Evan Soltas
Aug 19, 2012

The Model


So today is a big day. Today, I'm launching what will become a regular feature on this blog, the real-time macro forecasting model. Eventually, I'll integrate it into the righthand column, just as Nate Silver does with his "FiveThirtyEight Forecast." (I have yet to come up with a good name. I'm open to suggestions in the comments.) The aim, in fact, is to build a macroeconomic equivalent of FiveThirtyEight.

The idea is that the macro community focuses on a few "big numbers": quarterly real GDP, unemployment, nonfarm payrolls, and CPI/PCE inflation (core and headline). And we get all of these other indicators during the interim, ranging from stock indices to initial claims for unemployment.

I've built a real-time predictive model for quarterly annualized growth in real GDP which I'll update daily, very much as Silver does with his model. As a cautionary note, it is in the very early stages, and the prediction intervals for GDP are very wide -- a root-mean-square error of 2.88 projecting the model back from 1948 through 2012, although it is roughly twice as accurate post-1980 -- but I intend on making it more accurate and adding more functionality, including a forecasting side to it, over the coming weeks.

The idea here is if stock markets drop 5 percent, or if nonfarm payrolls surprise positively, I will have a new forecast from my model which reflects that change -- and I'll be able to tell you what we can expect across a panel of macroeconomic indicators and with what predictive confidence.

Here's a very quick summary of how it works. My model takes as inputs a number of weekly and monthly data sets -- percentage change in payroll employment, in industrial production, in real personal consumption expenditures, in initial claims for unemployment benefits, and in the S&P 500 index, and levels of the ISM survey, the Philadelphia Fed leading index, and the St. Louis financial stress index. It completes a univariate regression for each of these variables with the quarterly change in real GDP. For those of you who aren't math types, what this means is that I figure out, for any given value of a variable, what the historical matching value of GDP growth should be. Then I weight each variable's GDP projection according to the coefficient of determination r^2 of that variable as a fraction of all variables. Again, for the non-math people, what this means in English is that I assess how accurate the relationship between all my variables and GDP growth has been in the past, and the stronger the relationship, the more weight I put on the GDP growth projection from that variable.

My model replicates a similar program developed by the Treasury Department, according to this academic report here, although I plan on moving well beyond what is described here over time.

Although it is not all too accurate right now, it has some appreciable features. First, it doesn't miss any movements in GDP growth which happen over multiple quarters -- what it appears to be missing is the "noise" in the quarterly growth series, as my model looks much more like the year-over-year graph for GDP. This "conservatism" in the projection is happening because my model is constructed to hedge its bets. For my model to move downward, it takes coordination across the indicators. The result is that my model comes very, very close to matching up with GDP growth over longer intervals, but it isn't so good at capturing individual bounciness in the quarters. By my inspection, my forecast captures in real time every recession, every boom, and every period of stable growth since 1948 -- which, given the real-time nature of the model, is not all that bad.

Much, much more detail on this model will be forthcoming over the next few days. I hope this will be a fun project for me and for my readers -- I'll be constructing the forecasting model under your watch, explaining the additions and changes as I go.

Right now, I only have the GDP side of the model in any presentable stage of development, but I'm working on the unemployment forecast model right now, and it is showing real promise. I plan on forecasting inflation as well in the first full version of the model.

And incidentally, I plugged in the data we have thus far for the third quarter of 2012: my model anticipates 0.5 percent quarterly annualized growth in real GDP.