Causal inference paper tackles extreme events with 25% better accuracy on tail risks

Standard double machine learning (DML) for causal inference deliberately suppresses extreme outcomes to stabilize bulk averages—a dangerous practice for high-stakes domains like finance or climate where rare, catastrophic events are the primary target. Eichi Uehara's new estimator tackles this head-on by emitting a structured tail-shape output alongside the standard dose-response function. Its key innovation, the PDHTE+JK diagnostic, evaluates per-treatment tail shape using outcome residuals centered by a pilot median, breaking the circular dependence that plagues existing methods. The output includes four quantities: tail shape ξ(t), deep-tail return levels, conditional shortfalls, and the recovered mean ADRF, plus an explicit refusal mechanism that declines extrapolation when extreme-value modeling is unsupported by data.

Empirically, the estimator outperforms kernel-weighted quantile regression across heavy-tailed panels: a 11% reduction in deep-tail (α=0.001) return-level MAE and a 25.5% reduction in conditional-shortfall MAE. In sample-scarce regimes (n≤2000), gains reach 20-29%. On freMTPL2 motor-insurance claims, it triggered an explicit extrapolation refusal on the log-claim scale—a capability neither quantile regression nor loss-only DML can produce. These advances make the method practical for insurance pricing, financial risk management, and climate loss modeling, where accurately estimating the tail of the distribution matters far more than getting the average right.