Exact and asymptotic goodness-of-fit tests based on the maximum and its location of the empirical process

TL;DR

This paper introduces a new goodness-of-fit test based on the maximum of the empirical process, providing exact and asymptotic distributions, and demonstrating improved performance over existing tests through simulations.

Contribution

It proposes a novel test statistic with an exact distribution under the null hypothesis, improving accuracy and power compared to traditional tests like Smirnov's.

Findings

The new test outperforms the Smirnov-test in simulations.

Exact distribution derivation enhances test accuracy.

Method applies to both one-sided and two-sided hypotheses.

Abstract

The supremum of the standardized empirical process is a promising statistic for testing whether the distribution function $F$ of i.i.d. real random variables is either equal to a given distribution function $F_0$ (hypothesis) or $F \ge F_0$ (one-sided alternative). Since \cite{r5} it is well-known that an affine-linear transformation of the suprema converge in distribution to the Gumbel law as the sample size tends to infinity. This enables the construction of an asymptotic level-$\alpha$ test. However, the rate of convergence is extremely slow. As a consequence the probability of the type I error is much larger than $\alpha$ even for sample sizes beyond $10.000$. Now, the standardization consists of the weight-function $1/\sqrt{F_0(x)(1-F_0(x))}$. Substituting the weight-function by a suitable random constant leads to a new test-statistic, for which we can derive the exact distribution (and the limit distribution) under the hypothesis. A comparison via a Monte-Carlo simulation shows that the new test is uniformly better than the Smirnov-test and an appropriately modified test due to \cite{r20}. Our methodology also works for the two-sided alternative $F \neq F_0$.