Goldman Muses On Snowfall; Elaborates On -100,000 Preliminary NFP Estimate

Tyler Durden's picture

When Goldman Sachs writes extended treatises on the impact of snowfall and makes regression analyses of NESIS snowfall scores with payroll impacts, you know Goldman economists i) take their cure from Larry Summers' public caveats, ii) have an extended sense of humor, and iii) have way too much free time on their hands. We share the following GS essay on the second derivative impact of snowfall on the economy, but here is the punchline: "Our models suggest a small decline in payrolls without incorporating the special impact of the snowstorms, or the boost from temporary Census hiring.  Adding approximately 30,000 Census hires suggests unchanged or slightly positive payrolls; subtracting the snowstorm effects suggests a payroll number in the -50,000 to -100,000 range. Our forecast of a decline of 100,000 payroll jobs assumes an impact at the higher end of this range.  Despite this, we are inclined to think the risk remains on the side of a still bigger impact from the snowstorm itself, for two reasons:  1) the larger effects observed in as-first-reported data suggest that lags in reporting itself could be part of a snowstorm’s effect, 2) both the high-impact January 1996 snowstorm and the February 2010 storm hit slightly earlier in the month than others, and this difference in timing might be important in terms of the impact on new hires added to payrolls in the survey week." Since Larry Summers, who one may venture has a pretty good advance look at the NFP #, has warned snowfall will not be "additive" to the NFP per se, any substantial downside surprise will merely be attributed to the vagaries of mother nature, which has put global warming on hold for the time being.

From Goldman Sachs:

The Snowstorms' Impact on February Payrolls (Tilton)

February was an unusually wintry month in the eastern United States, with three significant snowstorms, two of which ranked among the highest-impact storms of the last half-century.  Financial markets are focused on how this may affect the upcoming economic numbers, particularly Friday’s payroll report.

Teasing out the impact of snowstorms on the employment news is a complex endeavor, because no two storms are the same in timing or magnitude.  Both extensive snow and colder-than-usual weather can delay hiring enough to have substantial impact on payrolls, more than a hundred thousand jobs on a few occasions.   Our forecast of a decline of 100,000 payroll jobs assumes the February snowstorms will delay at least this many net hires.  Assuming a return to more normal weather, March payrolls should post a substantial rebound. 

February was an unusually wintry month in the eastern United States.  Three significant snowstorms affected a large part of the East Coast population, with the main snow periods on February 4-7, 9-11, and 25-26. The first two of these each ranked among the 25 highest-impact Northeastern snowstorms of the last half-century, according to the National Oceanic and Atmospheric Administration (NOAA).

Weather has the potential to affect a wide variety of economic variables, including retail sales, housing activity, and even industrial production, as we have shown in prior research (see “What’s With the Weather?”, GS US Economics Analyst 07/02, January 12, 2007.)  However, the main focus of financial markets in the near term is Friday’s employment report—the nonfarm payroll number in particular.  Numerous journalists and economists have offered up estimates of the likely impact of the storms.  In the comment that follows, we explain the basis for a snowstorm effect on payrolls and gauge how large it might be this time.

In theory, the main impact of a snowstorm should be a delay in hiring (and possibly firing) activity.  Suppose business shuts down during a period (call it month 1) when payrolls would normally be expanding, and this shutdown delays some hiring into month 2.  If employment would otherwise have been on some steady underlying trend, the effect of the snowstorm will be to push payroll growth below that trend in month 1, with a sharp rebound above trend growth likely in month 2 as employers complete unfinished hiring from the month before.

A quick back-of-the envelope calculation illustrates that the impact could in theory amount to several hundred thousand jobs.  In recent expansions the typical February non-seasonally adjusted net job gain has been around 700,000-800,000 jobs; monthly gross job gains are probably around three times this amount (although the exact number is unclear as the data are only available quarterly).  If a major snowstorm shut down half of United States businesses’ hiring and firing activity for half of the payroll period (i.e roughly two weeks), then the impact would presumably be ½ x ½ x 700-800k, or 175,000-200,000.  If for any reason the hiring process was more susceptible to disruption or delays than firing, then the impact could be as much as three times as large.  Of course, the assumption of a total shutdown in hiring for half the country seems too extreme, but these calculations give a sense of the potential upper bound of impact.

Of course, bad weather might also keep existing employees from making it to work.  This would cause a reduction in hours worked and, for employees paid hourly, a reduction in earnings as well.  Some employees conceivably might not work at all during the payroll period.  However, the absence of these workers should not have any impact on the payroll number reported by the Labor Department.  As we understand it, all individuals listed on company payrolls are counted as employed regardless of how many hours (even zero) they worked during the payroll period.  So, difficulty in commuting or inability to work due to the weather should not have a direct impact on the payroll count.

Therefore, to gauge the impact of the storms on employment we need to focus on how much of an impact they have on net hiring.  The most straightforward way to go about this is simply to look at what happened after major snowstorms in the past.  Not including February’s storms, the NOAA reports eight “major” snowstorms within the past two decades.  (See http://www.ncdc.noaa.gov/snow-and-ice/nesis.php?sort=nesis_desc#rankings for a list.)  As luck would have it, most occurred during or just before the payroll survey week (the week containing the 12th day of the month), and therefore might have had an impact on payrolls.  We exclude two storms that occurred shortly after the payroll survey week; in each case, there were at least three weeks to go before the next payroll survey, so it seems likely that employers would have had time to make up most delays by that point.

The remaining six episodes happen to include the three highest-impact snow events as calculated by the NOAA’s Northeast Snowfall Impact Scale (NESIS): the “storm of the century” in March 1993, and the severe storms of January 1996 and February 2003.  The NESIS is a population-weighted metric of snowfall designed to “give an indication of a storm’s societal impacts,” so it is a very useful metric for our purposes.

As a rough first estimate of the impact of a storm, we simply take the change in payrolls during the month of the storm versus the average change over the previous three months.  We can then plot this payroll impact versus the NESIS score of the storm to gauge the relationship between storm intensity and employment impact.  In the exhibit below, the points plot the storms and their respective impact; the dotted line is the regression line through these points.  The storms range from “major” storms with a NESIS scale of 4-6 to the 13.2 figure of the March 1993 storm.  The deviation of payrolls from their prior three-month trend (we use as-first-reported changes here) ranges from about +75,000 to -325,000. 

Exhibit 1: Major Snowstorms and Payroll Deviations from Trend

 

The shaded area indicates the likely NESIS score of the combined storms on February 4-7 and February 9-11.  The preliminary scores of these storms are 4.3 and 3.9 respectively, or 8.2 together.  However, the construction of the NESIS scale is nonlinear; if the underlying data from the two storms were processed together the result could be somewhat higher (we have no way of doing the calculation ourselves).  One might also—but probably less plausibly—argue the joint score should be lower than the sum given the brief respite for cleanup between them.  The dotted regression line between the points suggests an impact of 120,000-200,000 on February payrolls.  Note, however, that the three major storms all had an apparent impact at or above the high side of this range, suggesting the potential for an even larger hit in February.

To cross-check this result, we added a historical “dummy variable” for snowstorms to a basic payroll forecasting model including other weather-related variables, estimated over the past two decades.  The dummy variable is significant (with a t-statistic of 2.8) and the fit of the overall model is reasonable, with a standard error of about 100,000.  Two interesting results emerge.  First, including a variable for snowstorms does not meaningfully diminish the impact of temperature variations on payroll growth; these remain highly significant in all formulations.  (For those unfamiliar with our prior work on the topic, warmer-than-usual winter temperatures provide a boost to seasonally adjusted payrolls, so a month-to-month change in temperatures that exceeds normal seasonal changes tends to temporarily increase payroll growth.  The corresponding relationship holds for cooler than normal temperatures.)  

Second, some experimentation shows that essentially all of the explanatory power of the snowstorm variable comes from the January 1996 episode.  When using as-first-reported payroll data, a simple dummy variable for this snowstorm by itself provides more explanatory variable than an indicator incorporating all six episodes.  Other indicators corroborate the outsized impact of the 1996 storm; for example, the number of people who missed work for weather-related reasons that month was nearly three times as large as in any other month in the past quarter-century.  The payroll impact of this episode is (slightly) less of an outlier in fully revised rather than as-first-reported data, suggesting that lags in reporting might themselves have been an issue.

Exhibit 2 below incorporates the knowledge from our model, assessing the snowstorm impact on payrolls with fully revised data, after removing the estimated impact of temperature changes.  The logic here is that snowstorms sometimes—but not always, and not this February—come along with much colder than usual weather, which can intensify their impact on employment.  The pattern, shown by the solid diamonds, is now quite striking—a tightly clustered group of points around a line that suggests a substantially smaller impact from the snow itself, perhaps 50,000 to 100,000 on the payroll report.

Exhibit 2: Snowstorm Impact on Revised Payrolls, Excluding Temperature Effects

 

So, what does this mean for our payroll forecast?  Our models suggest a small decline in payrolls without incorporating the special impact of the snowstorms, or the boost from temporary Census hiring.  (Our standard models do include the effects of temperature changes, which actually should help offset rather than intensify the snowstorm impact.  Since it was substantially colder than usual in the weeks before the January payroll survey, and roughly normal in the weeks up to the February payroll survey, we think January payrolls probably already incorporated a negative weather effect—albeit from temperature rather than snow—which otherwise would have seen a positive “payback” in February.)  Adding approximately 30,000 Census hires suggests unchanged or slightly positive payrolls; subtracting the snowstorm effects suggests a payroll number in the -50,000 to -100,000 range. 

Our forecast of a decline of 100,000 payroll jobs assumes an impact at the higher end of this range.  Despite this, we are inclined to think the risk remains on the side of a still bigger impact from the snowstorm itself, for two reasons:  1) the larger effects observed in as-first-reported data suggest that lags in reporting itself could be part of a snowstorm’s effect, 2) both the high-impact January 1996 snowstorm and the February 2010 storm hit slightly earlier in the month than others, and this difference in timing might be important in terms of the impact on new hires added to payrolls in the survey week.

Wednesday’s ADP Employment Report may provide a useful clue as to the weather effect on February payrolls.  We normally give it little weight because of its spotty track record in forecasting real-time payroll changes.  However, a simple model incorporating temperature effects and the ADP variable suggests that the ADP data are affected by unseasonal variation in temperature, so it’s not a stretch to think they would be affected by other forms of inclement weather.  If the storm did have an exceptionally large impact on hiring, the ADP report should at least hint in that direction.

Andrew Tilton