A Funny Forecasting Story

Last night, I knew I was on tap to talk about Wisconsin GDP numbers today (Wisconsin Public Radio report here). I wondered what a naive forecast would imply, so I regressed first difference of log Wisconsin GDP on first difference of log US GDP and a lagged Wisconsin GDP first difference, to obtain this forecast…(red square)

Figure 1: Wisconsin GDP, latest release (black), and previous vintage (gray-red), and forecast based on previous vintage (red square) and 95% confidence interval (gray +), all in millions of Ch.2012$, SAAR. BEA, and author’s calculations.

The 2020Q2 error was -0.0015 (-0.15%), or 419 million Ch.2012$ SAAR — barely anything.

But, the annual benchmark revisions changed upward the level of the 2018-19 GDP by about 0.8% (log terms), so it’s surprising that the forecast was actually so close. I guess it goes to show that one can forecast the outcome just right by accident.

More discussion of the specifics of the Wisconsin GDP numbers in this post.


Econometric details:

The forecast is based on this regression, estimated over the 2005Q3-2020Q1 period. Let y be real GDP.

Δywit = -0.0016 0.1312Δywit-1 + 1.0710Δyust

Adj-R2=0.65, SER = 0.0047, N = 59. Significance at 5% msl using HAC robust standard errors denoted by bold face. Q-stat 8 lags, 8.451 [p-value = 0.391]