Wednesday, September 27, 2006
Noise vs. Signal in Economic News: New Home Sales Edition
by Tom Bozzo
As a rule, though especially for reports of esoteric economic activity indexes, hard-to-measure stuff such as consumer confidence, I often wish for some discussion of typical variability of the data and generally some "so what" discussion. (*)
In this case, the difference between +4.1% and -3% in annualized home sales has a reasonable claim to practical significance. But statistically, helpful information provided by the Census Bureau with the release suggests that the month-to-month changes in new home sales are too noisy to be worthy of much attention — the margin of error on that 4.1 percent is +/- 15.5%. The forecasters almost might as well throw darts for their predictions of the monthly change.
In early stories from AP and Reuters (the former appearing on the NYT website), only Reuters noted the result that annualized sales have plunged 17.4% year over year (+/- 11% — so it's both qualitatively and statistically significant). There is no real joy to be found in the large-ish margin of error on that figure, as year-to-date new home sales are down 15.7% (+/- a modest 4.4%). This is an adverse turnaround of more than 20 percentage points from August 2005.
Nevertheless, this is reported as part of a good economic news stream. Clap louder?
Update: The actual NYT story correctly suggests that there is no "there" there in the headline new home sales figures, but the closest thing to an accurate quote mined from the economic commentariat is a characterization of this month's data as a "dead cat bounce." If any of the experts claimed that the positive surprise could have been statistical noise, it went unreported in the paper of record.
--------------------------
(*) After all, when it's reported that the Richmond Fed's Central Atlantic manufacturing index rises from a "tepid" 3 to 9, they're obviously not saying that conditions are 300% better. But the Richmond Fed provides, by BLS or Census standards, minimal discussion of survey methodology, sampling error, and potential sources of nonsampling errors.
The Census Bureau this morning surprised the economic forecasting community by reporting a month-to-month rise in new home sales from July to August — estimated sales rose 4.1 percent versus a reported expected decline of 3 percent.
As a rule, though especially for reports of esoteric economic activity indexes, hard-to-measure stuff such as consumer confidence, I often wish for some discussion of typical variability of the data and generally some "so what" discussion. (*)
In this case, the difference between +4.1% and -3% in annualized home sales has a reasonable claim to practical significance. But statistically, helpful information provided by the Census Bureau with the release suggests that the month-to-month changes in new home sales are too noisy to be worthy of much attention — the margin of error on that 4.1 percent is +/- 15.5%. The forecasters almost might as well throw darts for their predictions of the monthly change.
In early stories from AP and Reuters (the former appearing on the NYT website), only Reuters noted the result that annualized sales have plunged 17.4% year over year (+/- 11% — so it's both qualitatively and statistically significant). There is no real joy to be found in the large-ish margin of error on that figure, as year-to-date new home sales are down 15.7% (+/- a modest 4.4%). This is an adverse turnaround of more than 20 percentage points from August 2005.
Nevertheless, this is reported as part of a good economic news stream. Clap louder?
Update: The actual NYT story correctly suggests that there is no "there" there in the headline new home sales figures, but the closest thing to an accurate quote mined from the economic commentariat is a characterization of this month's data as a "dead cat bounce." If any of the experts claimed that the positive surprise could have been statistical noise, it went unreported in the paper of record.
--------------------------
(*) After all, when it's reported that the Richmond Fed's Central Atlantic manufacturing index rises from a "tepid" 3 to 9, they're obviously not saying that conditions are 300% better. But the Richmond Fed provides, by BLS or Census standards, minimal discussion of survey methodology, sampling error, and potential sources of nonsampling errors.
