Counterfactual Causal Inference for Credit‑Limit Modeling (Mono‑CFR)

The talk is organized into three parts: (1) research paradigms of causal inference, covering correlation vs. causation, the three basic assumptions, and the distinction between randomised and observational data; (2) the evolution of causal inference frameworks from randomised experiments to uplift models (S‑learner, T‑learner, X‑learner) and to representation‑learning approaches; (3) the Mono‑CFR counterfactual credit‑limit model designed for continuous treatments in the financial domain.

In the first part, the presentation reviews Judea Pearl’s three‑layer causal ladder (association, intervention, counterfactual) and distinguishes four ways relationships can arise: causal, confounded, selection bias, and reverse causality, illustrating each with real‑world examples such as exercise‑cholesterol and credit‑limit‑default.

The second part explains how uplift models estimate heterogeneous treatment effects and how representation‑learning methods (e.g., DeR‑CFR, VCNet) balance feature distributions across treatment groups by separating variables into instrumental (I), confounding (C), and adjustment (A) sets, enabling unbiased causal estimation from observational data.

The third part details the Mono‑CFR architecture, which consists of a credit‑limit propensity network that predicts the latent treatment propensity μ(T|X) to break the dependence between X and ΔT, and a risk‑monotonic network that enforces a monotonic relationship between ΔT and default risk Y using an ELU+1 activation whose derivative is always non‑negative.

Mono‑CFR models the profit function (profit = revenue – risk) and estimates per‑user revenue and default risk across credit‑limit buckets, accounting for confounding credit‑worthiness factors. The model learns a dose‑response curve ΔT → Y, validates it through offline interpretability checks and small‑traffic uplift experiments, and demonstrates that a 30% increase in credit limit can reduce default amount by over 20% while boosting loan volume and overall profitability.

Future work aims to separate instrumental and adjustment variables in a model‑free manner to improve risk migration for high‑risk user groups, and to continuously iterate the model by rolling observational data, validating with randomised samples, and feeding business decisions back into the system.