Causal Auditing of RL Loan Pricing Policies Using Process Mining

The challenge

A lender's RL agent sets loan offer terms at the point of sale, but the reward signal is confounded: historical rate assignment depended on the loan officer, the process variant the application took, and applicant-risk proxies that also affect acceptance. The objective was to build a pipeline that produces a causally-corrected reward model and then audits the trained policy to see what it actually learned to depend on.

Approach

• Reconstructed applicant journeys from the real BPI Challenge 2017 event log with PM4py, deriving process-mining features (variant, resource, offer position, elapsed time, prior offers) that become confounder candidates in the causal model.
• Built the demand model with EconML's LinearDML: a doubly-robust estimator combining a propensity model P(rate | X) and an outcome model P(accept | X, rate), producing reward(x, r) = P_hat(accept | x, r) * interest_revenue(r) - servicing_cost_proxy(x).
• Trained the RL agent inside an Apache Pekko actor-based simulator where each in-flight application is an actor, then audited the learned policy with causal-learn PC discovery, pm4py token-based conformance replay, and black-box feature-sensitivity analysis.
• Implemented distributed process mining with Apache Beam: per-case DFG construction for discovery and per-case token-based replay for conformance checking, parity-verified against single-machine pm4py on 193,849 events / 31,509 applications.
• Orchestrated the training and audit as a Kubeflow pipeline on Kubernetes, with versioned artefacts and Mermaid-diagram exports so every finding is reproducible from a single run ID.

Outcome

The pipeline successfully ties the three fields together: process mining enriches confounder candidates, causal inference estimates the reward, and RL learns a policy that is then audited back against the same causal and process-mining tools. A concrete audit finding was that the trained policy converged to a nearly applicant-independent rate, showing near-zero sensitivity to CreditScore, NumberOfTerms, and process-state features despite their discovered causal paths to acceptance. The DML estimate itself proved sensitive to nuisance-model specification , linear nuisances gave a negative theta (-10.4), while random-forest nuisances gave positive estimates that trend toward zero with better fit , which may indicate the presence of unknown confounders causing a real limitation for a further study.