Conditional and Unconditional Forecasting

Sometimes I wonder how others manage to write short posts. In my earlier post about forecasting, I used an analogy with medicine to make the point that an inability to predict the future does not invalidate a science. This was not the focus of the post, so it was a single sentence, but some comments suggest I should have said more. So here is an extended version.

The level of output depends on a huge number of things: demand in the rest of the world, fiscal policy, oil prices etc. It also depends on interest rates. We can distinguish between a conditional and an unconditional forecast. An unconditional forecast says what output will be at some date. A conditional forecast says what will happen to output if interest rates, and only interest rates, change. An unconditional forecast is clearly much more difficult, because you need to get a whole host of things right. A conditional forecast is easier to get right.

Paul Krugman is rightly fond of saying that Keynesian economists got a number of things right following the recession: additional debt did not lead to higher interest rates, Quantitative Easing did not lead to hyperinflation, and austerity did reduce output. These are all conditional forecasts. If X changes, how will Y change? An unconditional forecast says what Y will be, which depends on forecasts of all the X variables that can influence Y.

We can immediately see why the failure of unconditional forecasts tells us very little about how good a model is at conditional forecasting. A macroeconomic model may be reasonably good at saying how a change in interest rates will influence output, but it can still be pretty poor at predicting what output growth will be next year because it is bad at predicting oil prices, technological progress or whatever.

This is why I use the analogy with medicine. Medicine can tell us that if we eat our 5 (or 7) a day our health will tend to be better, just as macroeconomists now believe explicit inflation targets (or something similar) help stabilise the economy. Medicine can in many cases tell us what we can do to recover more quickly from illness, just as macroeconomics can tell us we need to cut interest rates in a recession. Medicine is not a precise enough science to tell each of us how our health will change year to year, yet no one says that because it cannot make these unconditional predictions it is not a science.

This tells us why central banks will use macroeconomic models even if they did not forecast, because they want to know what impact their policy changes will have, and models give them a reasonable idea about this. This is just one reason why Lars Syll, in a post inevitably disagreeing with me, is talking nonsense when he says: “These forecasting models and the organization and persons around them do cost society billions of pounds, euros and dollars every year.” If central banks would have models anyway, then the cost of using them to forecast is probably no more than half a dozen economists at most, maybe less. Even if you double that to allow for the part time involvement of others, and also allow for the fact that economists in central banks are much better paid than most academics, you cannot get to billions!   

This also helps tell us why policymakers like to use macroeconomic models to do unconditional forecasting, even if they are no better than intelligent guesswork, but I’ll elaborate on that in a later post.