Automatic Audit Judgment Using Causal Inference and Counterfactual Risk Estimation

Abstract

This paper addresses the instability and insufficient interpretation of automatic judgment in audit scenarios caused by large data scale, complex triggering mechanisms, and the coexistence of bias and drift. It proposes an automatic audit judgment algorithm framework that incorporates causal inference. The method models audit trigger signals as processing variables, uses transaction behavior, account status, and time context as observation features, and takes risk outcomes as the judgment target, constructing a unified process from processing mechanism modeling to bias-free discriminant learning. First, propensity probability modeling characterizes the probability of trigger signals occurring under given feature conditions to mitigate the interference of selection bias and confounding factors on the judgment boundary. Then, inverse probability weighting is introduced to achieve sample reweighting, and a dual robust counterfactual estimation is combined to improve the reliability of effect characterization. Based on this, a risk discriminator is trained to output a risk score, and causal consistency constraints are added to promote the alignment of the judgment output with counterfactual differences, thereby enhancing the auditability and verifiability of the results. The framework was compared with several related methods under the same data and process settings. The results show that the proposed method achieves more balanced performance on commonly used classification evaluation indicators and can provide an audit clue organization method that takes into account both risk ranking and mechanism explanation, providing practical technical support for digital auditing and intelligent risk control applications.