fig. 2. Whale curves as illustration [Storbacka, 1998]
The second methodology that was used by us is the methodology for constructing
risk management used in consumer lending [37], [38]. In a generalized form, it is shown
in fig.3.
129
Assessments by
Risk of inflow of Verification Collection
scoring and credit Monitoring
borrowers procedures activity
reports
Fig. 3. Constructing risk management methodology
The methodology includes the following blocks. The first block is the risk analysis
of the incoming flow. It can be assessed based on the scoring of the credit bureaus.
Then it is possible to get a comparison of the risk level of the incoming flow to the
lender and the market as a whole. Also, you can compare the incoming flows by the
channels of attracting borrowers and products proposed by lenders. The second block
includes checking the borrower against various databases (for example, the database of
lost and stolen passports). The third block includes the most dynamically developing
borrower appraisal system. The most common model combines credit bureau scoring
with application scoring. The combination can be both in matrix form and in the form
of a single scoring, which has both application and behavioral characteristics.
Monitoring is an important component, which in the segment of short-term loans shows
the likelihood of prolongation or obtaining a second loan. The final in the given scheme
is the collection activity.
The combined use of customer profitability analysis methodological approaches and
credit risk management allowed us to obtain several results presented below.
3. Results and discussion
Our study was based on data on loans issued by several nonbanker lenders in segment
Pay-day-loans as on-line as off-line. As part of the initial analysis, we examined a
period of one year and estimated the income that clients brought on the borrowing.
Credit relations in this segment have specific features in comparison with the banking
segment. One of these features is the frequency of cases when the borrower overpays
on the loan. This happens for several reasons. In this segment, there is a large
percentage of borrowers who have a high risk and quite often experience problems with
loan payments. As a result, they use loan rollover, during which they pay interest. Also,
in case of delay, they must pay fines and commissions. Thus, in this segment, some
borrowers overpay the loan amount several times. At the same time, there is a fairly
large number of borrowers who do not pay on loans. A typical dependence of
profitability is shown in fig. 4.
The main difference from the banking segment is the increase in the graph on the
left. In the banking segment, this curve on the left is flat.
Considering from the point of view customer profitability analysis it is logical to
combine profitability analysis with risk analysis. Because here we have high risk and
correspondingly high return.
130
Fig. 4. Typical profitability of nonbanking borrowers
First, our approach includes the expansion of clusters, based on the Whale curve.
Analysis of payments specificity we grounded to divide borrowers into 4 clusters: A,
B, C, D. Cluster A corresponds to the most profitable customers which provide 100%.
Our research confirms the classical Pareto principle 20%:80%. Approximately 20% of
borrowers generate 100% of the profit. But of course, in reality it may be quite small
percentage of borrowers which generate 100%, as example, 15%, 10% or sometimes
3%. Cluster B involves borrowers who tend to pay “fair and square”. There is not so
much overpayment. Borrowers from these two clusters are used as “Good” in the credit
risk modeling (especially in credit scoring construction).
Borrowers from clusters C and D are typically considered as “Bad” in credit risk
management. Our approach includes separation borrowers which pay something (they
are in cluster C) and borrowers who “had not paid a penny” (they are is cluster D). The
logic of such clustering is effective for two reasons. The first reason is that borrowers
from cluster C “better” because they seek to pay. Borrowers from cluster D didn’t have
any plans to pay. The second reason lies in the strategy of working with borrowers from
these clusters. Borrowers which have characteristics D should be carefully rejected at
the application stage. The strategy of improving collection procedures should be
applied to borrowers from C. Our researches demonstrate the level of recovery of 40%-
60% for the borrower from cluster C.
The illustration of our clustering approach for analyse credit portfolio of non-bank
lenders is presented in fig. 5.
131
Fig. 5. Clustering credit portfolio for non-bank lender
The basic result of our research lies in linking risk assessment and profitability
measurement of nonbanking borrowers (major payday loan borrowers). This result is
presented in Table 2. Initially, we have structured borrowers with the help of credit
scoring. It was application scoring of lenders which involved as characteristic
parameters from borrower`s credit histories (collected by credit histories bureau).
Scoring estimates borrower from 0 (high risk) and low risk (1000). There were
segmentation borrowers into risk classes presented in Table 2 (with step 100 scores,
vertical columns). The cut-off of applied scoring was 400 scores, so there was
consideration of borrowers from segment. The bad rate curve is the indispensable part
of any scoring. This curve indicates % of “Bads” (C+D) at the borrowers of the
corresponding risk class.
Table 2. Correspondence between risk classes and clusters A, B, C, and D
A 18,7% 22,0% 20,9% 17,4% 15,7% 11,3%
B 58,1% 65,2% 71,7% 74,8% 75,9% 88,7%
19,7% 10,9% 7,7% 5,7% 5,0% 1,5%
Cut-off zone
C
D 3,5% 1,9% 3,3% 2,2% 1,8% 1,5%
Net Profit
per
250 602 647 553 466 429
Borrower,
UAH
Bad Rate 23,2% 12,9% 11,0% 7,8% 6,8% 3,0%
Risk classes ≤400 (400-500] (500-600] (600-700] (700-800] (800-900] (900-1000]
After that, we estimated the percentage ratio of borrowers from clusters A, B, C, and
D into scoring classes. Results are provided in horizontal rows. What are the main
results? Borrowers from cluster A constitute a higher percentage ratio at the riskier
classes! The percentage ratio of A to the classes of good borrowers is lower.
Borrowers from cluster B demonstrate a monotonic increasing of percentage ratio
from high-risk classes to low-risk classes.
132
Borrowers from cluster D demonstrate a monotonic decreasing of percentage ratio
from high-risk classes to low-risk classes. This is natural, but changes are not so much
as for C. Percentage ratio of C essentially decreases in this direction.
The main economic effect that was identified in our research is the following.
Increasing risk of the borrower interconnects with higher percentage ratio of borrowers
from cluster A (which are overpayment). This led to the financial objective: it is
important to find optimal correspondence between risk and payments from borrower
who essentially overpaid. In other words, it logically finds maximum output from
borrowers from different scoring classes.
Analysis of indicated financial objectives necessitates estimation of net profit from
one borrower (average) from different risk classes. The results can see in fig. 6.
700 647 25,0%
23,2% 602
600 553
20,0%
Net Profit per Borrower, UAH
500 466
429
Bad Rate, %
15,0%
400
12,9%
300 250 11,0%
10,0%
200 7,8%
6,8%
5,0%
100
3,0%
0 0,0%
(400-500] (500-600] (600-700] (700-800] (800-900] (900-1000]
Scoring classes
Fig. 6. Net profit per one borrower via credit scoring
The basic conclusion of credit portfolio profit analysis grounded by A, B, C, D
clustering leads to the following specificity. Borrowers with high scoring generate not
so many net profits. Borrowers with more risky scores generate maximum net profits.
Risk classes with low scores (but on the right side from cut-off) demonstrate decreasing
in net profits. Because they already include many bad borrowers and overpayments do
not cover losses from them. The basic conclusion lies in the fact that most profitable
borrowers are involved in more risky classes where overpayments cover losses in
maximum.
Another clustering approach that was applied in our research grounded on fuzzy
clustering. Fuzzy clustering characterizes by the property that data points can be
involved in different clusters. The abovementioned analysis demonstrates that
borrowers from the average risk scoring class can generate losses and overpayments.
This is one of the basic reasons to choose fuzzy clustering instead of classical “hard
clustering”.
133
It was choosing three indicators for run fuzzy clustering procedures. First is the score
of borrowers which indicates risk level. The second indicator corresponds to the net
profit level. The third indicator is loan amounts.
The applying of package “ppclust” from R demonstrated the following clustering
(fig. 7).
Fig. 7. Results of fuzzy clustering applications
The cluster average characteristics are presented in Table 3.
Table 3. Clustering results
Size of clusters Scoring values Loan amount (UAH) Net profit (UAH)
Cluster 1 52,08% 610,33 1564,03 270,91
Cluster 2 34,69% 602,90 3428,81 500,69
Cluster 3 13,23% 603,31 5914,17 1813,39
Clusters were also analyzed by cross-sections, which provides a more deeply looking
inside (fig.8).
134
Fig. 8. Clustering results by using pairs of the features
One of the business strategies to develop lending in the considered sphere may be
realized by combining risk estimation and net profit-generating by the borrower.
Typical high net profit generates borrowers with recurrence loan receive. So, the focus
may be concentrated to develop CRM with such a category.
4. Conclusions
Our research identifies the frameworks for customer profitability analysis for borrowers
of nonbanking financial institutions. The nature of such borrowers includes high credit
risk and at the same time high profit from effect “overpayment”. Overpayment has a
positive correlation with risk (negative correlation with score value). This provides a
new approach to the CPM. It is logical to focus marketing efforts on borrowers from
cluster A in parallel with risk assessment. CPM with borrowers from cluster B leads to
strategy increase profit. “Bad” borrowers we proposed to divide into clusters C and D.
Here should apply strong risk rules for cutting of D borrowers at the application stage.
The prospects of development CPM for this framework we find in the constructing
system of estimation borrowers at the application stage. The estimation supposes to
assess the probability that the borrower belongs to cluster A, B, C, or D.
Fuzzy clustering which was applied demonstrates differences of clusters.
Differences focus on loan amounts and net profits.
One of the crucial applications proposed in article clustering can be used for forming
business strategies to lend clients of non-bank lenders. This direction is part of our plans
for researches.
135
References
1. In the Game: Traditional Financial Institutions Embrace Fintech Disruption. Harvard
Business Review Analytic Service, 2019. Retrieved from
https://hbr.org/resources/pdfs/comm/mastercard/Fintech.pdf.
2. Pobrić, A. (2014). Measuring customer profitability: The applicability of different concepts
in practice. Ekonomika preduzeća, 62(3-4), 187-200.
3. Storbacka, K. (1997). Segmentation based on customer profitability—retrospective analysis
of retail bank customer bases. Journal of Marketing Management, 13(5), 479-492.
4. Anandanatarajan, D. K. (2019). Customer Reationship Management–A Strategic Tool for
Marketing. IJRAR Volume 6, Issue 2.
5. Patalano, R., & Roulet, C. (2020). Structural developments in global financial intermediation:
The rise of debt and non-bank credit intermediation.
6. Chernenko, S., Erel, I., & Prilmeier, R. (2019). Nonbank lending. National Bureau of
Economic Research.
7. Fleckenstein, Q., Gopal, M., Gutierrez Gallardo, G., & Hillenbrand, S. (2020). Nonbank
Lending and Credit Cyclicality. NYU Stern School of Business.
8. Kondova, G., & Bandyopadhyay, T. (2019). The Impact of Non-bank Lending on Bank
Efficiency: Data Envelopment Analysis of European Banks. International Journal of Trade,
Economics and Finance, 10(5), 108-112.
9. d'Avernas, A., Vandeweyer, Q., & Darracq-Pariès, M. (2020). The growth of non-bank
finance and new monetary policy tools. Research Bulletin, 69.
10. Han, J. H. (2017). Does Lending by banks and non-banks differ? Evidence from small
business financing. Banks & bank systems, (12, № 4), 98-104.
11. Eichholtz, P., Mimiroglu, N., Ongena, S., & Yönder, E. (2020). Banks, Non-Banks, and the
Incorporation of Local Information in CMBS Loan Pricing. Swiss Finance Institute Research
Paper, (19-58).
12. Bédard-Pagé, G. (2019). Non-bank financial intermediation in Canada: An update (No. 2019-
2). Bank of Canada Staff Discussion Paper.
13. Lee, M. (2018). Non-Bank Lending to Firms: Evidence from Korean Firm-Level Data. The
Journal of Industrial Distribution & Business, 9(9), 15-23.
14. Rateiwa, R., & Aziakpono, M. J. (2017). Non-bank financial institutions and economic
growth: Evidence from Africa's three largest economies. South African Journal of Economic
and Management Sciences, 20(1), 1-11.
15. Vasileva, V. (2019). Development Of Consumer Lending By Non-Bank Credit Companies
In Bulgaria. Народностопански архив, (1), 65-76.
16. Soukal, I., Hamplová, E., & Haviger, J. (2021). Effectiveness of Regulation of Educational
Requirements for Non-Bank Credit Providers in Czech Republic. Social Sciences, 10(1), 28.
17. Sanfilippo-Azofra, S., Torre-Olmo, B., & Cantero-Saiz, M. (2019). Microfinance institutions
and the bank lending channel in Asia and Latin America. Journal of Asian Economics, 63,
19-32.
18. Mathur, P. (2019). Overview of machine learning in finance. In Machine Learning
Applications Using Python (pp. 259-270). Apress, Berkeley, CA.
19. Damodaran, S., Kavin, S., Keerthi, K. U., Madhumathi, J., & Mythili, P. V. (2019,
November). Empowering MSMEs Through Digital Lending. In 2019 International
Conference on Digitization (ICD) (pp. 249-253). IEEE.
20. Guryanova, L., Yatsenko, R., Dubrovina, N., Babenko, V. (2020). Machine learning methods
and models, predictive analytics and applications. CEUR Workshop Proceedings, 2649, pp.
1–5.
136
21. Derbentsev, V., Matviychuk, A., Datsenko, N., Bezkorovainyi, V., Azaryan, A. (2020).
Machine learning approaches for financial time series forecasting. CEUR Workshop
Proceedings, 2713, pp. 434–450.
22. Kuzmenko, O., Šuleř, P., Lyeonov, S., Judrupa, I., Boiko, A. (2020) Data mining and
bifurcation analysis of the risk of money laundering with the involvement of financial
institutions. Journal of International Studies, 13(3), pp. 332–339.
23. Kiv, A., Hryhoruk, P., Khvostina, I., Solovieva, V., Soloviev, V., & Semerikov, S. (2020).
Machine learning of emerging markets in pandemic times. CEUR Workshop Proceedings,
2713, pp. 1–20.
24. Sova, Y., & Lukianenko, I. (2020, September). Theoretical and Empirical Analysis of the
Relationship Between Monetary Policy and Stock Market Indices. In 2020 10th International
Conference on Advanced Computer Information Technologies (ACIT) (pp. 708-711). IEEE.
25. Manasov, J., & Ivanovska, L. P. (2018). User preferences for banking services offered by
non-banking companies and tech giants. Journal of sustainable development, 8(20), 35-50.
26. Pokorná, M., & Sponer, M. (2016). Social lending and its risks. Procedia-Social and
Behavioral Sciences, 220, 330-337.
27. Liberman, A., Neilson, C., Opazo, L., & Zimmerman, S. (2018). The equilibrium effects of
information deletion: Evidence from consumer credit markets (No. w25097). National
Bureau of Economic Research.
28. Babenko, V., Panchyshyn, A., Zomchak, L., Nehrey, M., Artym-Drohomyretska, Z.,
Lahotskyi, T. (2021). Classical Machine Learning Methods in Economics Research: Macro
and Micro Level Example. WSEAS Transactions on Business and Economics, Vol. 18, 2021,
Art. #22, pp. 209-217. https://doi.org/10.37394/23207.2021.18.22.
29. Agarwal, S., Alok, S., Ghosh, P., & Gupta, S. (2020). Financial Inclusion and Alternate
Credit Scoring for the Millennials: Role of Big Data and Machine Learning in Fintech.
Working Paper, National University of Singapore.
30. Venegas, P. (2018). Risk scoring for non-bank financial institutions. Available at SSRN
3280738.
31. Nyangena, B. O. (2019). Consumer credit risk modelling using machine learning algorithms:
a comparative approach (Doctoral dissertation, Strathmore University).
32. Papouskova, M., & Hajek, P. (2019). Two-stage consumer credit risk modelling using
heterogeneous ensemble learning. Decision support systems, 118, 33-45.
33. Monil, P., Darshan, P., Jecky, R., Vimarsh, C., & Bhatt, B. R. (2020). Customer
Segmentation using machine learning International. Journal for Research in Applied Science
& Engineering Technology. Volume 8 Issue VI, 2104-2108.
34. Costea, A., & Bleotu, V. (2012). A new fuzzy clustering algorithm for evaluating the
performance of non-banking financial institutions in Romania. Economic Computation and
Economic Cybernetics Studies and Research, 46(4), 179-199.
35. Cuadros-Solas, J., & Rodríguez-Fernández, F. (2019). A Machine Learning Approach to the
Digitalization of Bank Customers: Evidence from Random and Causal Forests.
36. Noche, B. (2014). ABC-/XYZ Analysis Introduction. Universitat Duisburg Essen. Duisburg:
Bernd Noche. Retrieved from https://www. unidue.
de/imperia/md/content/tul/download/en_ss2015_lm01_le_abc_analysis. pdf. 2014.
37. Kaminskyi, A., Pysanets, K. (2017). Audit of risk management system in consumer lending.
Journal Association 1901 "SEPIKE". 18 Edition, 133-140.
38. Kaminskyi, A., Motoryn, R., Pysanets, K. (2017). The effectiveness of the use statistical data
of credit histories bureaus in risk management systems. Probability in action Vol. 3, 139-156.