VECM forecast not fit reality

[Emilie_R]

Hi everyone!

I am trying to forecast real wages in France using 2 variables : real wages and inflation (cpi). I work with their log.

The 2 variables are I(1).
I modelised a VAR model using varlagselect :
@varlagselect(det=constant,crit=aic,lags=24) 2000:1 *
# dlsalr dlipc

Then i used the cointegration test :
@JOHMLE(DET=RC,LAGS=4)
# dlsalr dlipc
(results : there is 1 cointegration relation)

That way, i obtained this model :
system(model=cointmodel1)
variables dlsalr dlipc
lags 1 to 4
det constant DUM200803 DUM200903
ect ecmeq
end(system)
estimate(residuals=vres2) 2000:1 *

The 2 dummies were added to smooth the effect of the crisis.

The issue is : when i test the ability to predict by forecasting 1 step ahead in sample, i got the graph attached to this post.
I don't know how can i improve it.

Thanks for your help!

[TomDoan]

Are your variables the inflation rate and the rate of wage changes? I'm a bit puzzled because you said you were taking logs. The series that you're forecasting certainly doesn't look like it's I(1).

[Emilie_R]

Sorry, i wasn't clear : this is real wage (hourly salary index) in log and it's differentiated once.

The variables have to be stationnary to be modelized, isn't it?
My teacher only explains the theory and it's the first time i use winrats and the first time i forecast. So, i'm kinda lost...

The graph compare the real wage (log + I(1)) and my in-sample forecast (from my VECM model). As we can see, and I don't know why, it doesn't fit the reality. The forecast errors are very high.

Thanks for your reply.

[Emilie_R]

Here are my data if it can help.
I have the datas for France and Germany (I only study France). Series are indices.
ipc is the cpi (inflation) ans salr is the real wage.

I got those datas from Bloomberg terminal.

[TomDoan]

First, you model a VECM based upon the original data, not the differences because the key (error correction) term needs the original data. See the discussion in Section 7.8 in the User's Guide.

It will be hard to analyze those because the extreme structural break in the 2008-2010 period. However, just on inspection it looks almost like those are beyond being cointegrated to actually being linked by construction (with nothing more than minor rounding errors).

francedata.png (17.61 KiB) Viewed 10081 times

[Emilie_R]

Thanks a lot Tom!

So i got everything wrong...
Thanks for your explanation. I need to model a var or vecm for a university project.
I'll try to do what i can with these series and if it's too hard, i'll change my subject.

At least, I know how to start!

[TomDoan]

Real wage will definitely not be cointegrated with nominal prices. (It's also unlikely that nominal wages would be cointegrated with nominal prices since that would imply that real wages are stationary, and there's no reason to believe that that's the case).