which shows that the classical runs test is based on a first-order (sign) autocorrelation quantity, hence
might fail to detect higher-order dependence. In order to improve on this, Dufour, Hallin and Mizera [37]
propose generalized runs tests for randomness, that are based on higher-order (sign) autocorrelations. Their
contribution can be considered as an important step towards a general methodology for the analysis of time
series with nonhomogeneous innovations (the runs test for randomness does not assume that observations are
identically distributed, but only that they share the same population median).
In the late 1980’s, runs have also proved useful for the problem of testing symmetry of a univariate
(absolutely continuous) distribution, as is shown by Cohen and Menjoge [38] and McWilliams [39]; both works
independently propose a common runs test that is usually referred to as McWilliams test in the literature.
Assuming that a sample of i.i.d. observations X1, . . . ,Xn is available, McWilliams’ test rejects the null that
the parent distribution is symmetric about a fixed centre (, say) for small values of