Model specification testing of time series regressions (Q1057606): Difference between revisions

Let \(\{z_ j\}\) be a vector time series process. A model specification test is suggested for this process testing the null hypothesis that the corresponding regression function has the form of the conditional expectation and so employs the maximum amount of explanatory power contained in the past of the process. This null hypothesis can be formulated as \[ E(z_ j,z_{j-1},z_{j-2},...)=F(z_{j- 1},...,z_{j-m},\nu_ 0), \] where F is a regression function depending on a parameter \(\nu_ 0\) (i.e. we have a specification in the form of a general time series regression). In the case that the process \(\{z_ j\}\) is strictly stationary this test is consistent with respect to the alternative that the null hypothesis is false. In the non-stationary case the alternative cannot be so general to guarantee the consistency of the test. The test is demonstrated on a numerical example concerning the specification of the previous type for simulated autoregressive processes. A simplified version of the previous test for testing the hypothesis that the errors are martingale differences is also presented in the paper.