Technical Summary: Counterfactual Analysis via Large Language Models

Problem Statement

Counterfactual analysis seeks to predict outcomes under hypothetical scenarios to inform decision-making, a task particularly challenging in online lending where the return on investment (ROI) under alternative interest rates is unobservable. While traditional methods such as matching, randomized controlled trials, and survival analysis exist, they often rely on specific structural assumptions or require subsampling. Furthermore, standard machine learning (ML) models, while optimized for predictive accuracy on observed data, may fail to capture the causal mechanisms required to generate valid counterfactuals, especially when borrower behavior changes in response to unobserved interest rate shifts. This paper investigates whether Large Language Models (LLMs), specifically GPT-3.5, can serve as effective tools for constructing counterfactual loan outcomes, leveraging their vast pre-trained knowledge and reasoning capabilities.

Methodology

The study utilizes a dataset of 894,013 unsecured personal loans from LendingClub (2013–2020). The methodology is divided into two primary phases: predictive performance validation and counterfactual construction.

1. Predictive Performance Validation

To establish the LLM's suitability for counterfactual generation, the author first assessed its ability to predict actual loan outcomes (status and ROI) under observed interest rates.

• Baseline: The LLM's performance was compared against advanced ML algorithms, specifically XGBoost (for loan status classification) and gradient-boosted regression (for ROI prediction).
• Prompt Engineering: The study systematically evaluated various prompting strategies to elicit better performance:
  • Role-Playing: Shifting the LLM's persona from a "borrower" (which exhibited social desirability bias) to a "credit risk expert."
  • Information Sets: Incorporating loan specifics, borrower profiles, platform-level statistics, and crucially, the predictions of ML algorithms into the prompt.
  • Reasoning Techniques: Implementing "Chain-of-Thought" (CoT) and "Tree-of-Thought" (ToT) prompting. The ToT approach simulated a panel of four experts iteratively refining their predictions and reasoning processes.
• Encompassing Test: To determine if the LLM merely regurgitated ML predictions or provided incremental value, the author employed a forecast encompassing test (Chong and Hendry, 1986). This regression tested whether the LLM's prediction contained information not captured by the ML model.

2. Counterfactual Construction

For the counterfactual task, the author generated hypothetical ROI outcomes under alternative interest rates (proposed by Gopal et al., 2024).

• Prompt Design: The LLM was prompted to act as a panel of four credit risk experts. The prompt included loan details, borrower profiles, platform statistics, the actual observed outcome under the original rate (serving as an anchor), and the counterfactual prediction from a mixture cure model (an existing ML approach).
• Reasoning Process: The LLM was instructed to engage in a ToT discussion, critiquing and refining the ML prediction based on its pre-trained knowledge and the specific context of the loan, ultimately outputting a consensus status and duration in JSON format.
• Validation: Since counterfactuals are unobserved, validity was assessed through logical coherence. The author checked if the LLM's predictions adhered to intuitive economic rules: (1) loans that are charged off should remain charged off if rates increase, and (2) loans that are fully paid should remain fully paid if rates decrease.

Key Results

Predictive Performance

• Baseline vs. Engineered Prompts: A simple "borrower" role prompt yielded poor results (R-squared of 1.97%, F1 score of 4.7%). However, prompt engineering significantly improved performance.
• Optimal Configuration: The "Tree-of-Thought" prompt with four experts, incorporating ML predictions and platform statistics, achieved an R-squared of 2.84% for ROI prediction. This closely approaches the 3.48% achieved by gradient-boosted regression.
• Incremental Value: The encompassing test revealed a statistically significant coefficient (5.19% at the 1% level) for the LLM prediction when combined with the ML prediction. This indicates that the LLM identifies meaningful information beyond what the ML algorithm captures, rather than simply repeating the ML output.
• Bias Correction: The study noted that raw LLM predictions suffered from bias (likely due to pre-training corpus biases), which was corrected via a linear regression adjustment before calculating R-squared.

Counterfactual Construction

• Divergence from ML: The LLM's counterfactual predictions diverged from the mixture cure model in 46.2% of cases regarding status and duration. The LLM often revised the ML prediction, providing justifiable reasons based on borrower credit utilization and historical averages.
• Logical Coherence: The LLM demonstrated strong adherence to causal logic. 99.6% of predictions followed the rule that fully paid loans remain paid if rates decrease, and 85.3% followed the rule that charged-off loans remain charged off if rates increase.
• Outcome Differences: Under alternative interest rates, the LLM predicted a charged-off rate of 9.65%, which was closer to the historical LendingClub rate (12.34%) than the mixture cure model's prediction of 6.22%. This suggests the LLM may correct potential over-optimism in the ML model.

Significance and Claims

The paper claims that LLMs represent a novel and capable approach for counterfactual analysis, particularly in complex, information-rich domains like online lending. The author highlights two primary contributions:

1. Prompt Engineering as a Critical Lever: The study demonstrates that prompt engineering (specifically role-playing, CoT, and ToT) is as crucial as model fine-tuning. These techniques effectively elicit logical reasoning and causal understanding from LLMs, allowing them to approximate the performance of specialized ML algorithms.
2. Human-Machine Collaboration: The results suggest that LLMs exhibit human-like cognitive behaviors, such as social desirability bias (when acting as borrowers) and the ability to critically evaluate and refine machine forecasts (when acting as experts). This supports the literature on human-machine collaboration, where LLMs can serve as proxies for human reasoning in complex tasks.

The author modestly concludes that while LLMs may not yet outperform the best predictive models in raw accuracy, their ability to generate logically coherent, causally reasoned counterfactuals makes them a suitable tool for decision-making contexts. They further suggest that LLMs could be used to simulate human responses in preliminary field experiments, offering a cost-effective method for experiment design. The paper does not claim LLMs are a universal solution for all counterfactual problems but positions them as a promising, flexible alternative to traditional econometric and survival analysis methods in specific contexts.