What is DfmSimulation.java actually doing?

[Liedo]

In the simulation tab, the user has the option to specify the dates in which the model has to be estimated (let's call them estimation dates):

The sofware (in particular DfmSimulation.java) is supposed to sequentially generate all the information sets that become available every time there is a new data release (let's refer to those time events as publication dates). The information regarding data releases is also specified by the user in the model tab:

What happens inside the DfmSimulation.java is that the method "mustBeEstimated" determines whether each one of those information sets will trigger a new estimation or not: the method "isNotBefore()" checks whether each publication date follows or coincides with an "estimation date", in which case the model is estimated with the information set available.

Thus, the estimation dates trigger the estimation of the model. However, they do not necesarily coincide with the publication dates, so what actually happens is that the model will be re-estimated each time a publication date follows an estimation date. During the simulation, when you click to see the details on the task being run (see picture below), you obtain the confirmation that data for a given information set has been generated.

What happens if one sets a very long evaluation sample and the publication dates are before the estimation dates ? In that case, the model will not be re-estimated until an information set available after the estimation date has been reached. So all the forecasts conditional on the recursively generated information sets will be based on the model parameters that are currently saved in the workspace. Bottom line: users interested in simulating out of sample forecasts from the model should take into account:

• The real-time data availability, which is given by the publication delay of each time series. The forecasts are sequentially generated conditional on the information set available at each publication date
• The fact that the model can be re-estimated (for example, once a year, or once every two years) . Those estimation dates should be decided by the user.