Addressing Selection Bias in Credit Risk Modeling

PDE By islever Editorial Last updated

Preparing data for AI and ML.

Short answer

C Remove the bias from the data and collect applications that have been declined loans.

Question

You work for a bank. You have a labelled dataset that contains information on already granted loan application and whether these applications have been
defaulted. You have been asked to train a model to predict default rates for credit applicants.
What should you do?
  1. Increase the size of the dataset by collecting additional data.
  2. Train a linear regression to predict a credit default risk score.
  3. Remove the bias from the data and collect applications that have been declined loans. Correct Answer
  4. Match loan applicants with their social profiles to enable feature engineering.

Key Takeaways

The main trap is choosing an algorithm (like linear regression) while ignoring that the training data is structurally incomplete due to survivorship bias.

This question highlights the critical importance of handling selection bias in machine learning datasets for credit scoring. The correct approach involves including declined loan applications to ensure the model learns from all applicant types.

Common mistake: Choosing B (Linear Regression) because it seems like a simple predictive tool, failing to realize that without data on declined applicants, no algorithm can accurately predict risk for the general population.

Expert Analysis

The core issue here is selection bias (specifically survivorship bias). The dataset only contains applicants who were already granted loans. It completely lacks data on applicants who were declined. A model trained solely on accepted applicants will learn the characteristics of 'good' or 'acceptable' risk, but it will have no knowledge of what constitutes a high-risk applicant that should be rejected. Consequently, when applied to new applicants, it cannot distinguish between low-risk and high-risk candidates effectively because it has never seen the high-risk patterns.

Option C addresses this fundamental flaw by suggesting the inclusion of declined applications. This allows the model to learn the full spectrum of creditworthiness, distinguishing between those who get loans and those who don't, thereby enabling accurate prediction of default rates for the entire population of applicants. Without this data, any model is inherently biased and incomplete.

Option B suggests linear regression. While one could technically fit a line to binary data, it is statistically inappropriate for classification problems where the outcome is discrete (default vs. no default). Logistic regression would be the standard choice for probability estimation, but even then, it fails due to the missing data in Option C. Comments highlighting B often miss the critical context of the dataset's composition. As noted in community discussions, the key trap is focusing on the algorithm (B) rather than the data integrity (C). The presence of social profiles (D) raises privacy issues and is not a primary solution to the structural data gap. Increasing data size (A) without addressing the missing segment of declined applicants merely scales the bias.

Exam Strategy

When faced with a scenario involving model training, always audit the dataset's representativeness before selecting an algorithm. If the data excludes a significant portion of the target population (e.g., only approved loans), prioritize correcting the sampling bias over tuning the model. Look for options that address data completeness and fairness first.

Frequently Asked Questions

Why is linear regression incorrect for predicting default rates?

Linear regression predicts continuous values, whereas credit default is a binary classification problem (default/no default). Logistic regression is more appropriate for estimating probabilities, but neither works well if the training data is biased.

What is survivorship bias in this context?

Survivorship bias occurs when the analysis only includes entities that 'survived' a process (in this case, loan approval). By excluding declined applications, the model misses crucial data about high-risk profiles that were rejected.

More PDE FAQ →

Community Votes

B
55%
C
20%
D
20%
A
5%

55% of anonymous learners picked answer B. Votes are pick records left by other test-takers — they are not the verified answer.

Discussion 51 contributions

Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.

From public community threads 30 of 51

  1. GHN74 👍 40
    A is incorrect as you need to work with the data you have available C is an optimisation not a solution D is ethically incorrect and invasion to privacy, there could be several legal implications with this B although oversimplified but is a workable solution
  2. szefco 👍 23
    Question says: "to predict default RATES for credit applicants". It is not binary classification, so Linear Regression would work here. I think B is correct answer.
  3. sumanshu 👍 21
    We have labelled data that contains whether a loan application is accepted or defaulted - So Classification Problem Data. We need to predict (Default Rates for applicants) - I think whether application will be granted or defaulted. - So Binary Classification. No option matches the answer. - if we mark 'B' - It should be Logistic Regression, Instead of Linear Regression
  4. woyaolai 👍 8
    I used to be a Credit Risk modeler and I think this question is stupid.
  5. JayZeeLee 👍 6
    D. A - Incorrect. There's no indication that size was the problem. B - Incorrect. The model is likely to be a classification model. So linear is not the best fit. C - Incorrect. You wouldn't need declined applications. How can you default without getting an approved application. It's not relevant.
Show 25 more comments
  1. dambilwa 👍 6
    Question not worded properly. Option [B] looks closest
  2. TVH_Data_Engineer 👍 5 Answered B
    To predict default rates for credit applicants using the labeled dataset of granted loan applications, the most appropriate course of action would be: B. Train a linear regression to predict a credit default risk score. Here's the rationale for this approach: Appropriate Model for Prediction: Linear regression is a common statistical method used for predictive modeling, particularly when the outcome variable (in this case, the likelihood of default) is continuous. In the context of credit scoring, linear regression can be used to predict a risk score that represents the probability of default. Utilization of Labeled Data: Since you already have a labeled dataset containing information on loans that have been granted and whether they have defaulted, you can use this data to train the regression model. This historical data provides the model with examples of borrower characteristics and their corresponding default outcomes.
  3. medeis_jar 👍 5 Answered C
    We have only loan acceptance data, we do not have information about folks not getting loans thus our data is biased.
  4. arnabbis4u 👍 5
    Answer B
  5. hamzam 👍 4
    I think B is the best answer
  6. MaxNRG 👍 3 Answered D
    A and B dont make sense - labelled dataset. C doesnt either, What should we do with the applications? 🙂 D looks correct, in order to find new features for the model.
  7. changsu 👍 3 Answered B
    B or D, and I think B because the credit score already means a profile and no more data given but approved applications and defaults.
  8. ducc 👍 3 Answered B
    I did work for a real project We always use linear regression or AUC for predicting default in loan processes. So I go for B
  9. zxing233 👍 3
    D you don't have the defaulted rate in hand only defaulted or not... it's classification
  10. rocky48 👍 3 Answered B
    B. Train a linear regression to predict a credit default risk score.
  11. Deepakd 👍 3
    SCORE is continuous. So linear regression is right.
  12. Azlijaffar 👍 2
    Data that you have is binary - Defaulted or not. You want default rates - Linear Regression. HOWEVER. The data that you have is for "ALREADY GRANTED" loan applications and whether they have defaulted or not. But you want to "train a model to predict default rates for credit applicants", which would include applicants who would not be granted the loan. If you just work on that dataset your model will not be as accurate as it wont have considered profiles of applicants that normally would not be granted those loans in the first place. Am I missing something here?
  13. prasanna77 👍 2
    we dont have to predict risk score,we have to predict default rates, B and C are not correct. We need to get data of person where the credit score is available and do train the dataset.Social profiles I believe is about persons.So manipulate data already present in loan department (risk score) also present and other data,Use feature engineering to predict
  14. Mathew106 👍 2 Answered B
    Linear regression is not the good way to solve such a problem, but you can totally apply linear regression to solve a classification problem. Just set the labels to numeric values 0 and 1 and linear regression will try to predict a value inbetween and round to the closest label (0 or 1). Totally not the way to go about it, but actually it's possible.
  15. Jlozano 👍 2 Answered D
    A-It does not suggest a data size problem. B-We need a score. Linear regression cannot be an option. C-We need accepted loan applications. D-Complex to implement, but the most correct answer.
  16. ladistar 👍 2
    The question asks about default RATES, as in you are predicting a continuous variable, not a discrete one (classification). This is a regression problem, so choice B.
  17. Aaronn14 👍 2
    You cannot predict rate. You predict a realization, which is either default or not. This question is terribly written.
  18. omar_bahrain 👍 2
    A is the correct answer. please read: https://www.yields.io/blog/how-to-fix-bias-in-machine-learning-algorithms/
  19. Besss 👍 2 Answered B
    default rates can be predicted with linear regression.
  20. Whoswho 👍 2
    Answer should actually be a logistic Regression model
  21. Azlijaffar 👍 2
    So answer should be C i think.
  22. Mr_Stark 👍 2 Answered B
    B is correct option
  23. yxyj 👍 2
    D is right
  24. carbonscave 👍 1 Answered C
    C is correct: The dataset suffers from a severe case of selection bias (specifically, survivorship bias) because it only contains data from applicants who already passed the bank's initial screening to receive a loan. If you train a model exclusively on accepted applications, it will not learn the characteristics of high-risk applicants who are typically rejected. To build an accurate model for all future credit applicants, you must include data from declined applications (a process known in the financial industry as reject inference) to eliminate this bias.
  25. kishanu 👍 1
    I guess it is A, as I think we have data insufficiency. The Labeled data is for classification, where the application is granted or not, whereas the question is to train a model to predict the credit score which is a regression model. So, we need to increase the size of the dataset, could be adding more features, labels etc..

Reader comments & corrections

No reader corrections yet — spotted an error or have a note? Share it below.

Log in to comment, report an error, or add a note about this question.

How to establish consistent BigQuery spend?

S-grade

Triggering Airflow DAGs Reactively in Private Cloud Composer

S-grade

Monitoring Cloud Bigtable for Cluster Scaling Decisions

S-grade

How to resolve Cloud Composer 2 worker pod evictions and task failures?

S-grade

← Back to PDE Study Guide