OK, so now I have the unemployment forecasting model in a good place. Let's talk about how it works and how well it works.
At least for now, it is really simple, working from only two indicators: the prior month's unemployment rate and percent change in initial claims. That means updates to my real-time projection will come out once a week; I will be looking for additional inputs to make this daily.
The model's single most important component is the autocorrelation between the unemployment rate in one month and in the prior month. And that should make sense: the unemployment rate doesn't go from 4 percent one month to 8 the next, such that one month doesn't impact the next. No, rather it moves slowly upwards and downwards, often tick by tick. Using the lag as a base, the model pushes unemployment up or down on the basis of initial claims data.
I also make some small technical adjustments for rounding and to eliminate serially correlated error. It's worth noting that these changes actually decrease the overall goodness-of-fit of the model, albeit slightly. In effect, some of the error is spread through the rest of my predictions as I adjust for serial correlation. This is preferable, however, as the lagged aspect of the model will mean that without such adjustment it won't "notice" consistent conditions of growth or recession. Otherwise, it would have always undershot the unemployment rate when it was rising and overshot it as it fell.
My model explains 95.2 percent of all variability in the unemployment rate from 1948 to 2012. To consider its accuracy another way, 23.6 percent of its guesses over the 774-month interval get it exactly right; 66.5 percent of the time, its guess is within 0.1 percentage point above or below the actual unemployment rate. To boot, over the last decade, its "within-one-tick" accuracy rate has been 75.0 percent.
Looking at prediction error more technically, the root-mean-square error (RMSE) of my model's guesses versus what actually happens is 0.207. That error is pushed upwards by some big misses early in the sample, and if one excludes the first decade or so of data points, RMSE goes to 0.15.
This all means that looking at a graph of my prediction versus the actual outcome, as I displayed for my less-accurate quarterly real GDP growth forecast, would not be a very informative exercise -- it is basically impossible to distinguish one from the other. But here you go anyway. You may have to zoom in.
It predicts an 8.2 percent unemployment rate for August, with an upward bias -- i.e. before rounding, the model expects an 8.24 percent unemployment rate. That result is largely consistent with good news over the last few weeks in initial claims data.
I would hugely appreciate suggestions, ideas, comments, or other information which you think might improve the model.