Does a Bad Claim Experience Cost the Policy? Claim Handling, Customer Trust and Retention on a Synthetic Indonesian Life and Health Portfolio

A claim is the moment a policyholder learns what the policy is worth, and a rejected, partly paid or slow claim may end the relationship. This preprint studies that question end to end on a fully synthetic portfolio of 40,601 life and health policies held by 30,000 customers over six years, with 9,214 claims in ten linked tables, planted mechanisms, a randomised service-recovery pilot and counterfactual re-runs of the same world that give the true effects. RQ1 estimates how much a partial payment, a rejection with an explanation, a rejection without one, slow handling and a withdrawn claim raise 12-month voluntary exit against smooth claims, using crude, regression, inverse-probability-weighted and augmented estimators with bootstrap intervals, E-values and an adjustment ladder; it shows that the usual comparison of claimants with all policies from issue reverses the sign (immortal-time bias), measures spillover to the customer's other policies, and uses staff shortage as an instrument for handling time. RQ2 builds early-warning scores at four dates from policy information only to three weeks after the decision, compares logistic regression, gradient boosting and an Explainable Boosting Machine on later claims, quantifies how much claim information adds, and shows how much leaked fields and undated records inflate accuracy. RQ3 evaluates the randomised recovery call (intention-to-treat and complier effects, per-protocol shortcuts, subgroups), compares uplift models and targeting rules by Qini curves and net value under explicit cost and margin assumptions, and sizes the next pilot. The findings are reported with their uncertainty, including the effects the data cannot detect. All data are synthetic: no data from IFG Life or any other insurer, bank or hospital were used, and no result is a finding about a real company. The archive contains the paper, the data with a data dictionary and ground-truth files, the code, a Jupyter notebook and the results, and regenerates every number in the paper with one script.