Benchmarking and temporal disaggregation

plug-in for JDemetra+: User guide

Version 2.0.0

November 23, 2015

Contents

1 Introduction 2
1.1 Overview of the plug-in . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Installing and navigating to the plug-in . . . . . . . . . . . . . . . 2
1.3 How to use this guide . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Contact details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Quick start guide 3

2.1 Benchmarking - Denton and Cholette . . . . . . . . . . . . . . . . 3
2.2 Benchmarking - Multi-variate Cholette . . . . . . . . . . . . . . . . 3
2.3 Calendarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Temporal disaggregation . . . . . . . . . . . . . . . . . . . . . . . . 4

3 Function guide 6
3.1 Benchmarking - Denton . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Benchmarking - Cholette . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Benchmarking - Multi-variate Cholette . . . . . . . . . . . . . . . . 8
3.4 Calendarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.5 Temporal disaggregation . . . . . . . . . . . . . . . . . . . . . . . . 9
1 Introduction

1.1 Overview of the plug-in

This plug-in allows the user to carry out benchmarking and temporal disaggre-
gation of time series in JDemetra+.
The following functions are made available in JDemetra+ using the benchmarking
plug-in:
Benchmarking Constraining higher frequency time series to lower frequency time series. Methods
available are:
• Denton
• Cholette
• Multi-variate Cholette
Calendarization Converting a flow time series of varying time intervals into a monthly, quarterly
or yearly time series.
Temporal disaggregation Converting a lower frequency time series to a higher frequency time series possibly
using a (single) higher frequency indicator series.

1.2 Installing and navigating to the plug-in

To install the benchmarking plug-in, navigate Tools>Plugins from the drop-
down menus. For more information see JDemetra+ Reference Manual Version
0.1 . If the benchmarking plug-in has been successfully loaded you should be able
to see additional functions in the Statistical methods drop-down menu. These
include Benchmarking and Temporal Disaggregation.

1.3 How to use this guide

This guide is split into two sections, a quick start guide and a function guide.
Quick start guide This provides an overview of using each function in the plug-in.
Function guide This provides details on the options available within each function, the method-
ology used, references and examples.

1.4 Contact details

All enquiries regarding this plug-in should be made through the Seasonal Ad-
justment Helpdesk http://www.cros-portal.eu/content/ess-seasonal-adjustment-
helpdesk
2 Quick start guide

This section provides instructions for starting using the functions included in this
plug-in. For more detailed descriptions of the functions see the function guide.

2.1 Benchmarking - Denton and Cholette

To benchmark a high frequency series to a low frequency series using either the
Denton or the Cholette method follow the steps described below.
1. Navigate, Statistical methods>Benchmarking>Denton (or Cholette).
This will open a new window.
2. Drag your high frequency series from your Providers window and drop it
on the Series section. This is the series that you want to constrain.
3. Drag your low frequency series from your Providers window and drop it on
the Constraint section. This is the series that you want to constrain to.
4. Alter the specifications if necessary, using the Specifications button in the
top right of the screen (default specification is described in more detail
below). For example, you may want your higher frequency series to sum
to the lower frequency series over one period of the lower frequency series.
Press the Apply button to apply any changes that you have made.
5. Your benchmarked series (high frequency series that now meets the specified
benchmarking constraints) is presented in a plot. You can click and select
the series for example to drag it into Excel.

2.2 Benchmarking - Multi-variate Cholette

To ensure mutliple contemporaneous and temporal constraints on multiple series
follow the steps described below.

1. Navigate, Statistical methods>Benchmarking>Multi-variate Cho-

lette
2. Drag all series that will be constrained and constrained to from your Providers
window and drop on the Drop data here section. Source data can be con-
tained in one file and dragged in all together or multiple files and each file
can be dragged and dropped individually.
3. To change the specification and input constraints, under Details, right click
Benchmarked series and click Properties.
4. To input constraints click the box containing three dots to the far right of
Constraints.
5. Write each constraint in the Item box and click Add to add the constraint.
Repeat if multiple constraints are being used. When all constraints have
been entered click Ok and then click Close on the Properties window.
 6. To run the benchmarking, click the green arrow.
7. Your benchmarked series (high frequency series that now meet the specified
benchmarking constraints) are presented in a table. You can click and select
the series individually or the whole table, for example to drag it into Excel.

2.3 Calendarization
Given a flow time series on varying time intervals, calendarization can be used
to produce a time series on a monthly, quarterly or yearly calendar basis. Daily
weights can be specified to allow for trading day effecst to be introduced into the
calendarized series.
1. Navigate, Statistical methods>Benchmarking>Calendarization. This
will open a new window called Calendarization.
2. Data can be entered manually or dragged and dropped from a spreadsheet.
3. To enter data manually, for each observation, enter the start and end dates
and the value, then click add. Continue until all values have been enetered.
4. To drag and drop values from a spreadsheet, create a spreadsheet with three
columns and no headings. The first column contains the start dates, the
second column the end dates and the third column the values of the variable
that is being calendarized. Select these data in the spreadsheet, then drag
into JDemetra+ and drop onto the Drop data here panel.
5. To specify daily weights, select the Daily Weights button and specify weights
for each weekday (note that the default is 1 for each day).
6. Select the desired Frequency for the calendarized series from the drop-down
menu. The options are monthly, quarterly and yearly.
7. Two charts and a table are produced. The chart in the top window entitled
Smoothed Data plots the daily values. The chart in the lower window in
the Chart tab entitled Aggregated Data plots the calendarized series. The
calendarized series and standard errors are provided in the Grid tab in the
lower window. You can click and select the series individually or the whole
table for example to drag it into Excel.

2.4 Temporal disaggregation

To derive a higher frequency time series from a lower frequency time series using
a regression model follow the steps described below.
1. Navigate, Statistical methods>Temporal Disaggregation>Regression
Model. This will open a new window called TsDisaggregationModel.
2. Drag and drop your low frequency series that you want to disaggregate into
the section for the Y data labelled Drop data here.
3. (Optional Step) Drag and drop your high frequency regression variable used
for disaggregation into the section for the X data labelled Drop data here.
Note that if no variable is used then by default a constant is used.
4. Alter the specifications if necessary, using the Specifications button in the
top right of the screen (the default specification is described in more detail

4
 below). For example you may want change the model for the error term
from the default AR(1) which implements the Chow-Lin method to another
model. Press the Apply button to apply any changes that you have made.
5. The results provided include a summary of the model, model residuals, a
chart and table of the main results (your new higher frequency series) and
a preview chart of the series prior to disaggregation.

5
3 Function guide

3.1 Benchmarking - Denton

Description Benchmarks a high frequency series to a low frequency series using the method
of Denton (1971).
Menu path Statistical methods>Benchmarking>Denton
Inputs Requires one high frequency series and one low frequency constraint series.
Specification options
Type Takes values Sum, Average, First or Last. This establishes the type of bench-
marking constraint. For example you might a quarterly time series measuring
turnover to sum to an annual time series measuring turnover within each year,
or you might want your monthly index series to average your annual index series
within each year, or finally you may want your quarterly stock variable to be
equal to an annual stock figure for the last quarter of the year.
Multiplicative Tick box for true or false. If the tick box is selected, a multiplicative benchmark-
ing is applied. If it is not selected, an additive benchmarking is applied. The
choice will depend on the relationship between the series.
Modified Denton Tick box for true or false. If the tick box is selected a modification to the original
Denton method is applied (recommended). If the tick box is not selected, the
original Denton method is applied.
Differencing Takes a numeric interger value of 0, 1, ... n for order n differencing. Specifies the
order of differencing. Typically a first order difference is used.
Outputs A high frequency series that meets the constraints in the low frequency series.
Methodology The method of Denton (1971) aims to preserve movements in the high frequency
series while meeting the constraints of the low frequency series.
The additive method preserves the period-on-period differences and the multi-
plicative method preserves period-on-period growth rates.
Because the method uses differences or growth rates in the series, there is a
problem at the beginning of a series when using the orginal method of Denton
(1971), as the first value is not available. The modified Denton method overcomes
this problem.
References F.T Denton (1971). “Adjustment of monthly or quarterly series to annual totals:
an approach based on quadratic minimization”. In: Journal of the American
Statistical Association 66.333, pp. 99–102
P.A Cholette and E.B Dagum (1994). “Benchmarking time series with autocor-
related survey errors”. In: International Statistical Review/Revue Internationale
de Statistique, pp. 365–377
3.2 Benchmarking - Cholette
Description Benchmarks a high frequency series to a low frequency series using the method
of Cholette (1979).
Menu path Statistical methods>Benchmarking>Cholette
Inputs Requires one high frequency series and one low frequency constraint series.
Specification options
Type Takes values Sum, Average, First or Last. This establishes the type of bench-
marking constraint. For example you might a quarterly time series measuring
turnover to sum to an annual time series measuring turnover within each year,
or you might want your monthly index series to average your annual index series
within each year, or finally you may want your quarterly stock variable to be
equal to an annual stock figure for the last quarter of the year.
Aggregation frequency Takes values Yearly, HalfYearly, QuadriMonthly, Quarterly, BiMonthly, Monthly.
Specifies the aggregation frequency for which the benchmarking constraint is to
be met.
Rho Takes values between -1 and 1. Determines the parameter value for the autore-
gressive model (AR(1)) for the error terms in the path series. The default value
of 1 is equivalent to Denton Benchmarking.
Lambda Takes values between 0 and 1. Parameter value that enables equivalence to
multiplicative (lambda=1) or additive (lambda=0) benchmarking.
Outputs A high frequency series that meets the constraints in the low frequency series.
Methodology The method of Cholette (1979) is a regression-based approach to benchmarking.
It is implemented in JDemetra+ as follows.
Let s be the high frequency series and a be the lower frequency series. These two
series satisfy equations 3.1 and 3.2.

s = θ + e, E(e) = 0, E(ee0 ) = Ve (3.1)

a = Jθ + , E() = 0, E(0 ) = V , E(e0 ) = 0 (3.2)

Where J is an aggregation matrix describing how the low and high frequency
series are related.
The covariance matrix Ve has the structure Ve = Ξλ ΩΞλ where Ξ is a diagonal
matrix of standard deviations and Ω contains the autocorrelations of the error
function.
The value of λ (lambda) can be specified by the user and takes values between
0 and 1. Typically it takes values of 0, 1/2 or 1. When λ = 0 this is equivalent
to the additive method and when λ = 1 this is equivalent to the proportional
method.
The error function implemented in JDemetra+ is an autorgressive process with
order 1 (AR(1)). The value of the coefficient rho is specified by the user and can
take values between -1 and 1.
References P.A Cholette (1979). “Adjustment methods of sub-annual series to yearly bench-

7
 marks”. In: Proceeding of the Computer Science 12 th Annual Symposium on the
Interface, ed. JF Gentleman, pp. 358–360

3.3 Benchmarking - Multi-variate Cholette

Description Used to ensure mutliple contemporaneous and temporal constraints on multiple
series.
Menu path Statistical methods>Benchmarking >Multi-variate Cholette
Inputs Mutliple time series of differing frequencies.
Specification options
Rho Takes values between -1 and 1. Determines the parameter value for the autore-
gressive model (AR(1)) for the error terms in the path series. The default value
of 1 is equivalent to Denton Benchmarking.
Lambda Takes values between 0 and 1. Parameter value that enables equivalence to
multiplicative (lambda=1) or additive (lambda=0) benchmarking.
Constraints Defines the benchmarking constarints between series. Two types of contraint can
be included: contemporaneous and temporal constraints.
Contemporaneous constraints take one of two forms:
• y = a1 ∗ x1 + . . . + an ∗ xn where y, x1 , . . . , xn are time series of the same
frequency.
• c = a1 ∗ x1 + . . . + an ∗ xn where c is a constant and x1 , . . . , xn are time
series of the same frequency.
• It is also possible to use a wild card constraint to simplify expressions. For
example the constraint c = x1 + . . . + xn can be written as c = x?.
Temporal constraints take the following form:

• S = sum(s) benchmarks a high frequency series s to a low frequency series

S.
The series on the left of the equations is called the binding series. It can be used
more than once as a binding series, but it cannot appear on the right hand side
of any equations.
Outputs Multiple time series meeting the specified contemporaneous and temporal con-
straints.
Methodology The multi-variate Cholette is an extension of the univariate Cholette method
(Cholette (1979)), using a diffuse restricted Kalman filter as defined in, for ex-
ample, Pizzinga (2009)
References
P.A Cholette (1979). “Adjustment methods of sub-annual series to yearly bench-
marks”. In: Proceeding of the Computer Science 12 th Annual Symposium on the
Interface, ed. JF Gentleman, pp. 358–360
A Pizzinga (2009). Diffuse Restricted Kalman Filtering. http://virtualbib.
fgv.br/ocs/index.php/sbe/EBE09/paper/viewFile/938/296

8
3.4 Calendarization
Description Converts a flow time series of varying time intervals into a monthly, quarterly or
yearly time series using the method of Quenneville, Picard, and Fortier (2013).
Menu path Statistical methods >Benchmarking >Calendarization
Inputs A flow time series of varying time intervals. Start and end dates of the time
intervals are required.
Specification options
Daily weights Weights specified for each day of the week. Takes numeric values. Default weights
are 1 for all days.
Frequency Takes values Monthly, Quarterly, Yearly. Sepcifies the frequency of the output
series.
Outputs Time series on the specified frequency and, where more than one observation is
used, associated standard errors.
Methodology Given a time series of varying time intervals, calendarization uses the method
of Quenneville, Picard, and Fortier (2013) to produce a monthly, quarterly or
yearly time series. The method uses the input data and as set of daily weights to
produce a daily series. The daily series is then aggregated to produce the output
series.
The daily weights form a daily indicator series used to create the daily time series.
These weights allow trading day patterns to be included in the calendarized series.
The daily time series is created using splines, for more detail see Quenneville,
Picard, and Fortier (2013).
References B Quenneville, F Picard, and S Fortier (2013). “Calendarization with interpolat-
ing splines and state space models”. In: Journal of the Royal Statistical Society:
Series C (Applied Statistics) 62.3, pp. 371–399

3.5 Temporal disaggregation

Description Converts lower frequency time series to higher frequency time series possibly using
a (single) higher frequency indicator series. Includes the methods of ordinary
least squares regression, Chow-Lin (Chow and Lin (1971)), Litterman (Litterman
(1983)) and Fernandez (Fernandez (1981)).
Menu path Statistical methods>Temporal disaggregation >Regression model
Inputs Requires one low frequency series Y and multiple high frequency series X.
Specification options
Estimation span Takes values All, From, To, Between, Last, First, Excluding. Specifies the span
of data used to fit the regression model.
Error Takes values Wn, Ar1, Rw, RwAr1. Specifies the errors process. Wn is white
noise and is equivalent to ordianry least squares regression. Ar1 is an autore-
gressive process equivalent to the method of Chow-Lin. Rw is a random walk
process equivalent to the method of Fernandex. RwAr1 is a random walk with
autoregressive errors equivalent to the method of Litterman.
Parameter Takes values between -1 and 1 if Ar1 or RwAr1 selected for error. Determines

9
 the value of the autoregressive parameter in the error process.
Constant Tick box for true or false. If ticked will include a constant term in the regression.
Trend Tick box for true or false. If ticked will include a linear trend in the regression.
Type Takes values Sum, Average, First, Last. Determines the relationship between the
high and low frequency series.
Default frequency Takes values Undefined, Yearly, HalfYearly, QuadriMonthly, Quartely, BiMonthly,
Monthly. Determines the frequency of the disaggregated series. If undefined will
use the frequency of the high frequency input series.
Precision Takes numeric values.
Method Takes values DKF, AKF Fixed, AKF Diffuse.
ML estimation Tick box for true or false.
Zero initialisation Tick box for true or false.
Truncated rho Takes numeric values.
Diffuse regression Tick box for true or false.
coefficients
Outputs High frequency series that meets the constarints of the low frequency series.
Output table also includes disaggregation errors.
Methodology Given a low frequency series y• and indicator variables at a higher frequency,
X, temporal disaggregation aims to produce a high frequency series y to that
satisfies y• = Cy for some aggregation matrix C.
The method assumes that the underlying high frequency series y satisfies the
regression equation 3.3.

y = Xβ + u, E(u) = 0, E(uu0 ) = V (3.3)

This means that the low frequency series must satisfy the regression equation 3.4.

y• = Cy = CXβ + Cu = X• β + u• , E(u• ) = 0, E(u• u0• ) = CVC0 = V•

(3.4)
JDemetra+ allows four different models for the errors u. These are:
• Wn - white noise. This is equalivalent to fitting an ordinary least squares
regression model.
• Ar1 - autoregressive model of order 1. This is equivalent to the Chow-Lin
method (Chow and Lin (1971)).
• Rw - random walk. This is equiavlent to the Fernandez method (Fernandez
(1981))
• RwAr1 - (1,1,0) ARIMA model. This is equivalent to the Litterman method
(Litterman (1983)).

10
References G.C Chow and A Lin (1971). “Best linear unbiased interpolation, distribution,
and extrapolation of time series by related series”. In: The review of Economics
and Statistics, pp. 372–375
R.B Fernandez (1981). “A methodological note on the estimation of time series”.
In: The Review of Economics and Statistics, pp. 471–476
R.B Litterman (1983). “A random walk, Markov model for the distribution of
time series”. In: Journal of Business & Economic Statistics 1.2, pp. 169–173

11