For some time now, economic data geeks have had a fun little tool to play around with — the GDPNow forecast by the Federal Reserve Bank of Atlanta of quarterly gross domestic product growth. Rather than wait three months for the official GDP data, the wonks at the Atlanta Fed broke GDP into its component parts. Tracking those as each piece was updated gave a ballpark reading of how the economy was doing in something closer to real time. The Atlanta Fed even came up with a nice marketing name for the forecasting project — “nowcasting,” which I suppose is better than the hindcasting so much of the financial and economic professional does.
Now there’s some competition. Liberty Street Economics, the blog of the Federal Reserve Bank of New York, said earlier this week that it too is getting into the nowcasting game.
The idea behind nowcasting is that we can extract information from regular data series, extrapolate in between data releases and get a real time look at GDP. Nowcasting is practically magic!
The Atlanta Fed released a white paper before GDPNow’s introduction. You can read here about the six-step approach the modelers used to construct GDPNow.
Official GDP data from the Bureau of Economic Analysis, aims to be accurate, though not very fast. It is released, updated a month later, then revised a month after that, becoming more accurate with each iteration.
The old joke seems to apply here: Do you want it fast, cheap or good? Pick two.
Consider the third quarter of 2015 — the Atlanta Fed GDPNow was looking for a 1 percent gain. The final number was double that. That’s an enormous miss.
But the miss came as no surprise to serious data wonks — the idea behind GDPNow was speed and keeping the cost structure low. Fast and cheap may be a thing, but there are inherent limitations. Trying to gauge GDP a month into a new quarter poses incredible challenges, especially if you care about getting it right.
Which brings us to the New York Fed’s version of nowcasting. It differs from the Atlanta Fed’s in subtle but important ways. At the moment, as you can see from the chart below, the New York Fed’s projections are a good deal less pessimistic about first-quarter GDP growth than the Atlanta Fed’s. Since the New York Fed just started doing this, we have no way to know if it will be any more accurate.
The New York Fed’s model puts growth at 1.1 percent for the first quarter. The Atlanta Fed, which had first-quarter growth at 0.1 percent as of April 8, yesterday raised its outlook to 0.3 percent.
I have my doubts about all of it.
Building a GDP model contains a lot of selection choices. Each one imposes biases on the final outcome. Every choice made for the sake of expediency (using preliminary data, how you weight it and extrapolating the series) changes the final outcome. It isn’t that you can’t come up with a reasonable estimate; it’s that whatever choices you make in constructing your model are going to have inherent differences from the full-blown GDP model. One measures actual data series, the other models it. The more nondata measurements you build in, the more modeling assumptions you make. This will determine both what your final numbers look like and how close they are to the actual BEA data.
As a result, the Atlanta Fed seems to get the broad directional move of GDP right, but not the actual number. Some months have been dead on, while others have had big snafus. There seems to be some randomness here.
The good news is, we now have two Fed bank nowcasting models. The competition between them might lead to more accurate forecasts.
The bad news is, given what appears to be a bullish and bearish modeling bias in the two, selective perception is likely to rule the day. Don’t like one nowcast? You can check the other one. In a free market, having more choices never hurt anyone. And if you’re a long-term investor, it almost doesn’t matter. If you’re a regular reader, you know where I’m coming from: Pay little or no attention to forecasts and stick to probabilities instead.
Originally: A Forecast About GDP Nowcasts