Initial claims of unemployment
Excepting the post-Katrina period, none of the predictions I've seen for weekly initial unemployment claims have done as well as a simple exponential moving average from previous weeks. Of course, there's a lot of noise in the numbers in general; that the moving average is such a good predictor indicates that variation in the number is rather less than the noise in it. This is part of the reason I find it really annoying to hear mention of the change in the number from one week to the next; almost all the variation is noise. If you're really desperate to extract a trend given a single new week's worth of data, you probably least badly off to watch the change in the prediction you make, from one week to the next, by using that moving average. So here's how it works.

Tomorrow's number is expected to be about 303 ¹₃ — pretty close to predictions I've seen around 305. Tomorrow the announcement will come out; to extract some kind of trend, look at the "surprise" — how much the number exceeds 303 ¹₃ (negative if it's below that), plus two thirds of the upward (negative if downward) revision in last week's number. This total surprise is typically plus or minus a few thousand; if it gets to ten thousand or so, that starts to be worth noting, though really only if it persists for a couple weeks. To get next week's estimate, add one third of that surprise back into this week's (i.e. to 303 ¹₃).

I'll edit this at some point after the announcement to demonstrate the calculation described above.

Edit: The number this week was 299 (by which I mean 299,000), or -4¹₃ from the expectation; the number for last week was revised up from 302 to 304, so add 1¹₃ for a total "surprise" of -3. Next week's estimate is 302¹₃. Thanks for playing.
- posted by dWj @ 9:39 PM