We can’t forecast next year’s Y unless Statistics Canada tells us next year’s X, and they can’t.

His point is that some variables are outside of the model, and to make a forecast in real time you have to forecast those variables.

Which is exactly what model proprietors do. In the traditional model, most government spending (apart from something like unemployment benefits, which depends on other variables in the economy) is forecast, based on legislation and budget proposals. The price of oil might be another variable that is projected by the model proprietor, rather than determined within the model.

Incidentally, lagged values of variables were really important, so that the ratio of known information to unknown information was actually pretty high for very-near-term forecasts.

Back in the days when macro models were evaluated (see Stephen McNees), one type of study looked at how the models would have performed if all the values of variables outside the model had been known perfectly in advance. In general, this did not make the model perform better. It could even make the model perform worse, because model proprietors were fudging their models in ways that corrected for many problems, including bad forecasts for variables outside of the model.

In other words, the problem for model-based forecasts was not that the proprietors could not forecast variables outside of the model. The problem was that, even given correct values for those variables, the errors intrinsic to the models were large.

## READER COMMENTS

## Nick Rowe

## May 31 2010 at 9:48am

Arnold: your posts on this sort of topic always make for depressing reading. Especially since I know you actually have experience with this stuff (I don’t).

So, it’s worse than I thought.

The model is Y(t)=R(X(t)).

I assumed they were putting a curvy ruler on X(t) to forecast X(t+1), then plugging that into the model to forecast Y(t+1).

But from what you are saying, I can only conclude that what they are

reallydoing is putting a curvy ruler on Y(t) to forecast Y(t+1), then using the inverse of the model to forecast X(t+1), and then pretending to work backwards again, and use the model to forecast Y(t+1). In other words, the model is only there for show. That’s the only way to make sense of the observation that they do better at forecasting Y(t+1) than they would if they actually knew X(t+1).Comments are closed.