I want to check auto-correlation structure in a residual series. I know I can use DW test. But for that series should be normal distributed. My residual series is not normally distributed. Tell me what other options are available.
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
Use the Ljung-Box test https://en.wikipedia.org/wiki/Ljung%E2%80%93Box_test to test the autocorrelations.
If there are significant autocorrelations you might investigate the following
1) add structure to your model possibly including segmenting the data to deal with changes in model parameters over time
2)Incorporating Weighted Estimation to deal with model error variance that changes deterministically at different points in time.
3)Introduce dummy predictors to deal with pulses/seasonal pulses/step shifts or local time trends that may reflect latent deterministic structure in the model errors.
4) possibly add additional arima structure as suggested by the acf currently under analysis
$\begingroup$
$\endgroup$
Fit a regression between residual value and its lag and check out its t stats.
model<-lm(residual~lag(residual))
