An omnibus goodness-of-fit test based on trigonometric moments

Abstract: We present a versatile omnibus goodness-of-fit test based on the first two trigonometric moments of probability-integral-transformed data, which rectifies the covariance scaling errors made by Langholz and Kronmal [J. Amer. Statist. Assoc. 86 (1991), 1077--1084]. Once properly scaled, the quadratic-form statistic asymptotically follows a $\chi_2^2$ distribution under the null hypothesis. The covariance scalings and parameter estimators are provided for $32$ null distribution families, covering heavy-tailed, light-tailed, asymmetric, and bounded-support cases, so the test is ready to be applied directly. Using recent advances in non-degenerate multivariate $U$-statistics with estimated nuisance parameters, we also showcase its asymptotic distribution under local alternatives for three specific examples. Our procedure shows excellent power; in particular, simulations testing the Laplace model against a range of $400$ alternatives reveal that it surpasses all $40$ existing tests for moderate to large sample sizes. A real-data application involving 48-hour-ahead surface temperature forecast errors further demonstrates the practical utility of the test. To ensure full reproducibility, the R code that generated our numerical results is publicly accessible online.