When Economics Failed

The tools the profession uses to forecast recessions still are a long way from being up to the task.

At least we know when hurricanes will hit.

Photographer: Mark Wilson/getty images

When people come up to me and declare that economists are charlatans, they usually mention how economists failed to predict the Great Recession. This is true. No mainstream macroeconomic model of the day managed to anticipate that the largest, longest economic downturn since the Great Depression was imminent, until it was already under way.

Macroeconomists typically respond that forecasting isn’t their job. The economy has all kinds of things going on at any given time, they say -- too much randomness and noise to allow a reliable forecast. The best they can do, macroeconomists will say, is to predict the effects of specific policies.

This defense is weak. If the economy is dominated by random noise, that noise will also permeate the data that is used to validate macroeconomic models. If forecasting is impossible, then picking the right policy-evaluation model will also be impossible. Also, the inability to forecast is often a clue that a model is just plain wrong.

So forecasting actually is important, and macroeconomic models are bad at it. There’s even a whole line of research dedicated to showing just how bad even the most advanced models are at predicting things like output and inflation.

My favorite paper in this literature is by Refet Gurkaynak, Burcin Kisacikoglu, and Barbara Rossi. In 2013, they took some of the most advanced modern macroeconomic models then available -- called DSGE, for dynamic stochastic general equilibrium -- and tested them against some very simple models called autoregressive (AR) models. DSGE models are very big, complicated things -- they rely on tons of assumptions about consumers, companies, technology and the government. An AR model is just a single equation -- just about the simplest forecasting model you can possibly imagine. In principle, the sheer complexity of the DSGE models should make them better at forecasting the macroeconomy.

But as it turned out, the super-simple AR models won as many comparisons as they lost. That means that top-of-the-line macroeconomic models, for all their complexity and difficulty, were close to useless for forecasting, at least as recently as four years ago.

In the intervening years, macroeconomists have been working hard to augment DSGE models with new elements that make them better at fitting past data. But macroeconomists are also increasingly experimenting with forecasting models that don’t make so many assumptions about economic relationships. In an interesting and highly sophisticated new paper, economists Andrea Carriero, Ana Galvao and George Kapetanios examined a wide class of macro forecasting models to understand how best to predict the economy.

Carriero et al. don’t just test a wide array of models. They also look at multiple developed countries -- something that relatively few macroeconomists, especially those based in the U.S., tend to do. The authors focus on predicting just two things -- output and inflation.

Their findings are discouraging. There is no type of model that is clearly the best. Different models work better at different time horizons, for different historical time periods, and for different variables. That implies that forecasting successes are mostly due to luck. Economists’ favorite models, such as DSGE, seem to be better at forecasting inflation at a few time horizons, but even this modest victory is probably accidental.

The good old super-simple AR model, as in past studies, wins about as much as it loses, despite going up against fearsomely sophisticated competitors. As the forecasting horizons get longer, the simplest models start to do just as well or better than the others.

What’s more, adding more than a few predictors doesn’t really improve forecasts. That’s discouraging, because it means that macroeconomic data just doesn’t have much useful information in it. The old Wall Street joke that “all financially useful data costs money, which is why macro data is free” seems to hold true.

All this adds up to a pessimistic conclusion -- recessions just aren’t very predictable from economic data. The reason economists couldn’t foresee the Great Recession isn’t that they’re blinkered or closed-minded or arrogant or stupid -- it’s because no one could predict it, at least not with the kind of macroeconomic data that now exist.

That in turn implies that much of macroeconomics itself, as currently practiced, is a dead-end pursuit.

But I’m not ready to subscribe to that level of pessimism just yet. Although recessions are difficult or impossible to spot in the near term, there might still be ways to see some warning signs developing a few years out. Research by some economists at the Federal Reserve Board has found that certain characteristics of debt markets -- credit spreads and the share of junk bonds in total debt -- can give warnings of recessions two years in the future.

So the future of macroeconomic forecasting -- and of macroeconomics itself -- might lie in a different direction than the one most researchers have been pursuing. Instead of focusing on consumption, investment and other easily measurable things, economists might try thinking more about subtle, long-term buildups of problems in financial markets.

This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

    To contact the author of this story:
    Noah Smith at

    To contact the editor responsible for this story:
    James Greiff at

    Before it's here, it's on the Bloomberg Terminal.