One possible strategy is to classify as random any sequence presenting a run whose length is large enough—
is strictly larger than four, say. This is justified by the fact that, when asked to produce artificially such a
random sequence, children are usually afraid of writing down long runs, while long runs actually appear quite
often (the probability that the length of the longest run is strictly larger than four is above 95%; see Figure 1
for an illustration). This explains why this classification rule tends to work well in practice.
From this teaching experiment, one may guess that runs are of interest for statistical inference. Run-based
tests of hypotheses—or simply runs tests—are based on the size of the longest run or, more often, on the
number of runs in a sequence. Runs tests figure among the oldest nonparametric procedures, as evidenced by,
e.g., [2]. They exhibit many advantages : they remain valid (in the sense that they meet the level constraint)
under very mild assumptions, they are often distribution-free, they are very robust against possible outliers,
they do not require any moment conditions, and they are easy to calculate and implement. Their main
drawback, at least in the standard univariate case, is that they are poorly efficient compared to parametric
Gaussian procedures, unless observations originate from a distribution with very heavy tails.
One possible strategy is to classify as random any sequence presenting a run whose length is large enough—is strictly larger than four, say. This is justified by the fact that, when asked to produce artificially such arandom sequence, children are usually afraid of writing down long runs, while long runs actually appear quiteoften (the probability that the length of the longest run is strictly larger than four is above 95%; see Figure 1for an illustration). This explains why this classification rule tends to work well in practice.From this teaching experiment, one may guess that runs are of interest for statistical inference. Run-basedtests of hypotheses—or simply runs tests—are based on the size of the longest run or, more often, on thenumber of runs in a sequence. Runs tests figure among the oldest nonparametric procedures, as evidenced by,e.g., [2]. They exhibit many advantages : they remain valid (in the sense that they meet the level constraint)under very mild assumptions, they are often distribution-free, they are very robust against possible outliers,they do not require any moment conditions, and they are easy to calculate and implement. Their maindrawback, at least in the standard univariate case, is that they are poorly efficient compared to parametricGaussian procedures, unless observations originate from a distribution with very heavy tails.
การแปล กรุณารอสักครู่..
