Predictive Modeling of Loan Hiring Propensity

Andressa de Souza; Ana Julia Righetto

DOI: 10.22167/2675-6528-202602667

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Faced with the financial sector’s challenge to improve efficiency in credit offering, machine learning predictive models were developed and compared to estimate customers’ propensity to contract personal loans. The methodology adopted was quantitative and applied, based on a public dataset with 5,000 records, obtained from the Kaggle platform. Four classification algorithms were trained and evaluated: Logistic Regression, Random Forest, XGBoost, and LightGBM, with a split into training and testing sets. Performance evaluation was based on AUC, Precision, Sensitivity (Recall), and F1-Score metrics, allowing for a robust technical analysis. The results indicated that the ensemble models presented superior performance to the reference model, with greater discriminatory capacity in identifying customers with a higher propensity to contract personal loans. Variable importance analysis highlighted that factors such as customer income, average credit card spending, and education level were among the main ones associated with the loan contracting decision. The application of machine learning models demonstrated potential to improve strategies in financial institutions, promoting greater efficiency in resource allocation and competitiveness in the market.

Keywords: financial institutions; machine learning; classification models.

1. Introduction

The financial sector faces the constant challenge of improving efficiency in credit offering and customer acquisition. Despite technological advancements, financial institutions still encounter practical limitations in conducting effective strategies, especially regarding the prediction of the probability of product uptake, such as personal loans (Špicas et al., 2023).

The absence of robust prediction mechanisms and poorly refined customer segmentation result in less differentiated pricing strategies and difficulty in identifying customers with a higher propensity to hire financial products. This limits the targeting of personalized offers and directly impacts profitability, as the prior identification of propensity to hire is crucial for efficient conversion.

In this scenario, the adoption of machine learning models emerges as a promising solution to mitigate such problems. These models allow for a predictive analysis of customer behavior, supporting more effective strategic decisions (Khodabandehlou and Rahman, 2017). The application of Data Science techniques in the corporate environment is fundamental to mitigate risks and extract relevant information from large databases, providing robust analyses for strategic decisions (Fávero and Belfiore, 2019).

The evolution of the use of machine learning models in the financial sector has been remarkable. Initially, simpler models such as Logistic Regression and Decision Tree were predominant. With the maturation of the field, more sophisticated methods, such as Random Forest and Support Vector Machines, were incorporated, offering gains in precision and efficiency. More recently, ensemble techniques, such as XGBoost and LightGBM, stand out for their ability to handle complex databases with a large volume of records (Khodabandehlou and Rahman, 2017).

Literature also highlights the influence of personal, socioeconomic, and behavioral factors on the decision to contract credit products (Contreras Pinochet et al., 2019). Factors such as age, income, education level, and length of relationship with the financial institution are frequently associated with this decision. Additionally, marketing aspects, including product characteristics, pricing policies, and promotional strategies, are relevant in the perception of value and differentiation between financial institutions in competitive markets (Baumann et al., 2007).

The ability to predict the propensity to contract credit products contributes significantly to customer acquisition efficiency. This allows for the prioritization of leads with a higher probability of conversion, the reduction of spending on ineffective marketing campaigns, and the optimization of pricing strategies. Thus, accurate prediction of customer behavior favors a more efficient allocation of resources and strengthens the competitiveness of financial institutions (Romanenko et al., 2024). In this context, the present work aimed to implement and compare the performance of different supervised classification algorithms applied to the prediction of customer propensity to contract loans, using a public database that simulates characteristics observed in real financial sector data, and to identify the variables that have the greatest influence on the decision to contract a loan through variable importance analysis, providing relevant insights for optimizing pricing strategies and customer segmentation in financial institutions.

2. Material and Methods

The present study was characterized by a quantitative and applied approach, aiming to implement and compare the performance of supervised classification algorithms. The objective was to predict the propensity of clients to hire loans and identify the variables of greatest influence in this decision, providing subsidies for optimizing strategies in financial institutions.

The database used, titled “Bank Personal Loan Modelling”, was publicly obtained on the Kaggle platform (Kaggle, 2017). The dataset, made available in 2017, contained 5,000 records. These records included demographic, financial, and behavioral information of clients from a financial institution, with the target variable, “Personal Loan”, being a binary indicator of the hiring (1) or non-hiring (0) of a personal loan.

Initially, an exploratory data analysis was performed to understand the database structure, identify behavioral patterns, and the distribution of variables. This stage also allowed for the evaluation of possible inconsistencies before the predictive modeling phase, ensuring the quality and suitability of the data for subsequent analyses.

The variables considered in the study included demographic data such as age, professional experience, number of people in the family, and education level. Financial information included income, average credit card spending, mortgage loan amount, and ownership of an investment account, certificate of deposit, or credit card. The target variable, “Personal Loan”, indicated whether or not a personal loan was taken out.

In the preprocessing step, variable names were standardized by removing spaces and converting to lowercase to avoid inconsistencies. The existence of missing values and the need to exclude variables without predictive value were checked. Variables were classified as numerical and categorical, even if they were already numerically encoded in the source database.

Next, the data were separated into training (70%) and testing (30%) sets, using the test set validation approach (James et al., 2023). This proportion was adopted to balance the number of observations for model training and maintain sufficient data for evaluation. The original imbalanced distribution of data was preserved during training.

For predictive modeling, four supervised classification algorithms were implemented. Logistic Regression was used as a reference model (Hosmer et al., 2013), along with more advanced ensemble techniques such as Random Forest (Breiman, 2001), XGBoost (Chen and Guestrin, 2016), and LightGBM (Ke et al., 2017). For hyperparameter definition, the default parameters of the libraries were chosen to ensure comparability and avoid overfitting (Fávero and Belfiore, 2019).

The stability of the results was evaluated using different random seeds, testing two distinct seeds to verify the consistency of the algorithms. Data processing and algorithm implementation were carried out in the Python programming language, within the Visual Studio Code (VS Code) environment. The Pandas, NumPy, Matplotlib, Seaborn, Scikit-learn, XGBoost, and LightGBM libraries were employed for data manipulation, visualization, and predictive modeling.

After the implementation of the models, a comparative performance analysis was carried out based on Accuracy, Precision, Sensitivity (Recall), and F1-Score metrics. Additionally, the ROC Curve and the Area Under the Curve (AUC) were used, considered suitable for binary classification problems with imbalanced classes (Fawcett, 2006). These metrics were derived from the Confusion Matrix, as described by Fávero and Belfiore (2019).

Accuracy measured the proportion of correct predictions relative to the total number of observations. Precision quantified the proportion of positive predictions that were made correctly, indicating the probability that a customer classified as likely would actually take out the loan. Sensitivity, or Recall, measured the proportion of actual positive cases correctly identified, reflecting the algorithm’s ability to find customers who would indeed take out the loan. The F1-Score, in turn, represented the harmonic mean between Precision and Sensitivity, being a fundamental metric for evaluating the model on imbalanced datasets (Fávero and Belfiore, 2019).

Additionally, the importance of variables (“feature importance”) estimated from tree-based models was analyzed. The objective was to identify the attributes that exerted the greatest influence on the estimation of the propensity to hire loans, providing subsidies for decision-making in financial institutions.

3. Results and Discussion

The exploratory data analysis revealed important characteristics of the dataset used for predictive modeling. Initially, an inconsistency was identified in the variable “experience”, which presented negative values in approximately 1% of the records. This finding was interpreted as a recording error, and the values were adjusted to zero, representing clients with no professional experience. Additionally, the target variable, “personal_loan”, which indicates the hiring of a personal loan, showed an unbalanced distribution, with 90.4% of clients not hiring the product and 9.6% hiring it. This asymmetry underscored the need to prioritize appropriate performance evaluation metrics for classification problems with unbalanced classes, such as Precision, Sensitivity, and F1-Score, instead of relying exclusively on Accuracy, as discussed by Fawcett (2006).

The histograms of the numerical variables, such as “income” (customer income) and “ccavg” (average credit card spending), indicated asymmetry in their distributions, with a concentration of observations in specific ranges and the presence of extreme values. This pattern is common in financial sector databases and reinforces the importance of robust models capable of handling this structure. The analysis of the distributions of these variables in relation to personal loan uptake revealed significant differences between the groups. Customers who took out loans had, on average, higher income and credit card spending, suggesting that these factors have relevant discriminatory potential in identifying the propensity to purchase the product.

The correlation matrix among the numerical variables identified a strong positive correlation between “experience” (professional experience time) and “age” (client’s age), an expected result, as professional experience time tends to increase with age. A moderate positive correlation was also observed between “income” and “ccavg”, which is consistent with financial behavior, where credit card spending capacity generally follows income. Although the presence of multicollinearity can influence the interpretability of coefficients in linear models such as Logistic Regression, tree-based algorithms like Random Forest, XGBoost, and LightGBM demonstrated robustness to this characteristic, minimizing its impact on predictive performance.

The performance evaluation of the predictive models was carried out based on metrics such as AUC, Precision, Sensitivity (Recall), and F1-Score, considered more suitable for scenarios with imbalanced classes, as observed in this study. Accuracy, although calculated, was analyzed with caution, as it can present high values even with a low capacity to identify the minority class, leading to imprecise interpretations. Sensitivity, which measures the proportion of identified true positives, was crucial for evaluating the model’s ability to capture customers with a real propensity for hiring, directly linked to potential revenue generation. On the other hand, Precision, which indicates the proportion of correct positive predictions, was fundamental for evaluating operational efficiency, minimizing unnecessary contacts.

The comparison between the classification algorithms revealed the superiority of models based on “ensemble” techniques in identifying customers with a higher propensity to hire loans. Logistic Regression, used as a benchmark model, presented the lowest Sensitivity, reaching 64.6%. This means that the model failed to identify 35.4% of the customers who would actually hire the loan, resulting in a significant loss of business opportunities. In contrast, the LightGBM, XGBoost, and Random Forest models demonstrated substantially superior performance in all evaluated metrics, confirming the effectiveness of ensemble approaches for this type of classification problem, as observed in the literature by Khodabandehlou and Rahman (2017).

Among the ensemble models, LightGBM and XGBoost stood out with the highest Precision and F1-Score values. Both algorithms achieved 95.6% Precision and 93.2% F1-Score, indicating that their positive predictions were highly reliable and that they presented an excellent balance between Precision and Recall. Random Forest also exhibited robust performance, with 97.0% Precision and 93.1% F1-Score, slightly below the other two in Recall. These results demonstrate that LightGBM and XGBoost are particularly suitable for use in marketing campaigns and pricing strategies, as they minimize the waste of operational resources while maximizing the identification of potential customers.

The ROC curves and AUC values corroborated the superiority of the ensemble models. Logistic Regression achieved an AUC of 95.1%, while Random Forest and XGBoost reached 99.7%, and LightGBM presented the highest value, with 99.8%. These values close to 100% indicate a high discriminatory capacity of the ensemble models in distinguishing between customers who do and do not contract loans. However, it is important to consider that this high performance may be partially attributed to the characteristics of the database used, which has a simplified nature and a relatively well-defined structure, with variables strongly related to the target variable, which may not reflect the same performance in more complex real-world environments, as pointed out by Fawcett (2006).

The superiority of LightGBM in Sensitivity and Accuracy has a significant managerial impact, especially in active loan offering scenarios. In a hypothetical scenario of a personal loan offering campaign for 100,000 customers, with a real conversion rate of 10% (10,000 customers with propensity to hire), the LightGBM model, with a Sensitivity of 91.0%, would identify approximately 9,100 of these customers. In contrast, Logistic Regression, with a Sensitivity of 64.6%, would identify only about 6,460 customers. This difference represents the loss of approximately 2,640 potential customers who would not be approached, directly impacting revenue generation. Financially, assuming an average loan amount of R$ 4,000 per customer, this gap translates into an additional potential of approximately R$ 10.56 million in credit volume granted.

In terms of operational efficiency, LightGBM’s higher Accuracy (95.6%) compared to Logistic Regression (83.0%) ensures a more effective allocation of the campaign budget. Considering positive predictions, LightGBM would generate approximately 4.4% of false positives, resulting in about 440 inefficient contacts for every 10,000 customers approached. Logistic Regression, in turn, would result in approximately 17.0% of false positives, meaning about 1,700 unproductive approaches. This difference highlights a significantly greater waste of operational resources and potential customer relationship wear when using the benchmark model. Thus, the higher computational complexity of tree-based models translates into greater economic efficiency, better utilization of commercial opportunities, and a competitive advantage for financial institutions, as highlighted by Fávero and Belfiore (2019).

The importance analysis of variables (“feature importance”), estimated by the Random Forest, XGBoost, and LightGBM models, revealed consistency in the main explanatory factors for the propensity to hire loans. The variables “income” (customer income), “ccavg” (average credit card spending), and “education” (educational level) consistently appeared among the most relevant in all ensemble models. These results reinforce the coherence of the models with expected behavior in the business context, as financial and socioeconomic characteristics are widely recognized in the literature as direct influencers of the decision to hire credit products (Contreras Pinochet et al., 2019). The identification of these factors allows financial institutions to refine their segmentation and offer personalization strategies.

Despite the promising results and the high predictive power of the selected models, the study presented some methodological limitations that deserve consideration. The main one lies in the use of a public database with a relatively low number of records (5,000), which may have contributed to the high AUC values, which may not be replicable in more complex real-world environments. Furthermore, the database was static, not capturing the evolution of customer behavior over time nor the impact of dynamic macroeconomic variables, such as fluctuations in the interest rate (Selic), inflation, or unemployment levels, which directly affect the propensity to take out loans. This characteristic limits the generalization of the results to more volatile market scenarios.

Another important limitation was the non-application of class balancing techniques, such as SMOTE, which may have restricted the performance of the benchmark model, Logistic Regression, and represents an opportunity for improvement in future studies. To validate the robustness and stability of machine learning algorithms in different economic scenarios, future work could benefit from using more comprehensive databases, including a wider range of temporal economic variables and indicators. Additionally, exploring hyperparameter optimization techniques, such as GridSearchCV, could maximize the predictive capacity of ensemble algorithms in real-world scenarios, further enhancing the practical applicability of the models in financial institutions.

In summary, the results obtained demonstrate that machine learning models based on “ensemble” techniques, especially LightGBM and XGBoost, significantly outperformed the Logistic Regression model in predicting customer propensity to hire loans. This superiority was evidenced by robust performance metrics, such as AUC, Precision, Sensitivity, and F1-Score, which indicated greater discriminatory capacity and efficiency in identifying potential customers. The analysis of variable importance confirmed that factors such as income, credit card spending, and education are determinants in this decision, providing valuable insights for optimizing pricing and customer segmentation strategies. The application of these models offers considerable potential to improve resource allocation efficiency and strengthen the competitiveness of financial institutions, contributing to more informed and strategic decision-making.

4. Conclusion

This study aimed to implement and compare supervised classification algorithms to predict customers’ propensity to hire loans, as well as to identify the most influential variables in this decision, in order to optimize pricing and segmentation strategies in financial institutions. It was found that models based on ensemble techniques, notably LightGBM and XGBoost, presented superior predictive performance compared to Logistic Regression, the benchmark model. This superiority was evidenced by metrics such as AUC, Precision, Sensitivity, and F1-Score, which indicated a greater discriminatory capacity in identifying customers with high propensity to hire. The analysis of variable importance revealed that customer income, average credit card spending, and education level were consistently among the most relevant factors. The application of these models demonstrates significant potential to enhance customer acquisition efficiency, optimize resource allocation, and strengthen competitiveness in the financial sector, by enabling more informed and strategic decision-making.

However, the results should be interpreted with caution, given the methodological limitations of the study. The main one lies in the use of a public database, with a relatively low number of records and static in nature, which may have contributed to the high AUC values and does not reflect the complexity of real environments, which include the evolution of customer behavior and dynamic macroeconomic variables. Furthermore, the non-application of class balancing techniques represented a restriction. For future studies, it is suggested to use more comprehensive databases, which incorporate a wider range of temporal economic variables and indicators, in order to validate the robustness and stability of the algorithms in more realistic scenarios. The exploration of hyperparameter optimization techniques can also maximize the predictive capacity of ensemble models, improving their practical applicability.

Bibliographic References

Baumann, C.; Burton, S.; Elliott, G. 2007. Predicting consumer behavior in retail banking. Journal of Business and Management 2(3): 79-96.

Breiman, L. 2001. Random forests. Machine Learning 45(1): 5-32.

Chen, T.; Guestrin, C. 2016. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. San Francisco, CA, USA. Anais… p. 765-794.

Contreras Pinochet, L. H.; Diogo, G. T.; Lopes, E. L.; Herrero, E.; Bueno, R. L. P. 2019. Propensity of contracting loans services from FinTech’s in Brazil. International Journal of Bank Marketing 37(5): 1056-1077.

Fawcett, T. 2006. An introduction to ROC analysis. Pattern Recognition Letters 27(8): 861-874.

Fávero, L. P.; Belfiore, P. 2019. Data Science for Business and Decision Making. Academic Press, London, UK.

Hosmer, D. W.; Lemeshow, S.; Sturdivant, R. X. 2013. Applied Logistic Regression. 3. ed. John Wiley & Sons, Hoboken, NJ, USA.

James, G.; Witten, D.; Hastie, T.; Tibshirani, R.; Taylor, J. 2023. An Introduction to Statistical Learning: with Applications in Python. Springer, New York, NY, USA.

Kaggle. 2017. Bank Personal Loan Modelling. Disponível em: https://www.kaggle.com/datasets/itsmesunil/bank-loan-modelling. Acesso em: 28 nov. 2025.

Ke, G.; Meng, Q.; Finley, T.; Wang, T.; Chen, W.; Ma, W.; Ye, Q.; Liu, T. Y. 2017. LightGBM: A highly efficient gradient boosting decision tree. In: Advances in Neural Information Processing Systems, 2017, Long Beach, CA, USA. Anais… p. 3146-3154.

Khodabandehlou, S.; Rahman, M. Z. 2017. Comparison of supervised machine learning techniques for customer churn prediction based on analysis of customer behavior. Journal of Systems and Information Technology 19(1/2): 65-93.

Romanenko, N.; Sharma, K.; Verma, S. 2024. Prediction of financial customer buying behavior based on machine learning. Journal of Artificial Intelligence General Science JAIGS 5(01): 125-131.

Špicas, R.; Neifaltas, A.; Kanapickienė, R.; Keliuotytė-Staniulėnienė, G.; Vasiliauskaitė, D. 2023. Estimating the acceptance probabilities of consumer loan offers in an online loan comparison and brokerage platform. Risks 11: 138.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform