EViews.com

Hi,

I have a task where I'll be modelling outflow in a savings product with respect to a number of explanatory variables through time. I went with an OLS-ARMA approach. There is autocorrelation in the errors, an AR(1) term was able to account for it. Furthermore, a number of these explanatory variables are lagged. I decided to put it in EViews to see what results would be like quickly, before I tried out-of-sample forecasting in MATLAB.

The results that I was seeing in EViews, with respect to the model itself and some dynamic forecasts looked promising. However, when I decided to read up more before trying to program some forecasts in MATLAB, it quickly became clear that these types of models (OLS with ARMA errors, particularly with lagged explanatory variables) are very tricky to do right. Now I tried to make sure that my variables were not correlated, but I assume there is a whole lot more I am missing. I would like to know is making a model like this feasible for someone not that experienced like me, and are the results I would have obtained likely to be correct or are they probably misleading and wrong?

Why not do the out of sample forecasts in EViews, and see if they look satisfactory?

Well I really need help on this. I cannot figure out how to do out-of-sample forecasts. By an out-of-sample forecast I mean not using any observations past the sample period, and instead specifying my own forecasts for the explanatory variables (for the 3-month forecasts I want to do, this is only necessary for one variable, as all the other variables are 'leading', or use lagged values so the forecasts for these values exist in the sample period).

I tried using the dynamic forecast in EViews, and at first the results looked really good, but then I discovered that it was using out-of-sample observations for the explanatory variables. This is not what I want.

Try this. In the out-of-sample period replace the values of the explanatory variables with your own forecasts for those variables. Since it's out-of-sample, this won't affect the estimation. This way EViews will pick up the values you want.