help with lags and autocorrelation

[selina gravette]

Hello!!! So i am trying to choose how many lags to do, and someone helped me with some code but i don't know how to interpret it..
I have daily data for stock returns for two cities

they've looked at the PACF and ACF (code is: "ac" and "pac") as well as done:
"dfuller returns, lags(12) regress"

they also put all 12 lags in the regression, like this:
"regress returns cloud temp sunhour l(1/12).returns i.month"

How do I choose the number of lags using all this? is it with whichever has a significant p>z value ?
or do i count on the pacf or acf graph and if so do i use pacf or acf?

they also did
lmalb returns, lags(5)

please help, im stuck on this!!

[Stefano Grillini]

Hi Selina,

I would first advice you to have a look at the FAQ where they say to report Stata codes using delimiters. You can do it by clicking the "#" symbol.
Your question is more about textbook theory rather than coding, so my advice would be to look up on Google/books. Said this,

- Both ACF and PACF provide you a graphical interpretation of the autocorrelation. It is mostly used for residuals. The idea behind is that autocorrelation should decay to zero quite quickly. If you see a slow decay then you have problems of autocorrelation.

- dfuller is a test of stationarity of the time-series. You might consider trying different lag lengths and see if you have a rejection of the null hypothesis of a unit root. It is more concerning the presence of a trend and/or a drift, rather than the lag length. If you have returns then you already have a differenced series (by taking the difference of subsequent log prices). In my experience, although very little, I have almost never seen a non-stationary series of returns unless for very very long time periods. REMEMBER: do not perform any linear models on non-stationary time-series.

- To assess the optimal lag length you might consider using varsoc

- Alternatively, you can try to fit a model such as AR, MA, ARMA, ARIMA and choose the optimal lag length using the AIC criterion. Your best model is the one that minimizes AIC.

Hope this helps

Stefano