Causal Auditing of RL Loan Pricing Policies Using Process Mining

The challenge

The pricing agent optimised margin against a simulator, so its live behaviour drifted into segments where uplift correlated with proxies for protected attributes. Compliance needed evidence, not intuition: which decision paths cause a higher price, and by how much.

Approach

• Reconstructed end-to-end applicant journeys from raw servicing and underwriting events into conformant event logs with PM4py, then mined variants and conformance deviations against the intended policy model.
• Modelled each mined variant as a treatment and estimated conditional average treatment effects on offered APR with EconML doubly-robust and causal-forest learners, including refutation tests for unobserved confounding.
• Built the batch feature and log-normalisation layer as Apache Beam pipelines so the same code powers backfills and daily incremental audits.
• Packaged the audit as a Kubeflow pipeline on Kubernetes with versioned artefacts, so every regulatory report is reproducible from a single run ID.
• Shipped a reviewer dashboard that ranks policy paths by effect size and confidence, letting risk officers accept, contest, or escalate each finding.

Outcome

Seventeen policy paths were shown to raise price without a causally defensible driver. Retraining with the audit as a constraint cut the disparity gap by 63% while holding portfolio margin flat, and the pipeline now runs nightly as a standing control.