Weather adjusting economic data

This paper proposes and implements a statistical methodology for adjusting employment data for the effects of deviations in weather from seasonal norms. This is distinct from seasonal adjustment, which controls only for the normal variation in weather across the year. We simultaneously control for both of these effects by integrating a weather adjustment step in the seasonal adjustment process. We use several indicators of weather, including temperature and snowfall. We find that weather effects can be important, shifting the monthly payroll change number by more than 100,000 in either direction. The effects are largest in the winter and early spring months and in the construction sector. A similar methodology is constructed and applied to data in the national income and product accounts (NIPA), although the manner in which NIPA data are reported makes it impossible to integrate weather and seasonal adjustments fully. M time series are affected by the weather. In the first quarter of 2014, real GDP contracted by 0.9 percent at an annualized rate. Commentators and Federal Reserve officials attributed part of the decline to an unusually cold winter and large snowstorms that hit the East Coast and the South during the quarter (Macroeconomic Advisers 2014; Yellen 2014). Similarly, the slowdown in growth in the first quarter of 2015 was widely ascribed to another exceptionally harsh winter and other transitory factors (Yellen 2015). While the effects of regular variation in 1. In November 2013, the Survey of Professional Forecasters expected a seasonally adjusted increase of 2.5 percent in 2014Q1. The original report for the quarter was 0.1 percent, later revised to -2.1 percent, and subsequently revised to -0.9 in the 2015 annual NIPA adjustments that included revisions to the seasonal adjustment process, as discussed in section III below. With a snapback rate of 4.6 percent in the second quarter, it is highly plausible that weather played a significant role in the decline. 228 Brookings Papers on Economic Activity, Fall 2015 weather within a year should, in principle, be taken care of by the seasonal adjustment procedures that are typically applied to economic data, these adjustments are explicitly not supposed to adjust for variations that are driven by deviations from the weather norms for a particular time of year. It is typically cold in February, depressing activity in some sectors, and seasonal adjustment controls for this. But seasonal adjustment does not control for whether a particular February is colder or milder than normal. Our objective in this paper is to construct and implement a methodology for estimating how the data would have appeared if weather patterns had followed their seasonal norms. Monetary policymakers view weather effects as transitory—given the long and variable lags in monetary policy, policymakers do not generally seek to respond to weather-related factors. It follows from this that the economic indicators they are provided with ought, as far as possible, to be purged of weather effects. Moreover, we argue that failing to control for abnormal weather effects distorts conventional seasonal adjustment procedures. The measurement of inflation provides a useful analogy. The Federal Reserve focuses on core inflation, excluding food and energy, rather than headline inflation. The motivation is not that food and energy are inherently less important expenditures but that fluctuations in their inflation rates are transitory. Core inflation is more persistent and forecastable, and indeed a forecast of core inflation may be the best way of predicting overall inflation (Faust and Wright 2013). In the same way, economic fluctuations caused by the weather are real, but they are transitory. We may obtain a better measure of the economy’s underlying momentum by removing the effects of abnormal weather. Economists have studied the effects of the weather on agricultural output for a long time, going back to the work of R. A. Fisher (1925). More recently, they have also used weather as an instrumental variable (see, for example, Miguel, Satyanath, and Serengeti [2004]), arguing that weather can be thought of as an exogenous driver of economic activity. Statis tical agencies sometimes judgmentally adjust extreme observations due to specific weather events before applying their seasonal adjustment procedures. Although there is a long literature on seasonal adjustment, we are aware of only a few papers on estimating the effect of unseasonal weather on 2. Even when agencies do this, their goal is just to prevent the anomalous weather from distorting seasonals, not to actually adjust the data for the effects of the weather. We discuss this in more detail later. MICHAEL BOLDIN and JONATHAN H. WRIGHT 229 macro economic aggregates. The few papers on the topic include those by Macroeconomic Advisers (2014), which regresses seasonally adjusted aggregate GDP on snowfall totals, estimating that snow reduced 2014Q1 GDP by 1.4 percentage points at an annualized rate; by Justin Bloesch and François Gourio (2014), who likewise study the relationship between weather and seasonally adjusted data; by Melissa Dell, Benjamin Jones, and Benjamin Olken (2012), who implement a cross-country study of the effects of annual temperature on annual GDP; and by Christopher Foote (2015), who studies weather effects on state-level employment data. None of these papers integrates weather adjustment into the seasonal adjustment process, however. This is what the current paper attempts to do. We focus mainly, but not exclusively, on the seasonal adjustment of the Bureau of Labor Statistics (BLS) Current Employment Statistics (CES) survey (the “establishment” survey), which includes total nonfarm payrolls. We do so because it is clearly the most widely followed monthly economic indicator, and also because it is an indicator for which researchers can approximately replicate the official seasonal adjustment process, unlike the NIPA data. We consider simultaneously adjusting these data for both seasonal effects and unseasonal weather effects. This can be quite different from ordinary seasonal adjustment, especially during the winter and early spring. Month-over-month changes in nonfarm payrolls are in several cases higher or lower by as much as 100,000 jobs when using the proposed seasonal-and-weather adjustment rather than ordinary seasonal adjustment. Using seasonal-and-weather adjustment increases the estimated pace of employment growth in the winters of 2013–14 and 2014–15. The plan for the remainder of this paper is as follows. In section I, we discuss alternative measures of unusual weather and evaluate how they relate to aggregate employment. This is intended to give us guidance on which weather indicators have an important impact on employment data. In section II, we describe seasonal adjustment in the CES and discuss how adjustment for unusual weather effects may be added into this— seasonal adjustment is implemented at the disaggregate level. In section III we extend the analysis to NIPA data. Section IV concludes. I. Measuring Unusual Weather and Its Effect on Aggregate Employment Data We need to construct measures of unseasonal weather that are suitable for adjusting the CES survey. We first obtained data from the National Centers for Environmental Information on daily maximum temperatures, 230 Brookings Papers on Economic Activity, Fall 2015 precipitation, snowfall, and heating degree days (HDDs) at one station in each of the largest 50 metropolitan statistical areas (MSAs) by population, in the United States from 1960 to the present. The stations were chosen to provide a long and complete history of data, and are listed in table 1. We averaged these across the 50 MSAs, with the averages weighted by population, determined from the 2010 census. This was designed as a way of measuring U.S.-wide temperature, precipitation, and snowfall in a way that makes a long time series easily available and that puts the highest weight on areas with the greatest economic activity. Weather, of course, varies substantially around the country, and it might seem more natural to adjust state-level employment data for state-level weather effects. We used national-level employment data with national-level weather because the BLS produces state and national data separately using different methodologies. National CES numbers are quite different from the “sum of states” numbers, because both state and national CES numbers are constructed by survey methods, whereas the national data use more disaggregated cells. Meanwhile, it is the national numbers that garner virtually all the attention from Wall Street and the Federal Reserve. Let temps denote the actual average temperature on day s, and define the unusual temperature for the day as temp temp temp s s s y y * 1 30 , 1 30 ∑ = = , where temps,y denotes the temperature on the same day y years previously. Likewise, let prp*s , snow*s , and hdd*s denote the unusual precipitation, snowfall, or HDD on day s, relative to the 30-year average. This is in line with the meteorological convention of defining climate norms from 30-year averages (World Meteorological Organization 2011). In assessing the effect of unusual weather on employment as measured in the CES, we want to take careful account of the within-month timing of the CES survey. The CES survey relates to the pay period that includes the 12th day of the month. Some employers use weekly pay periods, others use biweekly periods, and a few use monthly periods. A worker is counted if she works at any point in that pay period. Cold weather or snow seems 3. The HDD at a given station on a given day is defined as max (18.3 t, 0), where t is the average of maximum and minimum temperatures in degrees Celsius. 4. An alternative measure of snowfall, used by Macroeconomic Advisers (2014), is based on a data set of daily county-level snowfall maintained by the National Centers for Environmental Information. This clearly has the advantage of greater cross-sectional granularity. However, these data only go back to 2005. Our data go much further back, allowing us to construct a longer history of snowfall effects and to measure normal snowfall from 30-year averages. MICHAEL BOLDIN and JONATHAN H. WRIGHT 231 Table 1. Weather Stations Used to Measure National Weather a MSA Station MSA Station New York Central Park San Antonio San Antonio Intl. Airport Los Angeles Los Angeles Intl. Airport Orlando Orlando Intl. Airport Chicago Chicago O’Hare Intl. Airport Cincinnati Cincinnati/Northern Kentucky Intl. Airport Dallas Dallas/Fort Worth Intl. Airport Cleveland Cleveland Hopkins Intl. Airport Philadelphia Philadelphia Intl. Airport Kansas City Kansas City Intl. Airport Houston George Bush Intcntl. Airport Las Vegas McCarran Intl. Airport Washington Washington Dulles Intl. Airport Columbus Port Columbus Intl. Airport Miami Miami Intl. Airport Indianapolis Indianapolis Intl. Airport Atlanta Hartsfield-Jackson Intl. Airport San Jose Los Gatos Boston Logan Intl. Airport Austin Camp Mabry San Francisco San Francisco Intl. Airport Virginia Beach Norfolk Intl. Airport Detroit Coleman A. Young Intl. Airport Nashville Nashville Intl. Airport Riverside Riverside Fire Station Providence T. F. Green Airport Phoenix Phoenix Sky Harbor Intl. Airport Milwaukee Gen. Mitchell Intl. Airport Seattle Seattle-Tacoma Intl. Airport Jacksonville Jacksonville Intl. Airport Minneapolis Minneapolis-Saint Paul Intl. Airport Memphis Memphis Intl. Airport San Diego San Diego Intl. Airport Oklahoma City Will Rogers World Airport St. Louis Lambert-St. Louis Intl. Airport Louisville Louisville Intl. Airport Tampa Tampa Intl. Airport Hartford Bradley Intl. Airport Baltimore Baltimore/Washington Intl. Airport Richmond Richmond Airport Denver Stapleton/Denver Intl. Airport New Orleans Louis Armstrong Intl. Airport Pittsburgh Pittsburgh Intl. Airport Buffalo Buffalo Niagara Intl. Airport Portland (Ore.) Portland Intl. Airport Raleigh Raleigh-Durham Intl. Airport Charlotte Charlotte Douglas Intl. Airport Birmingham Birmingham Airport Sacramento Sacramento Executive Airport Salt Lake City Salt Lake City Intl. Airport a. This table lists the 50 weather stations used to construct national average daily temperature, snowfall, and HDD data. Each weather station corresponds to one of the 50 largest MSAs by population in the 2010 Census. b. Stapleton International Airport was replaced by Denver International Airport in 1995. 232 Brookings Papers on Economic Activity, Fall 2015 most likely to affect employment status on the day of that unusual weather, but it is also possible that, for example, heavy snow might affect economic activity for several days after a snowstorm has ceased. Putting all this together, temperature/snowfall conditions in the days up to and including the 12th day of the month are likely to have some effect on measured employment for that month. The further before the 12th day of the month the unusual weather occurred, the less likely it is to have affected a worker’s employment status in the pay period bracketing the 12th, and so the less important it should be. It is hard to know a priori how to weight unusual weather on different days up to and including the 12th day of the month, but, on the other hand, it seems likely that unusual weather after the 12th day of the month ought to have little effect on employment data for that month. In solving this problem, we try to let the data speak. Our proposed approach assumes that the relevant temperature/precipitation/snowfall conditions are a weighted average of the temperature/precipitation/snowfall in the 30 days up to and including the 12th day of the month, using a Mixed Data Sampling (MIDAS) polynomial as the weights to avoid overfitting. We want to use this specification to collapse the daily weather data that we have into monthly weather measures. We will spell out the details of the MIDAS polynomial and its estimation below. MIDAS polynomials were proposed by Eric Ghysels, Pedro Santa-Clara, and Rossen Valkanov (2004, 2005) and by Elena Andreou, Ghysels, and Andros Kourtellos (2010) as a device for handling mixed frequency data in a way that is parsimonious yet flexible—exactly the problem that we face here. The presumption is that unusual weather on or just before the 12th day of the month should get more weight than unusual weather well before this date. In addition to temperature, precipitation, snowfall, and HDDs, there are two other weather indicators that we consider. First, as an alternative way of measuring snowfall, the National Centers for Environmental Information produce regional snowfall indexes that measure the disruptive impact of significant snowstorms. These indexes take into account the area affected by the storm and the population in that area, for six different regions of the 5. There are actually ways in which weather after the 12th could matter for CES employment that month. For example, suppose that a new hire was planning to begin work on the 13th and the 13th happens to be the last day of the pay period. She would be counted as employed in that month. But if bad weather caused the worker’s start date to be delayed, then she would not be defined as employed in that month. However, we do evaluate the possibility that weather just after the 12th could affect employment for that month. MICHAEL BOLDIN and JONATHAN H. WRIGHT 233 country. See Paul Kocin and Louis Uccellini (2004) and Michael Squires and others (2014) for a discussion of these regional snowfall impact (RSI) indexes. They are designed to measure the societal impacts of different storms, which make them potentially very useful for our purposes. They have the drawback that they do not cover the western part of the country, but there are only two big cities that are not covered and that receive significant snowfall: Denver and Salt Lake City. Any snowstorm affecting a region has an index value, a start date, and an end date. We treat the level of snowfall in that region as being equal to the index value from the start to the end date, inclusive. For example, a storm affecting the southeast region was rated as 10.666, started on February 10, 2014, and ended on February 13, 2014. We treat this index as having a value of 10.666 on each day from February 10 to 13, 2014. For each of those days, we then create a weighted sum of the six regional snowstorm indexes to get a national value, where the weights are the populations in the regions (from the 2010 Census). We then used this RSI index as an alternative to the average snowfall. Second, the household Current Population Survey (CPS) asks respondents if they were unable to work because of the weather. We seasonally adjust the number who were absent from work in month t, using the default X-13 filter, and then treat this variable, abst, as an additional weather indicator. We first estimate eight candidate models giving the effects of different weather measures on aggregate employment. Intuitively, we are simply interested in regressing monthly aggregate not seasonally adjusted (NSA) employment onto a weighted average of daily weather data, where the weights give the best possible fit. This is intended as a precursor to incorporating weather effects in CES seasonal adjustment. However, weather is only a very small part of what drives aggregate employment. We also want the model to allow for trend and seasonal components. I.A. Eight Candidate Models Each of our eight candidate models is an “airline model”—the default model in the first stage of the X-13—fitted to aggregate NSA employment, but augmented by weather variables. Each model specifies that there are trend and seasonal components that are nonstationary and consequently require taking first differences and differences from the same month one 6. This is the number with a job, not at work, in nonagricultural industries (series LNU02036012). 234 Brookings Papers on Economic Activity, Fall 2015 year earlier. After this differencing, the employment data are driven by weather effects and by moving average errors. The specific model is of the form L L y x L L t t t ( ) ( ) ( )( ) ( )( ) ′ g = + q + Q e 1 1 1 1 1 , 12 12 where yt is total NSA employment for month t, L is the lag operator, and et is an independent and identically distributed error term. The eight models differ only in the specification of the regressors in xt. The specifications that we consider are as follows: SPECIFICATION 1: TEMPERATURE ONLY There are 12 elements in xt, each of which is ∑ = w temp j s j j * 0 30 interacted with one of 12 monthly dummies, where day s is the 12th day of month t, and where w B j a b j 30 , , =      