Framing the Loan Default Problem

Defining the Business Problem and Objectives

First, let's establish why a financial institution would invest in predicting loan defaults. It's not just about predicting an outcome; it's about managing a fundamental business risk.

To see this in a real-world context, please watch the following segment from a talk by a data scientist at Zopa, a peer-to-peer lending company. Pay close attention to the three main reasons they give for needing to assess and manage credit risk.

As you saw, the objectives go beyond simple profit and loss. They include:

• Lending responsibly: Avoiding putting borrowers in a financially difficult situation.
• Protecting investors: Ensuring that the people lending money are confident they will be repaid.
• Maintaining reputation: Being seen as a credible and responsible lender in the market.

These business objectives are often formalized within a regulatory framework. Financial institutions must calculate and provision for potential losses.

To understand this formal context, please read the "Business problem" section of the following document from DataRobot.

The key takeaway is the formula for Expected Credit Loss:
$ECL=PD\times LGD\times EAD$

Our project will focus on modeling the Probability of Default (PD), which is the component that machine learning is best suited to predict. By building an accurate PD model, we directly address the business objective of minimizing Expected Credit Loss.

Identifying and Understanding Stakeholders

Now that we know why we're building this model, the next question is who it's for. A model is only useful if it's understood, trusted, and used by the right people. This requires a stakeholder analysis.

The following short video provides an excellent, general-purpose framework for identifying, prioritizing, and understanding stakeholders in any project.

Now, let's apply this thinking to our loan default scenario. Who are the specific stakeholders in a credit risk project? The DataRobot document you viewed earlier provides a clear list.

Combining the framework from the video with the roles from the document, we can profile our key stakeholders:

• Decision Executors (Underwriting Team): These are high-interest stakeholders. They will use the model's output daily to make decisions on loan applications. Their primary motivation is efficiency and confidence in their decisions. They need the model's output to be clear and easily interpretable.
• Decision Managers (Chief Risk Officer, Portfolio Managers): These are high-power, high-interest stakeholders. They are responsible for the overall financial health and risk exposure of the loan portfolio. They are motivated by regulatory compliance, portfolio performance, and strategic risk management. They need to trust the model's validity and understand its impact on the business's bottom line.
• Decision Authors (Data Scientists, Model Risk Monitoring Team): This is your team. You are high-interest and have influence over the project's technical direction. Your motivation is to build an accurate, robust, and fair model. You must also ensure the model is well-documented and can pass audits from managers and regulators.

By understanding these different perspectives, you can design a project that meets everyone's needs, which is essential for successful adoption.

Translating to a Supervised Classification Task

The final and most crucial step in project framing is translating the business problem into a well-defined machine learning problem. This involves moving from vague business goals to precise technical specifications.

A systematic approach is essential here. Industry frameworks like the Cross-Industry Standard Process for Data Mining (CRISP-DM) provide a roadmap. The work we've done so far falls squarely into its first phase: Business Understanding.

The following resource provides a structured set of questions that a data scientist should ask to bridge the gap between business needs and model design.

As the article highlights, we must define several key components to create a "model design":

1. The Business Problem: Minimize credit loss by identifying applicants likely to default.
2. The ML Problem Type: We can frame this as a binary supervised classification task.
3. The Unit of Analysis: The model will make a prediction for each loan application.
4. The Target Variable (Operational Definition): This is the most critical definition. What, precisely, constitutes a "default"? This is a business decision. A common industry standard is to define a default as a loan that is 90 or more days past due.
   • Our target variable, is_bad or default_status, will be 1 if the customer defaults (e.g., reaches 90 days past due) and 0 otherwise.
5. The Performance Window: Over what period after the loan is issued will we check for this default condition? This could be 12 months, 24 months, or the life of the loan.

Let's see a practical example of how Zopa defined their target variable.

This translation is the cornerstone of the entire project. Having a clear, unambiguous definition of the target variable, agreed upon with business stakeholders, ensures that you are building a model that solves the right problem.

Conclusion

In this lesson, we established the critical foundation for our loan default prediction project. You have learned how to move from a high-level business need to a concrete, machine-learning problem statement.

Key Takeaways:

• Business Objectives: The primary goal is to manage credit risk, driven by the need to lend responsibly, protect capital, and maintain a strong market reputation. This is formalized by the need to estimate and minimize Expected Credit Loss (ECL).
• Stakeholder Framing: A successful project must meet the needs of diverse stakeholders, including underwriters (executors), risk officers (managers), and the data science team (authors).
• Problem Translation: We translate the business problem into a binary supervised classification task by:
  • Identifying the unit of analysis (a loan).
  • Creating a precise operational definition for the target variable (e.g., a loan is in "default" if it is 90+ days past due).
  • Establishing a performance window to observe the outcome.

Next Steps:

Now that we have defined what we want to predict, the next logical step is to determine how we will measure success. In the next lesson, we will establish an evaluation framework, selecting metrics that are sensitive to the business costs of misclassification in a loan default scenario.