An Interpretable Gradient-Boosted Risk Model for Detecting Scam-Induced IRA Distributions Among Investors Aged 65 and Older

Older investors who are deceived by romance, impersonation, and investment scams frequently fund the scammer by withdrawing money from individual retirement accounts (IRAs).These withdrawals are authorized by the account owner, so identity verification and credential controls do not stop them.We develop and validate an explainable risk model that scores IRA distribution requests for the probability that the request was induced by a scam.Using administrative records from a national retail brokerage and IRA custodian covering 412,806 distribution requests submitted by 187,340 account owners aged 65 and older between January 2021 and December 2024, of which 1,962 were later confirmed as scam-induced, we compare seven model families under a strict temporal holdout.A gradient-boosted tree model achieved an out-of-time AUROC of 0.919 and identified 59.7% of scam-induced requests while alerting on 1% of all requests, against 9.7% for the custodian's incumbent rules engine.A monotone-constrained variant lost almost no discrimination while producing explanations that compliance reviewers judged more coherent.SHAP decomposition showed that new or recently changed destination accounts, contact-detail changes in the preceding fortnight, unusual liquidation of long-held positions, and the requested size relative to historical distributions carried most of the signal.In a within-subject reviewer experiment with 24 fraud analysts, adding feature-level explanations to the risk score raised decision accuracy from 0.77 to 0.84 and cut median review time by 22%.Performance was stable across age bands and rurality, though false-positive rates were modestly higher for the oldest owners.We discuss implications for trusted-contact outreach and temporary holds under FINRA Rule 2165.