I know that it is a long shoot but I have an econometrics problem need your help. I model the annual industrial production (which is monthly accumulated (t, t+12) on some variables (business and stock market variables also accumulated for the last year( t-12, t). My objective is to investigate if this factors can predict the industrial production. I use both time series OLS and GARCH models. And adjusted for the auto correlation using Newey-west with Ols and Bollereseve and Wooldridge (the case of GARCH(1,1). Both of them are correct the standard errors. However my results is different in these two models OLS give significant results while GARCH not. My first question , do you have explain for this variation in the results. My second question, in all cases the serial correlation is still found after the correction for both OLS and GARCH is this normal? (even I try inclusion of the first lag of the dependent variable in the model, however including this lag destroy the significance of other variables.RegardsM.Elgammal

Can any one answer my questions. I realy need help. ThanksM. Elgammal

1. Of course the standard errors of OLS and Garch will be different. The OLS standard errors assume homoscedastic normal residuals. You can see the impact by simulating Garch and estimating with OLS2. You should write down your model so we can understand what you do. As you describe it, it appears that you use overlapping 12-month intervals, which you fit on exogenous variables using OLS/Garch. This will give wrong standard errors [Campbell-Lo-MacKinley have that somewhere in their book]. Serial correlation a'la MA can come through that.If you use non-overlapping intervals, you should check if your series are I(0) first. I get the feeling that they are not, or that the lagged industrial production is more relevant than your other explanatory variables. Saying that including the lag destroys the significance is irrelevant: the fact that you don't like it doesn't mean it's not better.If you include AR(1) and still get serial correlation in residuals, I would blame some temporal aggregation somewhere [like overlapping intervals]Hope that this helpsKyriakos

thanks for your reply my model as following dip(t, t+12)= a+B1 mkt(t,t-12)+ b2 def(t,t-12) + b3 Value premium(t, t-12)+........+e(t, t+12)I just replicte the work of liew and vassalou (2000) but monthly rather than quarterly. and following them I used the newey-west consistent coeffecient covariance to overcome the problem of serial correlation in the OLS ( Eviews option) however after using this option i still find auto correlation in the residuals. does this method remove the autocorrelation or only corect the bais in the standared error. For GARCH I use Bollerslev and Wooldridge and I still have serial correlation. Yes even I used AR(1) I can find serial correlation.M.Elgammal

Try to model first differences. It can overcome some problems. Try it and let me know

It ia already in the first difference

AFAI remember Newey-West corrects for serial correlation, it does not remove it. So you would still find it in the residuals.But:1. Your serial correlation comes from a specific source, which is the overlapping interval. Since you know its source I would prefer to tackle it more explicitly [as in Campbell-Lo-MacKinley] rather than using a generic method like NW2. NW is asymptotic, and would work better for large samples and relatively low serial correlation. But your overlapping intervals which are 5/6ths of each observation suggests to me that it is not sufficient. Plus non Gaussianity etc. Why don't you produce Monte Carlo/bootstrapped standard errors?3. Why don't you put an AR(1) term to account for this serial correlation? I am sure you can afford losing one observation.Kyriakos