As a follow up to Dave’s recent post Cowen and Krugman and Volcker, Oh My! I’d like to present this little factiod about the Philly Fed and recent prognostications (via Zero Hedge).
And here is Wall Street’s economist brigade again proving they are worth every penny based on their predictive skills: we just had an 8 standard deviation miss to consensus.
If you look at the chart you can see a nice Gaussian curve on the right centered around 2% and the real number far off to the left at -30.7%. Hot air explains with a quote from Business Week:
The Federal Reserve Bank of Philadelphia’s general economic index plunged to minus 30.7 this month, the lowest since March 2009, from 3.2 in July. The August gauge exceeded the most pessimistic projection in a Bloomberg News survey in which the median estimate was 2.
Just to give you some idea. Six standard deviations is about one chance in a billion. In mathematical notation a one with 9 zeros following it. 1E9. Eight standard deviations is only about a million times less likely than one chance in a billion. The actual number being on the order of 2E15. You can learn more about the probability numbers at a place that is discussing the 2008 meltdown [pdf].
All very interesting. But what does it tell us? Simply that the guys running the show haven’t a clue. They aren’t RCH close (a Navy term – look it up). They aren’t in the right ballpark. Heck. They aren’t even in the same universe. And the economy? I’d say Dave has it right. A double dip is very likely.
Comments
5 responses to “Eight Standard Deviations Off”
Only a double dip? Maybe… if we’re lucky.
I once had a girlfriend, way back in the hazy days of high school, possessed of RCH.
Now I’m not one of these fancy-schmancy economists, I just build Bayesian models for sciencey stuff like genetics. However, I know that there’s a huge difference between the standard deviation of a bunch of point estimates and a pooled standard deviation.
Each of the economists presumably should be providing approximate variances on their estimates, the person building the consensus should then visit Wikipedia:
http://en.wikipedia.org/wiki/Pooled_variance
Alan,
You work with what you have. To the best of your ability. And not only that, how do you rate chances? Per hour? Per day? Week? Month? Year?
So four standard deviations or eight? They were way off.
Ask yourself why? Bad data? You could get some real time that would show changes of magnitude of that order by just watching truck toll collections daily. Or possibly some other metric for a sanity check.
Electrical power use if you could correct for weather.
Something.
So if it is not bad data then you are left with incompetence. Because even if the data is “adjusted” there are limits to how much adjusting you can do before it gets noticed. You can fudge 5 points. But 32.7? People notice.
The improbability only proves that the errors are not normally distributed, not that it’s a bad model.