Robust M-Estimation for Spliced Power-Law Models Under Cusp-Type Non-Regularity
Statistical modeling of heavy-tailed data with bounded or finite body behaviour often relies on composite or spliced distributions characterized by structural thresholds. Because the probability density function features a sharp directional shift (a cusp) at the join point, the classical log-likelihood function exhibits a discontinuity in its first derivative, causing standard Cramér-Rao regularity conditions and maximum likelihood frameworks to break down. To overcome this non-regularity, we develop a robust, log-transformed asymmetric Huber M-estimation framework for the structural threshold, coupled with conditional shape maximum likelihood estimators. We derive a closed-form parameter-balancing condition that analytically guarantees Fisher consistency under structural tail asymmetry. Furthermore, we establish the joint asymptotic normality of the hybrid Maximum Likelihood and M-estimator system and derive its explicit sandwich covariance structure. A computationally efficient profiled Z-estimation algorithm is implemented to reduce the joint estimation to a one-dimensional root-finding problem. Finite-sample properties are validated through Monte Carlo simulations, and the practical utility of the framework is demonstrated on the benchmark Danish Fire Insurance claims dataset.