Multicollinearity issues in time series regression

[Javier Perez Iglesias]

Hi there. I am doing my master thesis with an empirical study about how traders within the agricultural commodities market form their expectations, to see if they form them in an extrapolative way. The thing is that when I run a linear regression model I get a very high R squared with a very high F statistic, which are signs of multicollinearity. After testing for vif I confirm that there is multicollinearity issues in my model, but I cannot get rid of variables as I am trying to explain the expectations on i.e., corn (with a unique survey data) regressed on 10 lags of the actual prices of Corn. I have tried to center the variables but the result does not change. Any recommendations? Thanks in advance.

[Carlo Lazzaro]

Javier:
lags are highly correlated. If this your only strategy, thhere' s nothing more to comment on (from my side, at least).

[Javier Perez Iglesias]

I runned the model in first differences and mi R square has dropped down and also mean VIF now in 1.5. Should I do it with first differences or continue with model in levels and highly correlated independent variables?

[Carlo Lazzaro]

Javier:
without any results, my gut-feeling is to go D.
That said:
1) I'd consider if the predictors that you plugged in the right-habd side of your regression equation are ebough to give a fair and true view of the data generating process you're investigating;
2) I'd take a look at the "tribal rules" currently on force in your research field;
3) last but not least: what is your suprevisor's opinion about your approach?

[Javier Perez Iglesias]

My supervisor is not replying my emails
As far as I know multicollinearity produces that i cannot trust the p values, but i can trust the coefficients?
Going first differences makes my coefficients less trustable?
I will check literature related to this but there is no much literature on expectations about the cereals market.

Thank you so much,
Javier

[Carlo Lazzaro]

Javier:
1) Very bad indeed;
2) yes, you can trust the coefficients but not standard errors if they appear "weird" (vs, what you usually read) and related stuff;
3) and 4): scientific literature/more experienced colleagues is the way to go.