I have a time series of data (about 300-750 elements, depending on the sample) and a model that has some random residues. I used the Kolmogorov–Smirnov test to make sure that the normality hypothesis can't be rejected, so I assume that the residuals are normally distributed. But now I guess I should test if they are independent of each other - so that there is no autoregression? Which test should I use (preferably one that is easily implementable in java)?
$\begingroup$
$\endgroup$
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
3
You could try Runs test for randomness. I am not familiar with Java functions. I got a link related to runs test that may useful to you. JAVA NPST. Apart from this you could do durbin-watson test or Ljung-Box portmanteau test. Most important is the visual check of time series plot ( I guess you have done this already).
-
$\begingroup$ If you go with Ljung-Box, which is easy to program, you might want to check out this question and answers. $\endgroup$– jbowmanCommented Jan 3, 2012 at 17:22
-
$\begingroup$ Hi, which flavor of the runs test? Calculating the runs when the value increases and decreases? $\endgroup$– GrzenioCommented Jan 4, 2012 at 13:09
-
$\begingroup$ You need to check both. The rejection rule is $R \leq C_1$ or $R \geq C_2$, where $C_1$ and $C_2$ are critical values obtained from table (or from normal distribution for large sample) $\endgroup$– vinuxCommented Jan 4, 2012 at 14:08
$\begingroup$
$\endgroup$
Another option is a permutation test. Compute a measure of autocorrelation, then randomly permute the values and recompute the same measure of autocorrelation. Do the permute and recompute step a bunch of times (like 2,000 overall) and compare the value for the original ordering to the permuted values to do the test.
answered Jan 3, 2012 at 18:24