Traditional Smoothing Techniques

Simple Moving Average:

n n
xt i +1 x
x t + 1 = or x t = t i + 1
i =1 n i =1 n

Centered Moving Average, assume n is odd:

( n 1) ( n 1)
2 2

x t i + xt + x t +i
x t = i =1 i =1

n
Weighted Moving Average:

W x i t i
x t = i =1
n

W
i =1
i

(or, of course, you could set up the Wi so that they

simply add to one.)

NoteLinear Moving Averages (MAs of MAs):

Consider a system of weights for a 7-point weighted
moving average {1,1,1,1,1,1,1}. Another 4-point moving
average with weights {1,1,1,1}. Then the 74 moving
average would have weights {1,2,3,4,4,4,4,3,2,1} and is
essentially the convolution of the two sets of weights.

1
(Single) Exponential Smoothing:

x t + 1 = x t + (1 ) x t
or
x t = x t 1 + (1 )( x t x t 1 )

Adaptive Response Rate Single Exponential Smoothing

(ARRSES). The advantage here is that is dynamic.:

x t + 1 = t x t + (1 + t ) x t

where:
et = x t x t
Et = et + (1 )Et 1, smoothed error
Mt = et + (1 )Mt 1, abs. val . of smoothed error
Et
t +1 =
Mt
= 0.2, a choice variable

Chows Adaptive Control Method:

Can be used for nonstationary data.
t is adapted by small increments so as to minimize
the MSE.

St = t x t + (1 t )St 1
bt = t (St St 1 ) + (1 + t )bt 1
1 t
and x t + 1 = St + bt
t

2
Winters Linear and Seasonal Exponential Smoothing:

x t + m = (St + bt m )It L + m
where
x
St = t + (1 )(St 1 + bt 1 )
It L
bt = (St St 1 ) + (1 )bt 1
x
It = t + (1 )It L
St
L: length of the seasonality
I: seasonal adjustment factor
St: smoothed value for the series
m: forecast period
: a weight to suppress randomness (often 0.05)
: exponential factor for smoothing (often 0.2)
: parameter (often 0.1)

DECOMPOSITION METHODS

x t = f (It ,Tt , Ct , Et )

xt: actual at time t

It: seasonal component (or index) at t
Tt: is trend component at t
Ct: cyclical component at t
Et: error or random component at t

This function f() can be additive or multiplicative,

yielding an additive decomposition or a multiplicative
decomposition.

3
Example, Additive Method:

1. Compute a moving average of length N, where N is

the length of the seasonality. This eliminates the
seasonality by averaging seasonally high periods
with seasonally low periods, and reduces
randomness as well.

2. Subtract the moving average from the series.

The MA is the trend plus cycle. The error is the
seasonal component.

3. Isolate the seasonal component by averaging

them for each of the periods making up the
complete length of the seasonality.

4. Identify the appropriate form of the trendlinear,

exponential, S-curve, etc., and calculate its value at
each period t.

5. Subtract the estimated trend from the

deseasonalized series to obtain the cyclical factor.

6. Subtract the seasonal, trend, and cycle

components from the original series to yield the
random component.

4
 CENSUS II X-11
Decomposition/Seasonal Adjustment Method
The Census Method I began in 1954, followed by twelve
experimental programs, named X-0, X-1, etc., of Method
II. This culminated in X-11.

U. S. Department of Commerce, Bureau of the Census. Julius Shiskin

(1955), based upon the ratio-to-moving average classical decomposition.

Shiskin, J., A. H. Young, and J. C. Musgrave. The X-11 variant of the

Census method II seasonal adjustment program. Technical Paper 15,
Bureau of the Census, U.S. Department of Commerce, 1967.

Shiskin, Julius. Seasonal Adjustment of Sensitive Indicators, 1978. In A.

Zellner, editor, Seasonal Analysis of Economic Times Series, pages 97-103.
U. S. Department of Commerce, Bureau of the Census.

X-11 was popular because:

It was relatively easy to use.
It did not require restating past values when new
data was released.
It handled extreme values well.
It used well-known moving averages methods for
estimating trend and seasonal components.
The asymmetric moving averages used near the
ends of the time series were thought to be tried
and true.
It had a clear-cut way of estimating trading day
effects.

Statistics Canada extended the method as X-11-ARIMA.

This method included the full X-11 method, but used
ARIMA backcasts and forecasts to provide optimal
estimates of data outside the data window to improve
estimates at the ends of moving averages.

5
X-11-ARIMA results in seasonal adjustments whose
revisions are smaller, on average, when they are
recalculated after future data becomes available.

In the additive decomposition case, extension with

optimal forecasts and backcasts for the half-length of
the symmetric seasonal filter minimizes revisions in the
mean square sense.

Bobbitt, L, and M. C. Otto. Effects of forecasts on the revisions of

seasonally adjusted values using the X-11 seasonal adjustment
procedure. In Proceedings of the Business and Economics Statistics
Section, pages 449-453, Alexandria, Virginia, 1990. American Statistical
Association.

X-11-ARIMA also added diagnostics for comparing

direct and indirect seasonal adjustments of series that
are aggregates of multiple component series.

WE START WITH THE ORIGINAL X-11 METHOD:

Step 1. Trading Day Adjustment

1. Determine the number of active days in each
month for the years of interest.
2. Compute the average number of trading days for
each month.
3. Divide the number of days in each month by this
average to get an adjustment factor.
4. Use the adjustment factor to adjust the monthly
figures.
5. This creates a value called original data adjusted
for trading days.

6
Step 2. Preliminary Seasonal Adjustment.
Seasonality Adjustment
1. Apply a 12-month MA to eliminate seasonality.
2. Average the MAs of 2 successive months to form
the 7th month value. This addresses the centering
problem.
3. Form the ratio of the original series to the MA
series.

Extreme Values
4. Calculate the 3x3 month moving average (3-month
average of a 3-month average).
a. This is roughly equivalent to a 5-month moving
average.
b. Strictly speaking, this should result in the loss
of 2 months at the beginning and end of the
series, but Census estimates replacements
for these.
5. Calculate the standard deviation of the centered
ratios from the 3x3 MA.
a. This is used to construct control limits to
identify extreme values.
b. If the centered MA > 3x3 MA 2s2, then replace
it with the average of previous and following
period.

Preliminary Seasonal Factor Estimation & Application

6. Replace the 6 month at the beginning and end of
the ratios by the nearest values in a neighboring
year.
7. Normalize years so that the ratios in each year
add to 12. (Average ratio is 1.)
8. Divide the preliminary seasonal factors into the
original data to obtain the preliminary adjusted
series.

7
Step 3. Refine Seasonal Adjustments.
1. Apply Spencers 15 month weighted moving
average to the seasonally adjusted data. This is a
5x5x4x4 moving average (quadruple MA)
a. Isolates the trend-cycle component.
2. Divide the original data by the trend-cycle
component
a. Seasonal and random factors remain. These
are called the final seasonal irregular ratios.
b. Normally Spencers Method would cause the
loss of 7 points at the beginning and the end of
the series, so Census replaces the lost data
points with estimates.
3. Replace the extreme values as above.
4. Estimate missing values.
5. Adjust (normalize) ratios.
6. Take 5-year averages of these final seasonal-
irregular ratios
7. These are the stable factors (seasonal indices).

Step 4.
1. Apply a 3x3 moving average (or 5x5 if it still looks
too random) to the final seasonal irregular ratios.
2. Estimate values for the 2 periods at the beginning
and end of the series that would be lost.
3. Take the last 2 values for each month, and form an
expected value. For example, for 1992,
x 1992 = [( x 1991 3 ) x 1990 ] / 2

4. Divide these final seasonal factors into the original

data to form the seasonally adjusted series.

8
Step 5. Final adjustment.
1. Calculate a 15 month MA to create the final
seasonally adjusted data.
a. This is an estimate of the trend cycle
component.

Step 6. Create a mountain of summary statistics.

The Census X-12-ARIMA includes X-11, but extends the

modeling and diagnostic capabilities.

X-11 ARIMA
RegARIMA Models
(Forecasts, Backcasts,
and Preadjustments)

Modeling and
Model
Comparison
Diagnostics

SEASONAL
ADJUSTMENT
(Enhanced X-11)

DIAGNOSTICS

9
The major methodological improvements of X-12-ARIMA
are:
New X-11 adjustment options
New diagnostics
New modeling capabilities emphasizing regARIMA
modeling . (RegARIMA is a linear regression model
with ARIMA time series errors.)

NEW X-11 ADJUSTMENT OPTIONS

New filter options, including:
o longer seasonal moving average,
o user specification of Henderson filters
o modifications to asymmetric moving averages
Option for pseudo-additive decomposition,
sometimes useful for series with periodically small
or zero values.
Improvements in trading day adjustments and
options for user-defined effects based upon
preliminary estimates of the irregular component.

NEW DIAGNOSTIC CAPABILITIES

Spectral estimates for detection of seasonal and
trading day effects
Revisions history diagnostics for assessing the
stability of seasonal adjustments.
Better diagnostics for deciding whether to use
direct or indirect adjustments for aggregate series.

New RegARIMA CAPABILITIES

Capability to add regression effects to the models
for forecast extension.
Use of RegARIMA models can potentially improve
forecasts and backcasts, and provide earlier outlier
detection capabilities.

10
TYPES OF DECOMPOSITIONS THAT MAY BE
SELECTED WHEN USING X-11
Multiplicative Decomposition
o Usually appropriate for series of positive
values in which the size of the seasonal
oscillations increases with the level of the
series.
o The seasonally adjusted series is obtained by
dividing the original series by the estimated
seasonal component.
Additive Decomposition
o More appropriate to stationary series.
o The seasonally adjusted series is obtained by
subtracting the estimated seasonal
component.
Log-additive Decomposition
o The additive decomposition of the logarithms
of the series being adjusted is exponentiated.
o Mainly used for research purposes. Requires a
bias correction.
Pseudo-additive decomposition

In the updated X-11 (in X-12), the Spencer MA is

replaced by the Henderson filter. This is either 9, 13, or
23 points and is symmetric. It is designed to
approximate a cubic fit to stationary data.

Spectral analysis of the Henderson filter reveals that it

has substantial power after the first seasonal frequency
(leakage?). As a result, Schips and Stier (1995) argue
that the Henderson filter exaggerates short-term cyclical
behavior. The 17-term Henderson filter is the shortest

11
one that does not result in a significant peak beyond the
first seasonal frequency.

12