Causal Inference and Bias Correction Methods in Ant Financial Risk Control

Background of Causal Inference – Causal effect estimation aims to quantify how a treatment (e.g., issuing a coupon) changes an outcome (e.g., conversion rate) beyond what would happen without the treatment. Randomized experiments provide unbiased estimates but are costly; observational data are abundant but suffer from confounding bias, illustrated by Simpson’s paradox.

Confounding Bias and Its Mitigation – When a hidden variable (e.g., user payment frequency) influences both the treatment assignment and the outcome, the observed effect is distorted. Mitigating this bias requires either (1) using pre‑intervention differences to model the bias or (2) leveraging a small amount of unbiased experimental data together with large biased observational data.

Time‑Series De‑biasing: Double Difference (DiD) and DiDTree – The double‑difference approach subtracts pre‑intervention outcome differences to remove confounding bias, assuming a parallel trend. DiDTree extends this by embedding a parallel‑trend loss into a tree‑based model and using gradient‑boosting to enforce the assumption locally in each leaf, enabling individual‑level effect estimation.

Experiment‑Data De‑biasing: Shrinkage Causal Trees and FAST – By fusing unbiased experimental data with biased observational data through shrinkage estimation, a weighted combination of the two models yields lower mean‑squared error. This principle is applied to causal trees: leaf‑level effects are estimated via shrinkage, and a parallel‑trend loss modifies the splitting criterion. The resulting Fast Shrinkage Tree (FAST) and its ensemble (FAST‑RF) achieve robust performance across simulated and real datasets.

Experimental Evaluation – Extensive simulations varying the degree of parallel‑trend satisfaction and treatment‑group imbalance show that DiDTree and FAST methods remain stable, outperforming traditional DiD, meta‑learners, causal forests, and neural‑network baselines. Real‑world experiments on Ant’s risk‑control data confirm the superiority of FAST‑RF in estimating average treatment effects.

Applications in Ant Group – In credit‑limit adjustments, price reductions, and other risk‑control scenarios, these causal de‑biasing techniques enable accurate uplift estimation without costly large‑scale randomized trials, directly improving business outcomes.

Q&A Highlights – The audience asked about the relationship between DiDTree and causal forests, handling non‑monotonic uplift rankings, and the role of parallel assumptions in leaf nodes; the presenter clarified the methodological distinctions and practical remedies.