The Power of Credit Scoring - Evaluating Machine Learning and Traditional Models in Swedish Retail Banking - Master’s Thesis in Finance (Spring 2023)

Aug 3, 2026

Upload by Oliver

Investigates and compares credit scoring models with an emphasis on machine learning versus traditional approaches in Swedish retail banking. Evaluates the proposed PLTR model, combining machine learning with logistic regression, and analyzes performance under different class-weight settings. Assesses trade-offs between predictive accuracy, interpretability, and economic impact. Results indicate random forest best predicts defaults, but interpretability is limited by model complexity. A penalized logistic regression is identified as the most practical substitute, offering improved interpretability with slightly lower accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages69

Reading Time2 h 54 min

AccessFree to download

Introduction

Literature review

Introduction of models

Classification models and their performance

Data processing and variable selection

Data splitting and resampling

Model choice and hyperparameter tuning

Ethics of credit scoring

Regulation for credit scoring in Sweden

Data

Introduction of the sample

Ethical overview of the sample

Data processing

Descriptive statistics

Splitting the data into train and test samples

Theory and method

Missing values and K-nearest neighbour

Cross-validation

Classification Methods

Penalized Logistic Tree Regression

Class Imbalance

Statistical measures of performance and interpretability

Economic measures of performance

Empirical strategy

Predicting the probability of default

Tuning hyperparameters

Class weighting

Results

Total Defaults

12 months defaults

5 months defaults

Prediction vs outcome

Discussion

Conclusion

Appendix

The Power of Credit Scoring: Evaluating Machine Learning and Traditional Models in Swedish Retail Banking Written by: Emma von der Burg Saga Str¨omberg Supervisor: Marcin Zamojski Master’s thesis in Finance, 30 hec Spring 2023 Graduate School, School of Business, Economics and Law, University of Gothenburg Abstract In this paper, we investigate and compare different credit scoring models, with special attention paid to machine learning approaches outperforming traditional models. We explore a recently proposed method called the PLTR model, which is a combination of machine learning and traditional logistic regression. In addition, we examine the models’ performance and analyze the economic impact for different class weights. The main purpose of this paper was to identify the most effective and practical approach for credit scoring in the Swedish retail banking context. The findings suggest that the model that most accurately predicts defaults is the random forest, but at a high cost of interpretability due to the models’ complexity. According to our findings, the optimal substitute for the random forest is a penalized logistic regression, as it compensates with interpretability, for slightly less accurate predictions. Acknowledgements We would like to express sincere gratitude towards our supervisor, Marcin Zamojski, for the valuable guidance and support throughout this project. This thesis would not have been the same without your feedback and honest opinions. We also want to thank the teachers at the University of Gothenburg for bringing us knowledge and inspiration throughout our studies at this school. Lastly, special thanks to Collector Bank for providing us with the data needed for the analysis in this thesis. Contents 1 Introduction 1 2 Literature review 3 2.1 Introduction of models ............................. 3 2.2 Classification models and their performance ................. 4 2.3 Data processing and variable selection .................... 6 2.4 Data splitting and resampling ......................... 7 2.5 Model choice and hyperparameter tuning ................... 8 2.6 Ethics of credit scoring ............................. 9 2.7 Regulation for credit scoring in Sweden .................... 10 3 Data 11 3.1 Introduction of the sample ........................... 11 3.2 Ethical overview of the sample ......................... 12 3.3 Data processing ................................. 12 3.4 Descriptive statistics .............................. 14 3.4.1 Descriptive statistics of the independent variables .......... 14 3.4.2 Descriptive statistics dependent variables ............... 16 3.5 Splitting the data into train and test samples ................ 17 4 Theory and method 18 4.1 Missing values and K-nearest neighbour .................... 18 4.2 Cross-validation ................................. 18 4.3 Classification Methods ............................. 19 4.3.1 Logistic regression ........................... 19 4.3.2 Decision trees .............................. 21 4.3.3 Random Forest ............................. 24 4.4 Penalized Logistic Tree Regression ....................... 25 4.5 Class Imbalance ................................. 26 4.6 Statistical measures of performance and interpretability ........... 27 4.6.1 Calculating error rates ......................... 29 4.7 Economic measures of performance ...................... 29 5 Empirical strategy 30 5.1 Predicting the probability of default ...................... 30 5.1.1 Tuning hyperparameters ........................ 31 5.1.2 Class weighting ............................. 31 6 Results 32 6.1 Total Defaults .................................. 32 6.1.1 Importance of Variables ........................ 32 6.1.2 Performance results ........................... 34 6.1.3 Class weight results ........................... 36 6.1.4 Opportunity costs vs credit losses ........

FAQs

Which credit scoring models are compared in the thesis?

The thesis compares traditional models such as logistic regression and decision trees with machine learning approaches, including random forest and a hybrid PLTR model that combines machine learning with logistic regression.

What is the main finding about default prediction accuracy?

Random forest provides the most accurate predictions of defaults, outperforming the other evaluated models.

Why is penalized logistic regression proposed as a practical alternative?

Penalized logistic regression is recommended because it preserves stronger interpretability than random forest while delivering slightly less accurate predictions, improving practical usability in credit scoring.

Aug 3, 2026

Upload by Oliver