Title: Application of AI in Credit Risk Scoring for Small Business Loans: A

case study on how AI-based random forest model improves a Delphi

model outcome in the case of Azerbaijani SMEs.

Author: Nigar Karimova

1
Table of Contents
Abstract .......................................................................................................................... 3
Introduction .................................................................................................................... 4
Literature Review ............................................................................................................ 5
Previous Research on Credit Risk Scoring. A summary of existing models .................................. 5
AI in Finance ............................................................................................................................ 7
Ethical concerns in AI ............................................................................................................... 7
Methodology ................................................................................................................... 8
AI model used .......................................................................................................................... 8
Results ...........................................................................................................................11
Discussion.......................................................................................................................14
Conclusion ......................................................................................................................16
Ethical considerations .....................................................................................................17
Declarations of Interest ..................................................................................................19
References......................................................................................................................20
Appendix 1. Delphi model Python code (a logistic regression) ..........................................22
Appendix 2. Random forest Python code .........................................................................22

2
Abstract
The research investigates how the application of a machine-learning random forest model
improves the accuracy and precision of a Delphi model. The context of the research is
Azerbaijani SMEs and the data for the study has been obtained from a financial institution
which had gathered it from the enterprises (as there is no public data on local SMEs, it was
not practical to verify the data independently).

The research used accuracy, precision, recall and F-1 scores for both models to compare them
and run the algorithms in Python. The findings showed that accuracy, precision, recall and F-
1 all improve considerably (from 0.69 to 0.83, from 0.65 to 0.81, from 0.56 to 0.77 and from
0.58 to 0.79, respectively).

The implications are that by applying AI models in credit risk modeling, financial institutions
can improve the accuracy of identifying potential defaulters which would reduce their credit
risk. In addition, an unfair rejection of credit access for SMEs would also go down having a
significant contribution to an economic growth in the economy.

Finally, such ethical issues as transparency of algorithms and biases in historical data should
be taken on board while making decisions based on AI algorithms in order to reduce
mechanical dependence on algorithms that cannot be justified in practice.

Keywords: Credit Risk Scoring, Artificial Intelligence, Random Forest Model, Machine Learning
in Finance, Delphi Model

3
Introduction
The application of AI in a broad number of fields has gained momentum in recent years which
has bolstered efficiency and effectiveness of organizational outcomes (Cao, 2022). Financial
system is not an exception which has benefited from multiple contributions of AI and ML.
Credit risk modeling is one of the areas which is prone to human errors (Cao, 2020).
Additionally, manual work that goes into the calculation of default scores through assessment
of the circumstances of each and every entry in the database is a cumbersome and time-
consuming process (Li, et al., 2023). AI is, however, not only powerful for its completion of
long processes in a relatively short period of time but also can capture complex relationships
better compared to traditional models. Traditional models rely on the static nature of
variables in the financial world which is not the case (Li, et al., 2023). A dynamic nature of
variables must be incorporated into the analysis in order to make more informed decisions.
Learning from the historical data, machine learning algorithms can provide a much more
effective outcome in terms of accuracy (Farahani and Ghasemi, 2024). AI models do not only
investigate internal factors such as performance measures of companies but also analyze
macroeconomic and political (external factors) factors which can have a consequential impact
on the performance of firms. SMEs are exposed to a great risk as changes in macroeconomic
and political conditions in a particular country can be especially harsh on smaller firms (Li, et
al., 2023). These benefits of the application of AI in financial analysis increased its use and the
applications in credit scoring are prevalent as well. Nevertheless, the caveat is that there
should be a large dataset relevant to a particular context in order for AI models to analyze
and learn from for an efficient model.

The role of credit scoring and the integration of AI into the process for the successful
operations of SMEs in emerging market economies are crucial (Popovych, 2022). Firstly,
traditional scoring models rely on historical records and do not capture a wide variety of
factors that can in combination contribute to the success or failure of SMEs. Therefore, more
comprehensive approaches including the integration of machine learning into the process
enable financial institutions to more accurately evaluate potential risks of SMEs and take into
consideration variables which fluctuate with season and under the influence of general
macroeconomic conditions (Popovych, 2022). Therefore, to appease unfair treatment of

4
SMEs by financial institutions due to a lack of long historical data which often leads to either
a rejection of loan application or high interest rates, AI in credit scoring holds a significant
potential.

By evaluating possible AI models that can be applied in credit risk modeling, the purpose of
this research is to introduce a AI-based model which improves the accuracy and precision of
traditional models used in credit assessment. Random forests model is the selected algorithm
for this purpose which adapts to real time data and incorporates non-linear complex
relationships among variables to a significant extent (Zhang, et al., 2018). By using historical
data to build simulations and learn from thousands of iterations, AI models also predict the
variations in external variables and their impact on the performance of SMEs instead of
focusing on limited historical data of the performance of SMEs themselves. The research aims
to see a practical improvement in the accuracy and prediction of the model used traditionally
for credit scoring after the application of random forest model. By providing the management
of SMEs and financial institutions with valuable insights, the model will be crucial in terms of
determining whether practical improvements can be obtained from the application of AI
which leads to a greater accuracy in credit scoring. Moreover, the identification of relative
weight of each of the risk-drivers will also allow SMEs to understand the most important risk
categories they encounter.

Thesis statement: This study analyzes the effectiveness of AI algorithms in improving credit
risk scoring models for SMEs by taking into consideration both advantages and ethical
concerns. The research looks, in particular, into the case of Azerbaijani SMEs which have
applied for bank loans in the last 5 years.

Literature Review
Previous Research on Credit Risk Scoring. A summary of existing models

The most fundamental of the available historical models for credit risk scoring is a regression
analysis (Bolton, 2009). A particular type of regression called logistic regression is applied to
predict the default of firms. Regression techniques involve the analysis of the impact of

5
selected independent variables (characteristics of the loan applicant such as debt to equity,
previous record of payments etc) and the dependent variable (default status) (Bensic, et al.,
2005). Based on the historical data on the independent and dependent variables, a
relationship between the two is identified allowing to predict defaults. Some limitations of
these models should be highlighted. Firstly, logistic regression depends on linearity
assumptions (Khemeais, 2016). The model assumes that there is a linear relationship between
independent and dependent variables and this assumption is not always valid. Therefore, any
deviations from linearity reduces the accuracy of logistic regression outcome (Bensic, et al.,
2005). Furthermore, regression analysis uses historical data only which is again one of the
limitations of this type of analysis for credit scoring as changes in the external environment
cannot be captured by the analysis tool.

Next, discriminant analysis should be highlighted as a frequently used model for credit score
modeling (Mvula, 2011). Discriminant analysis analyzes key characteristics of the companies
studied and groups them into those likely to default and those which are in a better financial
condition. The model has been in use for a long period of time but its assumptions are also
restrictive as it relies on normal distribution of explanatory variables (Mylonakis, 2010). The
model is also prone to have a distorted outcome under the condition of high volatility. In
these cases, discriminant analysis performs poorly illustrating that it has a high accuracy
under relatively stable conditions.

Finally, Altman’s Z score model is a predictive model for default probability. The model
incorporates financial ratios and using company fundamental information, it is possible to
predict the likelihood of default (Panigrahi, 2019). A wide range of financial ratios go into the
calculation of Z score and depending on the outcome score, the likelihood of default can be
predicted. Some drawbacks of the model also exist as it has been mostly applied to large
corporations and there can be volatility in emerging markets which cannot be reflected well
by this model (Altman, et al., 2017).

In sum, different models have been applied thus far for credit score modeling but each has
certain shortcomings.

6
AI in Finance
AI has revolutionized finance and has been applied in credit score modeling as well
extensively (Cao, 2022). AI in credit risk assessment has been used on multiple fronts. Firstly,
the capability to analyze a vast amount of data in a short period of time has been one of the
most attractive features of the application of AI. Additionally, the complexity of analyzed data
is particularly impressive with AI as it can effectively divulge relationships between complex
variables, capture non-linearity and provide an accurate outcome for predictions (Hilpisch,
2020). AI models also reduces noise in the data indicating that these models select the most
important variables in the dataset.

AI also includes data which is unrelated to the direct activity of a borrower despite having a
significant impact on the borrower default (Buchanan, 2019). These variables are external
factors such as overall economic conditions in the economy, trends in the market, consumer
behavior and even political risks. By capturing real-time data and disclosing its impact on risk
factors have increased the prominence of AI in financial markets (Cao and Zhai, 2022). Deep
learning models and neural networks are examples of these algorithms (machine learning)
which have increased the effectiveness of credit scoring by financial institutions.

Ethical concerns in AI

Depending on the algorithms and provided input data, ethical issues might be present in the
use of AI for credit risk modeling. For example, existing discrimination in lending to a
particular group of individuals can be reflected in AI solutions if these biases are a part of
historical data used in training (Svetlova, 2022).

Regarding transparency, there is a literature on concerns about the algorithms used in deep
learning models (Max, et al., 2021). These models are referred to as a black box and have
raised concerns among consumers and other users who are not certain of the exact
procedures the model uses and whether there are biases. The use of more explainable
techniques such as LIME and SHAP has been advocated in order to ensure that AI models are
more explainable (Bhattacharya, 2022).

7
Data use has serious implications. Therefore, relevant regulations have developed in this area.
However, technology usually precedes regulations and with regards to the use of AI, the
adherence to regulatory requirements has been rather ambiguous as the biggest developers
of AI algorithms have operated mostly with a lack of transparency (Dumouchel, 2023). GDPR
requirements have stated that usage of personal data must meet certain standards such as
being based on unequivocal and informed consent. Nevertheless, there are certain concerns
regarding whether all information used for AI data training and models is based on informed
consent (Schuett, 2023). The issues of content and transparency have been present in regards
to the use of AI and it can be argued that they are not resolved yet as AI is still in significant
progress and similar issues are expected to emerge with time.

To summarize the literature review, models in credit scoring preceding AI use have been
regression analysis, discriminant analysis and Altman’s Z score model. These models have had
serious drawbacks which mostly boil down to normality of data and linearity of relationship
between variables. These features reduced the effectiveness of the mentioned models and
increased preference for AI in the credit risk assessment process. Secondly, AI has found its
way into credit risk assessment due to its ability to process a vast amount of data and
capabilities regarding the analysis of non-linear relationships between variables. Credit
scoring outcomes (such as precision and accuracy) have improved as a result of the use of AI
in finance. Finally, ethicality remains an issue in AI credit score modeling because of the
potential biases existing in the database used in the analysis and the way algorithms are built
as there is a low level of transparency in the process.

Methodology

AI model used

The model used in the research is random forest model is a machine learning
model/algorithm which allows researchers to enhance the accuracy of predictions which is
suitable in this research as well (Van, et al., 2016). The power of the model lies with the fact

8
that it incorporates multiple decision-trees and captures non-linear relationships between
variables which is one of the shortcomings of traditional models. Considering that there are
multiple dynamic risk factors impacting SMEs in Azerbaijan and in general, the model provides
useful insights into the complex nature of relationships between variables.

Several steps are taken in the model to arrive at a final outcome (prediction). Firstly, the
model takes sample data from the original data and creates training datasets. Randomness
of the selection of data points from the original data and replacement contributes to the
robustness of the model. This step is bootstrap sampling.

Next, for each training dataset, a decision-tree is built. Decision trees incorporate a number
of key variables such as revenue growth, leverage and so on. Based on the trained data
relevant to each dataset, a decision is made by the model which is the most suitable (Van, et
al,. 2016).

Decisions for each decision tree is made individually but aggregated for a final result. The final
result reflects the probability of default and tells whether the company will default or not.

The value of random forest models is high in the case of SME default projections because
these models are capable of handling non-linear data and relationships. Oftentimes,
relationships between variables cannot be analyzed accurately assuming linear relationships.
Therefore, random forest models are a powerful tool in this regard.

Additionally, allocation of scores for the importance of each variable (in terms of which
factors drive the company’s default risk) plays a role in the analysis of SMEs. SME managers
and creditors to these organizations can obtain insights from the model regarding which
variables have contributed to the risk of default most.

Data sources

Data source for the research was a local bank in Azerbaijan which in consultation with SMEs
allowed to access the data for SMEs used in the analysis. The same bank applied traditional

9
credit scoring (Delphi method) to assess the risk of default of the analyzed SMEs in this study.
5-year data was collected for the research as the data for longer than 5-year horizon did not
exist in the database (either because SMEs evaluated no longer had a relationship with the
bank for more than 5 years or some of them stopped their operations altogether due to a
default or being acquired by other firms).

The following historical data was analyzed as an input in the research:

-Revenue growth (historical revenue growth figures were taken into account). Revenue
growth figures would be a useful starting point for the model to assess the future of the
operations of SMEs. Furthermore, net profit margin of companies have also been selected as
a variable as revenue might be misleading as a before costs measure.

-Cash flow variance: Cash flow is the most important metric in regards to credit assessment
as cash flow health of the firm is the indicator of its ability to service its debt as well (Zhang,
et al., 2018). Therefore, cash flow variability should be included in the calculation of default
probability. Operating cash flow is the selected proxy in this research.

-Debt to equity ratio: This ratio illustrates the level of indebtedness (the proportion of debt
in capital structure vs equity) and can be informative in terms of evaluation of the risks of
companies. Excessive debt level can result in high risks as well preventing the firm from being
able to pay off its debt. Therefore, it is crucial that debt to equity ratio is chosen as one of the
variables in the model.

-Commodity price sensitivity: Commodity price sensitivity can be a crucial determinant of the
risks faced by companies in particular small firms. Commodity prices on the world markets
such as oil and gas, wheat and coffee can determine revenue level and profitability of SMEs.
Therefore, based on the companies in the data, the most relevant commodities have been
selected in order to measure the sensitivity of revenue of SMEs to the prices of these
commodities. Oil price is the most pertinent due to the fact that Azerbaijan is an oil exporting
country and its economic performance depends significantly on the price of oil. Therefore, a
drop in the price of oil in the world markets can reduce profitability of companies in the

10
country. Additionally, grain and cotton prices have been selected as a significant variable in
regards to potentially explaining the revenue variability of Azerbaijani SMEs. These variables
have been picked because SMEs in the sample are from agricultural sector also having an
exposure to oil and gas prices. Sensitivity is measured by 5-year correlation coefficient
between the prices of relevant commodities and revenue of SMEs.

-Market conditions: To measure the impact of market conditions, GDP growth rate has been
integrated into the analysis as an input.

Model evaluation
Accuracy and precision of the model are metrics used for the evaluation of the model
performance.

Ethical considerations
To ensure that there are no ethical issues in the model (biases), the data for the model has
been selected so that it captures all available SME data provided instead of selecting
companies from certain sectors or with certain characteristics (for example, poor performers
vs good performers).

Results

The results of the study have been presented in a comparative fashion. Firstly, the outcome
of a Delphi model has been presented and discussed. Next, a random forest model
implemented in the research as an improvement to the traditional method has been laid out.
The comparison between key metrics facilitated an understanding of an improvement
achieved as a result of the application of AI model.

It should be mentioned that variables (predictors) for each model are the same in the analysis.
Thus, as discussed in methodology section, independent variables of the model are revenue
growth, cash flow variance, leverage (debt to equity ratio), net profit margin, sensitivity to

11
commodity prices (oil, cotton and grain have been included in the form of correlation
coefficients between each of these commodity and revenue of SMEs) and default dummy (0
for no default and 1 for default that occurred).

Python code for the analysis can be found in Appendix 1.

Logistic regression has been applied (as one of the traditional methods of credit score
modeling) as the Delphi model used in practice relied on a regression analysis according to
the provider of the model (financial institution that applied this model to calculate credit
scores of SMEs). Considering this fact, a logistic regression has been run (See Appendix 1 for
code). Python codes have been run to calculate accuracy, precision, recall and F-1 scores.

To elaborate on the metrics, precision of the model refers to how often the prediction of the
true positive by the model is correct. If there are both true and false predictions, the precision
of the model is calculated by dividing the number of total correct positive predictions by the
number of total predictions made (Provenzano, et al., 2020).

Precision is the accuracy with regards to the target class. However, accuracy of the model
shows an overall accuracy of the model.

Furthermore, the ability of the model to find all of the objects in the target class is called a
recall in machine learning which is also compared between two models (Provenzano, et al.,
2020).

F-1 score, finally, is a balanced measure (harmonic mean) of precision and recall
demonstrating an overall robustness of the model.

Next, Python code for a machine learning model (random forest) has been presented in
Appendix 2 which is a step-by-step process for calculation of each metric (accuracy, precision,
recall and F-1.

12
Table 1. Key performance indicators
Performance metric Delphi model Random forest (AI model)
Accuracy 0.69 0.83
Precision 0.65 0.81
Recall 0.56 0.77
F-1 0.58 0.79
This table compares the performance metrics (accuracy, precision, recall, F-1 score) of the
Delphi model and the Random Forest model in predicting SME defaults.

To interpret the comparisons, Delphi model has an accuracy of 0.69 vs 0.83 of random forest
model. In this case, the application of machine learning algorithm has resulted in an
outperformance over a traditional model. Delphi model predicted the defaults with 69%
accuracy whereas random forest model has an 83% accuracy which is considerably high.
Accuracy refers to being able to identify defaulting SMEs.

Additionally, precision adds another perspective in terms of understanding the performance

of the two models. Precision of Delphi model is 65% whereas random forest model has
achieved a precision rate of 81%. This significant improvement demonstrates a higher level
of accuracy when machine learning is integrated into the analysis. The interpretation of the
result is that random forest model has been able to identify true positives with a greater
precision. Therefore, the model classified defaulting and non-defaulting SMEs to a higher
accuracy by reducing false positives.

Recall of the model looks at the analysis from another angle as it captures the ability of the
model to determine false negatives (those SMEs which had been identified as a low risk but
actually defaulted). Recall of the model for Delphi model is 56% whereas it is 77% for random
forest model. Therefore, random forest algorithm is much more effective in terms of its ability
to decrease the number of false negatives.

Finally, F-1 score is a balanced metric between recall and precision (false positive and false
negative reduction capabilities of models) and it is 58% for Delphi model as opposed 79% for
random forest model (Provenzano, et al., 2020). As this metric can be considered as a

13
measure of an overall performance of the model, again, random forest model is much more
effective in terms of predicting defaults of SMEs accurately.

The results of the model have been compared for two models which are a Delphi model and
a random forest model. Delphi model was traditionally used through the application of a
logistic regression by the financial institution providing the SME data for the research. A
random forest model is developed and applied in this study to improve the outcome of Delphi
model. As observed from the results, there has indeed been a significant improvement in the
outcome of the model (accuracy, precision, recall and F-1) as a result of the use of machine
learning.

Discussion
To discuss the findings from the perspective of SMEs in Azerbaijan and broadly, it should be
mentioned that Delphi and other historical models are known for a lower level of
effectiveness in terms of prediction accuracy and precision. This is due to the fact that
historical models rely on linear relationships between variables (Provenzano, et al., 2020).
This assumption is breached in reality as financial variables which are both specific to the firm
and external (such as macroeconomic and political factors) are often non-linear and complex
in their relationship (Panigrahi, 2019). This shortcoming is fixed by random forest models
which is reflected in the outcome of this study. As was found in the analysis of both models,
a logistic regression had a much lower level of accuracy, precision, recall and F-1 scores
compared to a random forest model. AI-based random forest model is more accurate as it
incorporates real-time data, better captures non-linear relationship among variables and
integrates changes in the relationship of variables. The dynamic nature of the model is much
more effective in regards to capturing the factors that can lead to a default of an SME.

In case of SMEs in general, a higher level of accuracy of the model in predictions indicates
that, credit risks of financial institutions are ameliorated considerably as a result of the
application of this model. Financial institutions can predict the accuracy of defaults of SMEs
and reflect this in their operations in order to avoid lending to companies which are
imminently on the brink of a default (Bensic, et al., 2005). Moreover, a more comprehensive
analysis and capturing of a larger number of variables along with their dynamism lead to a

14
better assessment outcome and SMEs are not unfairly treated as a result of a lack of robust
model. Initially, a Delphi model was used which usually attaches a higher risk to SMEs because
SMEs in Azerbaijan do not have comprehensive information going back for a considerable
period of time. Taking this into consideration, financial institutions usually reject loan
applications from SMEs or charge them a higher interest rate in order to cover for potential
risks. However, as a random forest model is more dynamic and can integrate future potential
of the enterprises better into the analysis, it is also more likely that these models with the
help of machine learning can provide a fairer assessment of the creditworthiness of SMEs.
This will lead to an allocation of funds to SMEs which merit it with their historical information.

If compared with an existing literature, it can be argued that the results are backed up by the
extant research in the area. The limitations of the historical methods have been illuminated
in the existing literature. For example, inability of logistic regression to analyze non-linear
relationship between variables has been elaborated on as a key limitation of this method and
shown as an impeding factor in terms of reducing the accuracy and precision of these forms
of models (Bolton, 2009). The results of this research also showed that predictive capabilities
of a logistic model are lower compared to a random forest model. From a different
perspective, literature also touched on the capabilities of AI models and highlighted that
these models analyze a vast amount of data regardless of the type of relationship between
variables and provide an accurate forecast outcome (Popovych, 2022). The results of the
research have also corroborated this line of argument in the literature.

With regards to limitations, there are certain shortcomings of the model which need to be
discussed. Firstly, the accuracy of the model is as good as its input. The inputs for key variables
of the model were provided by the financial institution (a local bank in Azerbaijan) in
agreement with SMEs in the data. As there is no publicly audited financial data on these SMEs,
it is possible that there are inaccuracies in the data as SMEs could have been interested in
manipulating the data for a better credit score outcome. Additionally, decision-making
rationale and the process by AI models cannot be easily explained (Svetlova, 2022). Machine
learning models are challenging to interpret and financial institutions could find it difficult
and sometimes unethical to explain to SMEs why their application for loan has been rejected.
Therefore, by simply referring to a higher accuracy of a particular model in justifying a loan

15
application decision can be difficult to substantiate. Furthermore, overfitting is another
concern that can reduce the feasibility of AI models as it is possible that models can attach
greater weights to historical data and short-term data variability in decision-making (Schuett,
2023). This, in turn, would impair the capability of financial institutions to assess a long-term
financial prospect of SMEs. Hence, any potential applicants of AI models in practice for credit
risk assessment should be aware of these limitations.

The findings also have broader implications for the use of AI in finance along with the context
of Azerbaijan. The results lent support to the use of AI in finance including credit risk
assessment by highlighting the fact that the accuracy and precision of the model rise
significantly when machine learning model instead of a traditional logistic regression is used.
Also, it is an evident that machine learning models need a considerable dataset in order to be
trained to achieve a reliable outcome. The practical implications from this fact are that relying
on a limited dataset for an accurate outcome with the use of machine learning can backfire.
Finally, ethical issues in the use of machine learning such as content for the inclusion of data
in training the model and transparency should be taken into account before interpreting and
applying the model outcomes in practice. However, the solution to some of these issues can
be beyond the control of banks such as transparency in machine learning algorithms as banks,
in particular in developing countries such as Azerbaijan, rely on ready algorithms instead of
determining the work principles of complex machine learning models. Nevertheless, banks
can take a greater control of issues by checking the input data for potential biases. This could
reduce pre-existing biases and ensure that model results are accurate.

Conclusion

To summarize, this study used SME data in Azerbaijan to compare two models for credit
scoring. The initial model is a Delphi model for credit scoring but has a number of limitations
which have been addressed by AI model. The selected AI model in the research is random
forest model which predicts the default probability (credit score) of SMEs using different
variables of importance for the performance of SMEs. The selected variables are then
simulated to build training data and create decision-trees. The decision-tree outcomes are

16
summarized to obtain the final score. The comparison of the predicted scores with actual
default figures in both models (Delphi and random forest) has revealed the accuracy and
precision of each model. This allowed to calculate the improvement that AI model has
achieved. Delphi model in the analysis was a logistic regression model which uses historical
data to predicts defaults. It is a linear model which can be accurate mostly under normal data
assumption and linearity of relationship between variables. Therefore, these limitations had
a negative impact on the model’s accuracy in comparison with a random forest model which
is more flexible and powerful.

It was revealed that random forest model increases the accuracy (overall ability to identify
defaulters) from 0.69 to 0.83. A significant improvement was also identified from the use of
a machine learning model in regards to the precision of the model (the ability to reduce false
positives). A change from 0.65 to 0.83 was recorded which highlights that financial institutions
can increase the precision of determining potential defaulters in time. False negatives have
also been reduced as the recall score rose from 0.56 to 0.77. Finally, F-1 which is a harmonic
weighted average between recall and precision is 0.79 in the case of a random forest model
which was 0.58 for Delphi model. Hence, a random forest model (a machine learning model)
has outperformed a traditional model on all categories.

The limitations of the model are that it relies on historical data provided by SMEs to a financial
institution can be inherently biased to present the financial conditions and performance of
firms favorably. Also, a random forest is a complex model based on machine learning and its
application to sift through potential defaulters might be challenging to explain by banks and
can carry potential ethical concerns as the work principles of machine learning algorithms are
not clear enough to substantiate decisions based on them.

Ethical considerations

Several ethical issues should be emphasized along with mitigating practices that must be
applied by responsible leaders.

17
Firstly, a bias in historical data and algorithms are well-known to be issues in the use of AI in
predictions (Svetlova, 2022). A bias in historical data can be the result of the discriminatory
practices that were used in lending, for example. To minimize a bias in this regard,
practitioners should evaluate the data for potential misrepresentations or discriminations
prior to its analysis. In the case of this research, the data is from SMEs to which a bank had
already allocated loans.

The research relied on historical financial data of firms and current macroeconomic data for
the analysis. However, one major issue regarding bias in this study is the fact that it is
impractical to check values provided for variables by SMEs as financial institutions should be
in more direct involvement with SMEs to verify the accuracy of the provided data.

Furthermore, accountability and transparency are major issues in justifying decisions made
by AI models (Schuett, 2023). If financial institutions are not capable of explaining the model
algorithms clearly, then making decisions on these models’ outcome can be unethical.
Therefore, models which are explainable (LIME and SHAP tools can be helpful in this regard)
should be used for decision-making instead of relying on ambiguous models to justify credit
decisions.

Fairness of models should be ensured as well by checking that models do not apply special
features creating an unequal inclusion for a certain group of people or companies (Zhang, et
al., 2018). Particular tools to check demographic parity can be incorporated into models, for
example. In the case of this study, as the research involves the analysis of SME data that have
been able to secure a loan from a financial institution, fairness issues are not applicable as
these SMEs are further evaluated to check if they are likely to default in the future.

In short, ethicality of the inputs and outcome of the models should be ensured by financial
institutions using them. This involves regularly checking the input data for biases, using
explainable tools for models and finally, taking the limitations of algorithms into account in
decision-making rather than mechanically deciding based on the models. In case of SMEs in
Azerbaijan, financial institutions that apply a random forest model in the future, should
ensure that data provided by SMEs accurately reflect the reality of companies. Additionally,

18
as discussed, tools that enable a clear explanation of models should also be integrated for an
ethical and justifiable decision-making.

Declarations of Interest
The author reports no conflicts of interest. The author alone is responsible for the content
and writing of the paper.

19
References

Altman, Edward I., Marta Iwanicz-Drozdowska, Erkki K. Laitinen, and Arto Suvas. "Financial Distress
Prediction in an International Context: A Review and Empirical Analysis of Altman’s Z‐Score Model."
Journal of International Financial Management & Accounting 28, no. 2 (2017): 131-171.

Bensic, M., N. Sarlija, and M. Zekic‐Susac. "Modelling Small‐Business Credit Scoring by Using Logistic
Regression, Neural Networks, and Decision Trees." Intelligent Systems in Accounting, Finance &
Management 13, no. 3 (2005): 133-150.

Bhattacharya, Anirban. Applied Machine Learning Explainability Techniques: Make ML Models

Explainable and Trustworthy for Practical Applications Using LIME, SHAP, and More. Birmingham:
Packt Publishing Ltd., 2022.

Bolton, Carl. "Logistic Regression and Its Application in Credit Scoring." PhD diss., University of
Pretoria, 2009.

Buchanan, Bruce G. Artificial Intelligence in Finance. Cambridge, MA: MIT Press, 2019.

Cao, Longbing. "AI in Finance: A Review." SSRN Scholarly Paper, 2020.

https://doi.org/10.2139/ssrn.3647625.

Cao, Longbing. AI in Finance: Challenges, Techniques, and Opportunities. New York: ACM Computing
Surveys, 2022.

Cao, Longbing, and J. Zhai. "A Survey of AI in Finance." Journal of Chinese Economic and Business
Studies 20, no. 2 (2022): 125-137.

Dumouchel, Paul. "AI and Regulations." AI 4, no. 4 (2023): 1023-1035.

Farahani, M. S., and Ghasemi, G. "How AI Changes the Game in Finance Business Models."
International Journal of Innovation in Management, Economics and Social Sciences 4, no. 1 (2024):
35-46.

Hilpisch, Yves. Artificial Intelligence in Finance. Sebastopol, CA: O'Reilly Media, 2020.

Khemais, Zied, Nesrine D., and Mohamed M. "Credit Scoring and Default Risk Prediction: A
Comparative Study between Discriminant Analysis & Logistic Regression." International Journal of
Economics and Finance 8, no. 4 (2016): 39.

Li, X., Alex Sigov, Leonid Ratkin, Leonid A. Ivanov, and Lina Li. "Artificial Intelligence Applications in
Finance: A Survey." Journal of Management Analytics 10, no. 4 (2023): 676-692.

Max, R., A. Kriebitz, and C. Von Websky. "Ethical Considerations about the Implications of Artificial
Intelligence in Finance." In Handbook on Ethics in Finance, 577-592. New York: Springer, 2021.

Mylonakis, John, and George Diacogiannis. "Evaluating the Likelihood of Using Linear Discriminant
Analysis as a Commercial Bank Card Owners Credit Scoring Model." International Business Research
3, no. 2 (2010): 9.

20
Mvula, Chijoriga M. "Application of Multiple Discriminant Analysis (MDA) as a Credit Scoring and Risk
Assessment Model." International Journal of Emerging Markets 6, no. 2 (2011): 132-147.
Panigrahi, C. M. A. "Validity of Altman’s ‘Z’ Score Model in Predicting Financial Distress of
Pharmaceutical Companies." NMIMS Journal of Economics and Public Policy 4, no. 1 (2019).

Popovych, Bohdan. Application of AI in Credit Scoring Modeling. Wiesbaden: Springer Gabler, 2022.

Provenzano, A. R., D. Trifiro, A. Datteo, L. Giada, N. Jean, A. Riciputi, G. L. Pera, M. Spadaccino, L.

Massaron, and C. Nordio. "Machine Learning Approach for Credit Scoring." arXiv preprint,
arXiv:2008.01687, 2020.

Schuett, Johann. "Defining the Scope of AI Regulations." Law Innovation and Technology 15, no. 1
(2023): 60-82.

Svetlova, Ekaterina. "AI Ethics and Systemic Risks in Finance." AI and Ethics 2, no. 4 (2022): 713-725.

Van Sang, H., N. H. Nam, and N. D. Nhan. "A Novel Credit Scoring Prediction Model Based on Feature
Selection Approach and Parallel Random Forest." Indian Journal of Science and Technology 9, no. 20
(2016): 1-6.

Zhang, X., Y. Yang, and Z. Zhou. "A Novel Credit Scoring Model Based on Optimized Random Forest."
In Proceedings of the IEEE 8th Annual Computing and Communication Workshop and Conference
(CCWC), 60-65, Las Vegas: IEEE, 2018.

21
Appendix 1. Delphi model Python code (a logistic regression)
# Logistic Regression (proxy for Delphi)
logistic_model = LogisticRegression()
logistic_model.fit(X_train, y_train)
y_pred_logistic = logistic_model.predict(X_test)

# Performance Metrics for Logistic Regression (Delphi)

accuracy_logistic = accuracy_score(y_test, y_pred_logistic)
precision_logistic = precision_score(y_test, y_pred_logistic)
recall_logistic = recall_score(y_test, y_pred_logistic)
f1_logistic = f1_score(y_test, y_pred_logistic)

print(f"Delphi Model (Logistic Regression) Performance:")

print(f"Accuracy: {accuracy_logistic:.2f}")
print(f"Precision: {precision_logistic:.2f}")
print(f"Recall: {recall_logistic:.2f}")
print(f"F1-Score: {f1_logistic:.2f}")

Appendix 2. Random forest Python code

import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score

# Simulate SME dataset

np.random.seed(42)
n_samples = 1000
data = pd.DataFrame({
'Revenue_Growth': np.random.uniform(-0.2, 0.2, n_samples), # Revenue growth
between -20% and +20%
'Cash_Flow_Variability': np.random.uniform(0.1, 0.5, n_samples), # Cash flow variability
10%-50%
'Debt_Equity_Ratio': np.random.uniform(0.2, 3, n_samples), # Debt-to-equity ratio
between 0.2 and 3
'Profit_Margin': np.random.uniform(0.05, 0.25, n_samples), # Profit margins 5%-25%
'Commodity_Price_Dependency': np.random.uniform(0.5, 1, n_samples), # Correlation
with commodity prices
'Industry_Sector': np.random.choice([0, 1], size=n_samples), # 0: agriculture, 1:
manufacturing

22
 'Default_Status': np.random.choice([0, 1], size=n_samples, p=[0.8, 0.2]) # 80% no default,
20% default
})

# Split into features and target

X = data.drop('Default_Status', axis=1)
y = data['Default_Status']

# Train-test split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=42)

# Random Forest Classifier

rf_model = RandomForestClassifier(n_estimators=100, random_state=42)
rf_model.fit(X_train, y_train)
y_pred_rf = rf_model.predict(X_test)

# Performance Metrics for Random Forest

accuracy_rf = accuracy_score(y_test, y_pred_rf)
precision_rf = precision_score(y_test, y_pred_rf)
recall_rf = recall_score(y_test, y_pred_rf)
f1_rf = f1_score(y_test, y_pred_rf)

print(f"\nRandom Forest Model Performance:")

print(f"Accuracy: {accuracy_rf:.2f}")
print(f"Precision: {precision_rf:.2f}")
print(f"Recall: {recall_rf:.2f}")
print(f"F1-Score: {f1_rf:.2f}")

23