In the 1950's, Julius Shiskin
and others at the United States Census
Bureau began using the electronic
computer to seasonally adjust
economic time series. By 1957,
more than 3,000 series had been
adjusted for seasonal variations.
The Census Bureau's method for
adjusting time series, the Census
II X-11, is today the most widely
used method for deseasonalizing time series data. Seasonally
adjusted data play an important
part in decisions on national
monetary and economic policies,
and are regularly reported in the
news media. Despite the widespread
use and acceptance of
seasonally adjusted data, seasonal
adjustment is not without its
critics. Some even go as far as to
question whether it should be done
at all.
What Is Seasonality?
If a periodic pattern observed in
a time series is shorter than a
year, and if this regular pattern of
fluctuations tends to repeat from
year to year, then this periodic pattern is usually called seasonal
movement or seasonality. For the
most part, seasonal effects are due
to either one or a combination of
the calendar and weather and
temperature related factors.
For example, the graph below
shows the unadjusted and the
seasonally adjusted Connecticut
Wages and Salaries series from
1980 to 1995. The saw-toothed line represents the unadjusted
series, which displays the yearly
seasonal peaks in the fourth
quarter. Notice that the seasonal
peaks become more pronounced by
the 1990's. Thus, though a given
seasonal effect occurs at the same
time of the year, from year to year
the magnitude can be changing.
The smoothed line represents the
Wages and Salaries series after
having the seasonal component
removed using the Census II, X-11
ARIMA Method.
The table below 'zooms in'
on the unadjusted and adjusted
Wages and Salaries for the four
quarters of 1995. For the unadjusted
series, Wages and Salaries
decline from the first to third
quarters, then in the fourth
quarter they reach their peak level
for 1995, $59.9 billion, which
reflects seasonal effects such as
the increased economic activity
during the Christmas season and
year-end bonuses. However, it may
also include nonseasonal effects
such as a response to anticipated
tax changes. This occurred in the
fourth quarter of 1992 after the
new administration announced its
federal tax change proposals. The
seasonally adjusted Wages and
Salaries series follows the same
pattern as the unadjusted series.
Notice, however, that the values for
the first three quarters are higher
than their corresponding values for
the unadjusted series, while the
fourth-quarter value of the adjusted
series, at $54.6 billion, is
significantly lower than that for
the unadjusted series. This presumably
is the result of filtering
out all seasonal influences, leaving
nonseasonal effects.
Table 1:
1995 |
Unadjusted |
Adjusted |
First Quarter |
53,498 |
54,078 |
Second Quarter |
52,232 |
54,062 |
Third Quarter |
51,103 |
53,861 |
Fourth Quarter |
59,869 |
54,643 |
* For workers covered by Unemployment Insurance.
The Theory Behind Seasonal Adjustment
Seasonal adjustment of economic
data is predicated on the
idea that an observed time series
can be decomposed into unobserved
seasonal and nonseasonal
components. The notion that a
series of observations over time is
composed of separate unobserved
components was widely applied by
seventeenth century astronomers.
Today, four so-called unobserved
components are familiar to the
time-series analyst: the seasonal
component (which was described
above), trend, cycle, and irregular.
The trend or secular trend is the
long-term tendency of a series to
rise or fall. Cycles are broad
contractions and expansions that
take place over a period of years (not within each year). The length
of time between successive peaks
(or troughs) is not necessarily
fixed. The familiar expansion-recession-recovery-expansion of the
business cycle is a good example of
cyclical movement encountered in
economic data. The movement left
over after accounting for seasonality,
trend, and cycle is called the
irregular component. Borrowing
from engineering terminology, the
irregular component is often called
the 'noise' in the series.
Early on, there was criticism of
the empirical methods now embodied
in the X-11 (and recently
released X-12). These methods use
moving-average techniques for
isolating and then removing the
seasonal component. Critics
proposed model-based methods as
a preferred alternative to the
empirical approach. Recently,
there has been interest in using
probabilistic models and other
methods to do seasonal adjustment
based on signal extraction, a
concept from the field of communications
engineering.
Why Seasonally Adjust Data?
A problem in using business
indicators has been the difficulty
of separating the underlying, more
meaningful cyclical movements
from other types of fluctuations.
Extracting such information to
determine the stage of the business
cycle helps in forecasting and
provides a factual basis for taking
steps to moderate the business
cycle. (In fact, the Federal Reserve
Board has effectively removed
seasonality from interest rates
through monetary policy). Three
main reasons for seasonal adjustment
are: (1) to aid in short-term
forecasting; (2) to aid in relating a
time series to other time series,
external events or policy variables;
and (3) to achieve comparability in
the series values from month to
month.
Criticisms of Seasonal Adjustment
Some economists contend that
the seasonal cycle is an important
source of variation that should be
studied in its own right, as the
business cycle is. In fact, they maintain that the business cycle
and the seasonal cycle display
similar characteristics. They refer
to it as the 'Seasonal Business
Cycle'. Their results argue against
removing seasonality from the data.
There are other problems that
arise from seasonally adjusting
data. Recent research suggests that
seasonal adjustment may affect the
results of statistical methods used
to test for the presence of trends in
the data. In his textbook on timeseries
analysis, Walter Enders
warns that though a standardized
procedure may be necessary for
government agencies such as the
United States Bureau of Economic Analysis
(BEA), which reports hundreds of
series, it may not be the best
procedure for an individual modeling
a single series. Further, seasonally
adjusted data may still contain
some residual seasonal pattern.
From an administrative and
procedural point of view, problems
arise from not incorporating changing
seasonal effects (as depicted in the graph), or from a mechanical
process which incorporates variation
not caused by seasonal phenomena,
or caused by (or masked
by) classification effects, into
calculating seasonal factors.
Changing seasonality can be
mitigated somewhat by concurrent
seasonal adjustment or breaking
the series, or both. The monthly
nonfarm employment estimates
are an example of where classification
effects arise. The Office of
Management and Budget (OMB)
ruled that those employed by the
tribal councils be classified in local
government. Any seasonal variation
in employment specific to
tribal councils or local governments,
or both, is now submerged
in one aggregated seasonal factor.
Temporary or one-time events can
result in a transitory spike in the
magnitude of a seasonal factor
(such as the fourth quarter 1992
spike noted earlier). However, this
can be accommodated by intervention
and concurrent seasonal adjustment (which the Connecticut
Labor Department does on its
own series).
To Adjust or Not to Adjust?
So how does one respond to all
of this? When it comes to the
series that are published by the
federal agencies such as United States Bureau of Economic Analysis, we
have no choice in the matter.
Many economic series, such as the
quarterly Personal Income series,
are seasonally adjusted by United States Bureau of Economic Analysis
and other federal agencies before
release to the public. For those
series developed by the Connecticut
Department of Labor, such as
the Wages and Salaries series,
both the adjusted and unadjusted
series are forecasted and published
in the "Connecticut
Economic Digest". While
recognizing the issues discussed
here are far from resolved, this
follows the practice of the federal
statistical agencies and adds
additional data on which to track
the economy in Connecticut.
The Connecticut Department
of Economic and Community
Development announced that
Connecticut communities authorized
717 new housing units in
September 1996, a one percent
increase compared to August
1996 when 710 were authorized.
The Department further
indicated that the 717 units
permitted in September 1996
represent an increase of 5.8
percent from the 679 units
permitted in September 1995,
and that the year-to-date numbers
are down 11.2 percent, from
6,385 in 1995 to 5,672 in 1996.
Reports from municipal
officials throughout the state
indicated that New London County showed the greatest
percentage increase in September
compared to the previous month;
122 percent. Tolland County
reported the greatest percentage
decline; 23 percent.
New Haven County documented
the largest number of new, authorized
units in September with 155.
Hartford County followed with 137
units and New London County had
131 units. Norwich led all Connecticut
communities with 44
units, followed by Waterford with
24 and Cheshire with 21.
The permit activity figure for
September included the following
statewide amounts by structure
type: detached single-family units,
632, attached single-family units, 4; two units structures, 2; three
and four-unit structures, 21;
structures containing five or
more units, 58.
Year-to-date totals indicate
that Hartford County has issued
the most building permits
through the first nine months of
1996 with 1,291, followed by
New Haven County with 1,173
and Fairfield County with 1,111.
Southington authorized 154 new
permits during this period,
followed by Waterford with 134,
Shelton with 133, Rocky Hill
with 131, and Wallingford with
130.
The statistical agencies of Mexico,
Canada and the United States
are cooperating in the design of the
North American Industry Classification
System (NAICS). This industry
classification system will be used
by the three countries and will
replace the Standard Industrial
Classification (SIC) system currently
used in the United States by
the Bureau of Labor Statistics and
other agencies, in the development
of labor force statistics and other
economic indicators.
In developing North American Industry Classification, the three
countries agreed that the following
principles would be followed:
- North American Industry Classification would be built on a
production-oriented framework.
Production units which use the
identical or similar processes
would be grouped together. The
current Standard Industrial Classification system groups businesses
by their products and
services.
- Special attention
would be given to new and emerging
industries, service industries in
general, and industries engaged in
the production of advanced technologies.
- Time series continuity
would be maintained
whenever possible.
However, changes in
the economy and
establishing comparability
between the
three countries would
have priority.
- North American Industry Classification would attempt
to be comparable
with Revision 3 of the International
Standard Industrial
Classification system at the two
digit level.
Structure
North American Industry Classification is a hierarchical system
similar to the Standard Industrial Classification, however it will
have a six digit structure rather
than the four digit Standard Industrial Classification structure.
Under North American Industry Classification there are 1,160
industries grouped into 21 Industry
Sectors, as compared to 1,005
industries grouped into nine Major
Industry Divisions under the Standard Industrial Classification
system. The North American Industry Classification structure is as
follows:
- XX Industry Sector
- XXX Industry Subsector
- XXXX Industry Group
- XXXXX Industry
- XXXXXX Country specific industry detail*
*Identifies important industries particular to the United States
Impact of the Change
A direct match to a North American Industry Classification
code will be possible for 661
current Standard Industrial Classification industries. The
remaining 344 Standard Industrial Classification industries
will be split. These split industries
will affect approximately
3.6 million units, with employment
of 55.5 million nationally.
(An assessment of how many
Connecticut employers are in
split industries has not yet
been made.)
It has been estimated that
511 industries will have a
potential time series break. Of
these, at least 256 will have breaks involving three percent or more of
the units classified, with employment
of 39.9 million nationally. To
preserve time series continuity, the
United States statistical agencies will
develop a crosswalk to link the Standard Industrial Classification
and North American Industry Classification using the dual coded
units to produce the data series
using both classification systems.
National time series data that will
be affected include industry
employment developed by the U. S. Department of Labor, Bureau of
Labor Statistics and personal
income data produced by the United States
Department of Commerce, Bureau
of Economic Analysis. In Connecticut,
the nonfarm employment
data series will be impacted, as
well as data on employment and
wages covered by unemployment
insurance.
In addition to the structural
changes and the shift in focus to a
production-based classification
system, another major difference
in the two coding systems is the
treatment of auxiliary establishments.
Under the Standard Industrial Classification system,
units that perform support functions
for other establishments of
an enterprise are assigned the
industry classification of the
establishments they support, with
an auxiliary code to further indicate the type of establishment.
Under North American Industry Classification, these establishments
will be classified according
to the economic activities they
themselves perform. For example,
under the Standard Industrial Classification system, the corporate
headquarters of a manufacturing
enterprise is assigned a
manufacturing code with the
auxiliary code for corporate headquarters.
Under North American Industry Classification, this
establishment would not be
classified in manufacturing, but would be classified
under the North American Industry Classification sector Management, Support and Remediation Services.
Implementation
Under one current
proposal, the
conversion to
North American Industry Classification will be phased in over a three year period
beginning in Fiscal Year 1998. Approximately
1.7 million units in these
split industries will be surveyed
each year during Fiscal Year 1998 and Fiscal Year
1999 and will be dual coded using
both classification systems. In Fiscal Year
2000, all new businesses and nonrespondents
will be surveyed and
North American Industry Classification codes will be applied to any
remaining non-respondents.
Beginning with the second quarter
2000 data, only North American Industry Classification codes will
be applied. An alternate implementation
schedule would complete
the coding in one year, with
the exclusive use of North American Industry Classification codes
beginning in the second quarter
1999. Either way, both the task
and the effect of conversion to
North American Industry Classification will be significant.
Connecticut's coincident
employment index moved,
once again, to its highest level in
the current recovery with the
release of the (preliminary)
August data, having not fallen on
a month-to-month basis since
December 1995. The leading
index, however, remained unchanged
from its July level.
The coincident index, a gauge
of current employment activity,
rose in August, continuing its
strong upward momentum that
has characterized this year. This
experience in 1996 follows weak
upward movement in the coincident
index in the early phases of
the current recovery. In sum, the
economy continues to climb out
of the Great Recession, and has
accelerated its pace in 1996.
Connecticut watchers can only
hope that this trek continues in
the near term.
The leading index, a barometer
of future employment activity,
has bounced around considerably
during 1996. It reached its highest level in the current
expansion in June. We shall
continue to monitor the leading
index closely to see if, and when,
it starts a strong downward
move. No such signal is yet
observable in the data.
The coincident employment
index rose from 83.0 in August
1995 to 88.8 in August 1996. All
four index components continued
to point in a positive direction on
a year-over-year basis with
higher nonfarm employment,
higher total employment, a lower
total unemployment rate, and a
lower insured unemployment
rate.
Let us examine the individual
components a bit more closely.
Over the past year, nonfarm
employment, which is based on
an employer survey, rose by only
1.1 percent. Total employment,
which is based on a household
survey, rose by 2.5 percent,
however. Finally, the insured and
total unemployment rates both
fell over the year by 13.5 and 14.5 percent, respectively. As a
result, the individual components
collectively produced a 7 percent
rise in the coincident index.
The leading employment index
rose from 87.6 in August 1995 to
88.4 in August 1996, or somewhat
below its previous peak of
June 1996. Two of the five index
components sent positive signals
on a year-over-year basis with
lower initial claims for unemployment
insurance and a lower
short-duration (less than 15
weeks) unemployment rate. One
component sent a negative signal
on a year-over-year basis with
lower total housing permits. The
final two components, average
work week of manufacturing
production workers and Hartford
help wanted advertising were
unchanged on a year-over-year
basis.
Source: Connecticut Center for Economic Analysis, University of Connecticut. Developed by Pami Dua [(203) 322-3466,
Stamford Campus (on leave)] and Stephen M. Miller [(860) 486-3853, Storrs Campus]. Tara Blois [(860) 486-4752, Storrs
Campus] provided research support.
Return to Top
|