On models and forecasts, and how one implies the other.

The economic calendar is packed with forecast releases, and each time a new one comes out, social media bursts forth with comments denigrating forecasts.

Common themes are:  forecasting and economic modelling are different activities;  forecasting is trash-talk, modelling or analysing the present is the serious business;  forecasting requires one set of skills, modelling another (viz discussion of whether you need 'foxes' or 'hedgehogs' for them).

My response to this is to point out that there is not a good separation between modelling and forecasting.  A model implies a forecast.

An example of what prompted me to write this is this piece by John Kay, which, while not sinning, is open to a mis-reading by sinners.

Tom Sargent was fond of repeating every lecture that 'a model is a probability distribution over a sequence'.  What he meant was that if you have written down a model, it will contain in it a statement about how likely certain things [defined in your 'sequence'] like output or inflation are to take on different values at different times.  Or, if you project into the past, it will tell you how likely it was that output and inflation in the past ended up being what it was.

This statement was always couched in terms of macroeconomic time series.  But it is more general than that.  The 'sequence' could be the set of individuals in a workforce, whose behaviour you haven't measured yet, but you are trying to predict from what you have measured.

Take the Bank of England's model as an example.  This model involves a decomposition of all past output and inflation into a set of 'shocks'.   The model will tell you the contribution productivity shocks were making at any point in time in the past.  But also, and since the model will tell us that shocks take time to have their full effect, we can work out without any extra maths, assumptions, or clairvoyance, what the model tells us about the chance of output and inflation being within certain ranges out into the indefinite future.

The Monetary Policy Committee don't just do this.  They introduce off model-judgements to bend the forecast to what they think is most likely.  But this doesn't mean that forecasting is something else apart from modelling.  It means that they are averaging across models;  or modifying the model.

So, when John Kay writes:

"A bane of this economist’s life is the belief that economics is clairvoyance. I should, according to this view, be offering prognostications of what gross domestic product growth will be this year and when the central bank will raise interest rates."

I'd say:  Fair enough.  But, without any clairvoyance, your understanding of the macroeconomy [your model] means that you are implying something - even if you haven't stated it explicitly - about what would, other things equal, given what you know now, happen to growth and interest rates in the future, and when.

Or when he writes:

"It is usually easy to move the subject on to something more interesting than macroeconomic forecasting."

I'd say:  maybe so.  But your conversation partners are failing to parse the logic of economics fully if they think that you really have changed the subject away from forecasting.

John is right to say that economics is not clairvoyance.  But economics - explanations of the economy's present workings - contains within it statements about the future.

There are of course lots of differences between models [or, equivalently, between forecasts].  Models that are chosen to best fit in sample.  And those chosen to best fit out of sample.  Models with or without explicit Bayesian priors.  These are models where the probability distribution over the sequence - to indulge in the Sargent language - is bent to achieve different criteria.

Following this line of thought, it doesn't make sense to talk of modelling being interesting but forecasting boring.  Or models requiring one set of skills and forecasts another.  Or forecasts, but not models, being trash.

Although I admit that it would be a bit of a turn off to try to pitch a column about probability distributions over sequences bent to achieve one criterion needing foxes, while those bent to meet another needed hedgehogs.