Credit risk management is central to the stability and profitability of financial institutions. This study applies binary logistic regression to a real-world dataset of credit card clients to identify predictors of loan default. Using SPSS for statistical modeling, we evaluated the contribution of demographic and financial variables including credit limit, past bill amounts, and repayment history. The model achieved an overall classification accuracy of 81.2%, with strong predictive power for recent payment behavior.
The most important things were the ones that made payments late in the last three months (PAY_0, PAY_2, PAY_3). The model calibration isn't perfect, but the results do give useful information on how to find borrowers who are likely to default. Logistic regression is a useful tool for risk analysts because it is easy to understand and see through. This is especially true in regulated environments where it is important for models to be clear. These results support the use of data-driven credit scoring models in decision-making processes.
c_1
3 Materials and Methods
4. Results and Discussion
5. Conclusion
References
Abstract
Credit risk management is central to the stability and profitability of financial institutions. This study applies binary logistic regression to a real-world dataset of credit card clients to identify predictors of loan default. Using SPSS for statistical modeling, we evaluated the contribution of demographic and financial variables including credit limit, past bill amounts, and repayment history. The model achieved an overall classification accuracy of 81.2%, with strong predictive power for recent payment behavior.
The most important things were the ones that made payments late in the last three months (PAY_0, PAY_2, PAY_3). The model calibration isn't perfect, but the results do give useful information on how to find borrowers who are likely to default. Logistic regression is a useful tool for risk analysts because it is easy to understand and see through. This is especially true in regulated environments where it is important for models to be clear. These results support the use of data-driven credit scoring models in decision-making processes.
Keywords : Credit risk, Logistic regression, Default prediction, SPSS, Credit scoring, Banking analytics, Risk management
1. Introduction
Predicting loan defaults is more than just a technical necessity in banking. It is necessary to guarantee lending institutions' financial stability. When a borrower defaults on a loan, the consequences are not limited to that one loan. They have an impact on capital reserves, future risk tolerance, and the overall lending approach.
The likelihood that a borrower will default has long been determined by credit scoring. Logistic regression has remained a prominent approach among the available techniques. Its clarity is what makes it appealing. Without depending on intricate models that might not be transparent, it enables analysts to comprehend which factors influence the result. This component is still essential for organisations that have to defend choices to clients and regulators.
Still, accurate prediction is far from simple. High risk profiles are not always visible in the raw data. Some borrowers with stable incomes may still default. Others with riskier financial backgrounds might not. Building a reliable model depends not only on the data itself but on how well it is prepared and tested. That is where SPSS becomes useful. It offers a structured environment to manage variables, handle missing values, and run logistic regressions with clarity and precision.
This study uses a real dataset of bank loan applications to build and evaluate a logistic regression model in SPSS. The objective is to identify which client characteristics most influence the risk of default. The goal is not to push statistical limits but to show how a classic, accessible method can still deliver value in real credit environments.
The main query is straightforward. Is it feasible to use a simple logistic regression backed by clear data and meticulous analysis to precisely identify clients who are likely to default ?
If the answer is yes, the model can give banks a useful tool that is easy to use and comprehend in addition to being efficient.
2. Literature Review
2.1 Overview of Existing Credit Scoring Methods
Credit scoring depends on being able to group borrowers based on how likely they are to default. Different statistical models have been made over the years to help with this job. Some put more importance on being able to understand something, while others put more importance on being able to predict something. This part looks at the most common ways to score consumer credit, mostly using the work of Hand and Henley (1997), which is still an important reference in this area.
· Logistic Regression
People often think of logistic regression as the basic method for credit scoring. It shows how likely it is that a borrower will default based on a set of input variables. The model presumes a linear correlation between the log-odds of default and the predictors. It doesn't need to assume that the data is normal or has equal variance, which makes it a good choice for financial data.
The fact that it is open is what makes it strong. It is possible to directly interpret the effect of each variable on the default probability using coefficients and odds ratios. Banks and other financial institutions really like this clarity because they often have to explain credit decisions to auditors or regulators.
Logistic regression does surprisingly well in many scoring situations, even though it is simple. But it does have some problems. It cannot capture intricate non-linear relationships unless interaction terms or transformations are manually incorporated. Also, multicollinearity between predictors can change the estimates of the coefficients.
· Linear Discriminant Analysis (LDA)
LDA was a popular method before logistic regression became the standard. It presumes a normal distribution of predictors and uniform covariance matrices among classes. In practice, these assumptions are seldom satisfied in credit data. However, LDA can still work well if the variables are well-behaved.
LDA builds a discriminant score that is used to sort borrowers, while logistic regression directly estimates probabilities. The absence of intuitive interpretability, in contrast to odds ratios in logistic regression, has resulted in its diminished use in credit scoring applications.
· Decision Trees
Based on the values of the explanatory variables, decision trees divide the data into groups. The outcome is a tree structure, with each internal node representing a variable split and each leaf representing a predicted class. One of the best things about decision trees is that they look and make sense. Analysts and people who make decisions can see how a classification was made, step by step. Trees can naturally capture how variables interact with each other and work with both numerical and categorical data. But they are sensitive to small changes in the data, which can make the trees unstable. To fix this, people use pruning methods or ensemble methods like bagging.
Hand and Henley talked about how basic tree structures can be too fit. Without control mechanisms, trees tend to fit the training data too closely and don't work well with new data. This is a big problem in credit scoring, where model stability is very important.
· Neural Networks
Neural networks, especially feedforward architectures, have been used for credit scoring with good results. They can show how complicated, non-linear relationships between input variables and the target outcome work. But their "black box" quality is still a big worry in the financial services industry.
A neural network may have better raw accuracy than a logistic model, but it isn't good for situations where decisions need to be explained because it isn't clear how it works. Hand and Henley said that regulatory scrutiny needs models that can explain why they give certain results. Neural networks have a hard time with that, even though they are flexible.
Also, neural networks need more work to prepare the data and set the parameters. When the datasets are small, they usually don't give a big advantage over simpler models.
· k-Nearest Neighbors (k-NN)
k-NN has been tested in scoring problems, but it is not used as often in practice. The method puts a new borrower into a class based on the class of most of the borrowers who are closest to them in the feature space. It is easy to understand, but it is affected by irrelevant variables and the choice of distance metric. Hand and Henley said that k-NN isn't strong enough to be used for credit scoring. It doesn't give a probabilistic output and doesn't tell you much about how important each variable is. It also has trouble with high-dimensional data, where the idea of "closeness" doesn't mean as much.
· Ensemble Methods and Hybrid Approaches
When Hand and Henley wrote their review, ensemble methods were becoming more popular but weren't being used very much. Since then, techniques like boosting and random forests have become popular. These methods use more than one model to make things more accurate and stable.
Hand and Henley predicted this trend by pointing out the problems with using just one model. In scoring situations, using a combination of methods can strike a balance between being able to predict and being able to understand.
· Model Selection and Practical Considerations
Hand and Henley advised examining model performance and operational demands together. An accurate but complicated model may not be as valuable as a simplified one that meets business goals. Even though machine learning has advanced, logistic regression is still the most preferred scoring system due to this tradeoff.
They underlined the importance of selecting the proper variables, checking the model, and ensuring good data. Overfitting models or applying incorrect predictors in credit operations might be harmful. Evaluation should include correctness, stability, fairness, and legality.
· Summary Table: Method Comparison
2.2 Justification for Logistic Regression in Binary Classification Tasks
Logistic regression is a direct and effective way to predict whether something will happen or not, like when a client defaults on a loan. Logistic regression is made for binary classification, while linear models work best for continuous outcomes. It doesn't put the data in a framework that doesn't work. Instead, it respects the outcome variable's nature and gives results that are both valid and easy to understand.
In linear regression, the values that are predicted can be less than zero or more than one. This is a problem when trying to model probabilities, which should always be between zero and one. Logistic regression gets around this problem by changing the response using the log odds function. This change makes an S-shaped curve that shows how likely an event is to happen as the predictor variables change.
Take the example of a client’s income in a credit scoring context. If income increases, we might expect the likelihood of default to decrease. Logistic regression does not assume a fixed drop or rise but calculates how each unit of increase in income changes the odds of default. If the model returns a coefficient of minus 0.5 for income, the odds ratio is exp(minus 0.5), which is approximately 0.61. This means that for each increase of one unit in income, the odds of default decrease by 39 percent, assuming other variables remain constant.
One of the greatest strengths of logistic regression is that this kind of interpretation is possible for each variable. Age, employment status, or loan amount can all be included, and the model will estimate how each contributes to the outcome. The coefficients are not abstract; they tell a story that managers and analysts can follow.
According to Hosmer, Lemeshow, and Sturdivant (2013), the model's diagnostic tools are another benefit. Analysts can determine the statistical significance of variables, the adequacy of the model in fitting the data, and the potential undue influence of specific observations. Deviance residuals, for instance, help find places where the model has trouble. This may lead to a reexamination of the data, which could improve accuracy or help you understand special cases better.
Additionally, the model handles categorical variables well. Imagine including the person's marital status. By putting marital status into dummy variables like married or not married, the model can determine if it affects default. Assuming no other changes, married clients are 20% more likely to default than the reference group if the odds ratio is 1.2. This data can help banks manage risk.
Another useful benefit is that it is flexible. Logistic regression lets you see how different variables affect each other. If the effect of the loan amount is based on the borrower's income level, the model can include an interaction term. This makes the model more complex without making it too hard to understand.
Remember that logistic regression doesn't require rigorous assumptions. It doesn't imply predictors are normally distributed or have the same variances, unlike discriminant analysis. This is good for real data because it doesn't always follow perfect patterns. A lot of the time, financial datasets have missing values, outliers, or distributions that are not even. Even when these things happen, logistic regression is strong enough to give accurate results.
There is a lot of software that works with the method. Analysts can run the model and check the outputs in SPSS, R, or Python. They can also make visualisations like ROC curves or classification tables. These tools help explain how well the model works to people who aren't technical.
Being open is very important in places where there are rules. Banks and other lenders should be able to explain how they decide who gets a loan and who doesn't. Logistic regression meets this need by giving decision-makers a model that is both statistically sound and easy to use. It avoids the problems that come with unclear methods while still being able to make predictions.
It does have some limits, of course. The connection between the predictors and the log odds needs to be linear. If the data indicates a more intricate relationship, transformations or interaction terms may be required. But these changes are easy to make and don't make the model less clear.
2.3 Role of Tools Like SPSS in Credit Risk Analysis for Business Users
Credit risk evaluation is a core responsibility in banking, and business users—such as loan officers or credit analysts—need reliable models to support decisions without relying solely on data scientists. SPSS was designed for precisely this audience. Its interface allows professionals to build, test, and deploy predictive models using logistic regression, even with minimal programming experience. The IBM example using the file bankloan.sav demonstrates how the process can be applied to real customer data to assess loan risk.
· Preparing the Data for Analysis
The file bankloan.sav contains data on 850 individuals, including both past borrowers and future applicants. SPSS allows easy filtering of the first 700 records (past customers) to be used for model training. A random selection tool built into the software creates a sample that avoids bias in selection.
SPSS also helps prepare variables: continuous variables like Income or Age can be normalized or categorized; categorical fields like Education or Marital status can be transformed into dummy variables through guided dialogs. Handling missing data, a frequent issue in credit applications, is made accessible via the Missing Values Analysis module.
- Quote paper
- Nabil Nakbi (Author), 2025, Assessing Credit Default Risk Using Logistic Regression. A Transparent Approach to Scoring with the UCI Dataset and SPSS, Munich, GRIN Verlag, https://www.hausarbeiten.de/document/1618055
