CausalML Part I: Your 5-Min Primer

This series on Causal Machine Learning (CausalML) offers a step-by-step guide through its basic concepts. We begin with the fundamental principles, moving on to the intricacies of causal models and their real-world applications. We'll delve into data-driven insights, focusing on effect heterogeneity and decision-making based on choosing policies. The series wraps up with a practical look at evaluating the reliability of causal estimates, making complex CausalML concepts accessible and applicable.

Introduction to CausalML

Causal Machine Learning (CausalML) is an advanced field intersecting traditional machine learning and causal inference. It focuses on understanding not just correlations but causal relationships between variables. It goes beyond predicting outcomes based on patterns in the data (machine learning) to understanding how changing one variable can affect an outcome.

Figure 1: CausalML combines the predictive power of ML algorithms with the causal inference methods traditionally used in econometrics

CausalML combines the predictive power of ML algorithms with the causal inference methods traditionally used in econometrics. A key aspect of econometrics aims at providing solutions to ‘what-if’ scenarios via experimental or observational data by examining causality thus leading to more informed decision making; Many of the models and methods known from ML such as decision trees, nearest neighbour methods etc. are adapted to a causal setting in CausalML where they are used to estimate treatment effects. (More on treatment effects later in the article)

---

Understanding Causal Models: A Simple Use Case

Case Study: Bank A's Credit Policy

Consider Bank A, keen on evaluating its new credit policy, which includes stricter loan approval criteria. Traditional machine learning methods might analyze historical loan approval and default data to identify patterns. However, these methods primarily reveal correlations (e.g., higher income correlating with lower default rates) and might inadvertently embed biases present in the data; Causal ML delves deeper, aiming to understand cause-and-effect relationships. For instance, it can help Bank A understand whether its stricter criteria genuinely reduce default rates or if this effect is due to other factors like economic shifts or customer demographics.

To understand this scenario better, lets examine two possible approaches the bank’s research and analytics team could take in choosing an effective credit policy.

Figure 2: Illustration of how a bank might approach choosing it credit policy using a traditional ML approach [Image by Author]

Traditional ML in this scenario will involve some process of data collection and preprocessing focusing on attributes like credit scores and financial history. Machine learning models, ranging from linear regressions to complex neural networks, are then trained on this data to predict credit risk. Key to this process is the model evaluation, where metrics like accuracy, precision, and recall are scrutinized to ensure robustness. The outcome of these predictions directly informs the bank's credit policy, dictating terms such as interest rates and loan amounts. The model undergoes continuous refinement with new data, ensuring its adaptability to evolving financial trends and customer behavior patterns.

Figure 3: An illustration of the same scenario but using CausalML approach this time, this can be more effective as causalML techniques can uncover different policies for different groups and help it better informed and targeted decision making. [Image by Author]

CausalML in this case would involve gathering detailed data, including not just traditional financial indicators but also variables presumed to have causal relationships with credit risk. The process involves causal discovery techniques, like constructing Directed Acyclic Graphs (DAGs), to identify and visualize potential causal pathways. This leads to the development of causal models that decipher the impact of various factors on credit risk, distinguishing between correlation and causation. The insights derived from these models are critical in shaping nuanced credit policies, as they reveal the underlying 'why' behind credit risk trends. This approach enables banks to design more effective, evidence-based policies, which are continually updated as new causal insights emerge, ensuring a more dynamic and informed approach to credit risk management.

---

Why CausalML?

CausalML isn't just about establishing cause-and-effect relationships; it's also crucial for understanding the nuances of these relationships in diverse scenarios. Here are some useful instances;

• Estimation of Heterogeneous Treatment Effects:

  • What it Means: Heterogeneous treatment effects refer to the variation in the impact of a treatment or policy across different subgroups or individuals. For instance, a credit policy might affect borrowers with different credit histories in varied ways.

  • Why It's Important: Understanding these differences is essential for tailoring policies or treatments to maximize effectiveness and fairness. In the banking scenario, this knowledge can lead to more personalized loan products, optimizing the balance between risk and customer service.

    Example: If a bank implements a new loan policy, CausalML can reveal how this policy differentially affects customers with high credit scores versus those with lower scores.

• Better Explainability In Scenarios where Black-box models are not sufficient:

  • What it Means: In credit risk modeling, enhanced explainability involves clearly understanding how a model determines risk levels, moving away from opaque 'black-box' approaches. This means breaking down a model's decision-making into understandable factors.

  • Why It's Important: Explainable models are essential in banking for compliance with regulations, effective risk management, and building customer trust. They allow banks to identify and communicate the specific reasons behind credit decisions.

    Example: For instance, a bank uses a CausalML-based model for loan approvals. Unlike a black-box model, this explainable model reveals how factors like employment history and spending habits influence risk assessments. This clarity aids in personalizing loan offers and refining credit policies based on solid, interpretable data insights.

---

Useful terminologies

To understand baseline terminologies that we will use in this series of articles, lets consider bank (B) that needs the to strengthen its lending policy, this is called a goal or outcome; Lets assume the bank decided to run a six(6) month experiment; the pool of customers would be customers the bank has decided salvageable (low-risk to mid risk-of-default customers), this called population or sample, and the method applied here is called a selection strategy. There experiment considers how the following policies would have impacted customer behaviour.

i) Lower credit limit: reduce the credit limit by approx 25%

(ii) Payment plan: split the debt up into six monthly installments

iii) Lower credit limit & Payment plan: application of both policies

These are called Treatments.

The bank evaluates the outcome of this experiment by its impact on a proxy variable named 'LTV - Lifetime Value' this is called a metric - more on this later. In order to observe this change, the bank has split the groups of customers into four groups, the polices were applied individual on three groups treatment groups, while the last group has no policy effect applied to it control group.

Causal Models: In CausalML, the models represent relationships between variables beyond mere correlation. In the above scenario, the model would estimate the causal impact of each policy, the model can then predict what might happen under each scenario. For instance, it might mean while credit limit leads to lower default rates for a particular group, a structured payment plan might be the best option for another group.

Causal Graphs: Causal graphs are visual representations used to depict and analyze cause-and-effect relationships between variables. They typically are depicted as a network of nodes and arrows, illustrating how one factor can influence another. They are crucial in identifying and understanding the direct and indirect causal pathways within complex datasets, providing a clear framework for analyzing how changes in one variable might impact others.

Decision Cycles: Decision cycles refer to the iterative process of implementing a policy, observing its outcomes, and refining the approach based on these observations. In our example, the bank might go through several decision cycles, adjusting its credit scoring algorithm and policies based on observed loan performance and other metrics.

Choosing Policy: This involves selecting the best course of action based on the insights derived from CausalML analysis. The bank would use the data from its treatment and control groups, along with the metrics, to decide whether to fully implement the new credit scoring system, modify it, or revert to the old system.

In the next part, we would explore

• Explore causal models while working with data from our use-case

• Explore effect modifiers and confounding variables

• Understanding doubly robust learning, fit a linear double robust learner.