1 anna ciammola, claudia cicconi francesca di palma istat - italy workshop on methodological issues...
TRANSCRIPT
1
Anna Ciammola, Claudia Cicconi
Francesca Di Palma
ISTAT - Italy
Workshop on methodological issues in Seasonal Adjustment
Luxembourg, 6 March 2012
Does the order matter?
On temporal aggregation and seasonal adjustment
Luxembourg, 6 March 2012 2/19
Outline of the presentation
Statement of the problemOur experimentResultsFinal remarks
Luxembourg, 6 March 2012 3/19
Statement of the problem Sometimes seasonally adjusted data, required at
quarterly frequency, can be derived from raw data available at monthly frequency
Two possible approaches for seasonal adjustment (SA)1.On monthly data (SA first, quarterly aggregation later)
2.On quarterly data (quarterly aggregation first, SA later)
The minimization of revisions of SA series can represent a criterion to choose between the two alternatives
Luxembourg, 6 March 2012 4/19
QNA: the framework QNA are derived applying temporal
disaggregation techniques with related indicators to annual dataChow-Lin (1971) and, occasionally, Fernandez (1981)
Quarterly unadjusted, working-day adjusted (WDA) and seasonally adjusted (SA) data are derived through three separate disaggregation processes1.Unadjusted NA annual data and quarterly short-term
indicators Unadjusted QNA2.WDA annual data and quarterly WDA indicators
WDA QNA Monthly indicators: WDA indicators are derived from Tramo at monthly frequency and then quarterly aggregated
Luxembourg, 6 March 2012 5/19
QNA: the framework
3. WDA annual data (as in step 2) and quarterly SA indicators SA QNA Quarterly indicators: SA indicators are derived from Tramo-Seats at quarterly frequency (processing quarterly WDA data)
Major domains where monthly reference indicators are available
Industrial production and foreign trade
Luxembourg, 6 March 2012 6/19
Aim of the analysis To investigate whether the order of temporal aggregation
(TA) and SA matters in terms of revisions of seasonally adjusted data
Previous analysis Di Palma and Savio (2000):Theoretical properties of revisions in the model-based
decompositionEmpirical analysis implemented omitting WDA
Our contributionConsidering WDA as part of the analysisDifferent indicators on revisions estimated on a longer time
spanMore general simulation exercise and empirical results on
industrial production indicators (IPI)
Luxembourg, 6 March 2012 7/19
TA of seasonal ARIMA models
Well documented in the literature on time series Wey (1978), Geweke (1978), ……, Silvestrini, Veredas (2008)
Quarterly aggregation (QA) of monthly data Invertible ARIMA QA Invertible ARIMAThe order of the autoregressive part of the model (stationary
and non-stationary) does not changeThe order of the moving average (MA) part of the model may
change
The airline model(0,1,1)(0,1,1)12 QA (0,1,1)(0,1,1)4 The seasonal MA parameter not affected by QA
Luxembourg, 6 March 2012 8/19
Our experiment
Aim Analysis of revisions on both SA data in level and q-on-q growth rates (GR), when SA is implemented before and after QA
Exercise on simulated seriesApplication on the indicators of industrial
production
Luxembourg, 6 March 2012 9/19
Revisions and their measures Revisions
For both SA data in levels and GR The target of the revision analysis is the concurrent estimates
(SAt|t or GRt|t) How concurrent estimates change when 1, 2, 3 or 4 quarters are
added (SAt|t+step i or GRt|t+step i) Revisions computed over a 12 year span (48 iterations)
Measures on revisions of quarterly SA levels and GR Mean of revisions (MR) Mean of absolute revisions (MAR) Root mean squared revisions (RMSR)
Quarterly SA data Monthly data QA SA (hereafter Q SA) Monthly data SA QA (hereafter M SA Q)
Luxembourg, 6 March 2012 10/19
Simulation exercise (1)
1. Airline models 25 models with and Θ = {-.1, -.3, -.5, -.7, -.9} 100 monthly series for each model (22 years) QA to derive quarterly series 48 iterations for each series, adding one new
quarter (three new obs. for monthly data)
2. Issue: simulation of series on which GR can be computed Initial conditions in the data generation process
different from zero Transformation of generated time series (with initial
conditions = 0) in indices our choice• Constant • Scale factor
Luxembourg, 6 March 2012 11/19
Simulation exercise (2)
3. Tramo-Seats processing Automatic identification of the ARIMA model Computation of revisions on quarterly SA
data
Luxembourg, 6 March 2012 12/19
Results on simulated series: quarterly SA data
Θ
-0.9 -0.7 -0.5 -0.3 -0.1
1 step
-0.9 0.51 0.64 0.73 0.72 0.92
-0.7 0.56 0.69 0.82 0.92 0.96
-0.5 0.72 0.79 0.89 0.57 0.99
-0.3 0.63 0.74 0.96 0.92 1.01
-0.1 0.60 0.81 0.83 0.85 1.02
4 step
-0.9 0.60 0.80 0.87 0.81 0.95
-0.7 0.63 0.80 0.87 0.95 0.97
-0.5 0.73 0.84 0.93 0.62 0.99
-0.3 0.65 0.78 0.98 0.90 0.96
-0.1 0.68 0.84 0.86 0.87 0.95
RMSR MSAQ
RMSR QSA
Luxembourg, 6 March 2012 13/19
Results on simulated series: q-on-q GR
RMSR MSAQ
RMSR QSA
Θ
-0.9 -0.7 -0.5 -0.3 -0.1
1 step
-0.9 0.40 0.38 0.42 0.63 0.80
-0.7 0.44 0.44 0.60 0.83 0.92
-0.5 0.60 0.69 0.86 0.92 1.01
-0.3 0.56 0.62 0.90 0.97 1.10
-0.1 0.50 0.78 0.80 0.93 1.13
4 step
-0.9 0.54 0.81 0.93 0.94 0.94
-0.7 0.62 0.81 0.90 0.98 0.94
-0.5 0.72 0.85 0.95 0.95 0.96
-0.3 0.67 0.82 0.96 0.92 0.92
-0.1 0.67 0.86 0.86 0.93 0.90
Luxembourg, 6 March 2012 14/19
Empirical analysis (1)
1. Industrial production indicators Total index and 16 industrial sectors WDA data Sample: 1990 – 2011
• 1990q1-2000q1: first estimation sample• 48 iterations for each series, adding one new quarter
(three new obs. for monthly data)
2. Partial concurrent approach with some constraints At the end of the year, current model and
identification of outliers in the last 12 (4) obs. Identification of a new model in case of diagnostics
failure, non-significance/instability of parameters
Luxembourg, 6 March 2012 15/19
Empirical analysis (2)
3. Current processing
Reg-Arima model fixed and parameter estimation run every quarter On monthly data On quarterly data
4. Computation of revisions on quarterly SA data and growth rates
Luxembourg, 6 March 2012 16/19
Results on real data: quarterly SA data
0 0.5 1 1.5 2 2.50
0.5
1
1.5
2
2.5One-step
Q-->SA
M--
>S
A--
>Q
0 0.5 1 1.5 2 2.50
0.5
1
1.5
2
2.5Two-step
Q-->SA
M--
>S
A--
>Q
0 0.5 1 1.5 2 2.50
0.5
1
1.5
2
2.5Three-step
Q-->SA
M--
>S
A--
>Q
0 0.5 1 1.5 2 2.50
0.5
1
1.5
2
2.5Four-step
Q-->SA
M--
>S
A--
>Q
RMSR - (SAt|t + step i − SAt|t) / SAt|t
Circle size is proportional to the sectorial weight (the biggest circle represents the total industrial index)
Luxembourg, 6 March 2012 17/19
0 1 2 3 40
1
2
3
4One-step
Q-->SA
M--
>S
A--
>Q
0 1 2 3 40
1
2
3
4Two-step
Q-->SA
M--
>S
A--
>Q
0 1 2 3 40
1
2
3
4Three-step
Q-->SA
M--
>S
A--
>Q
0 1 2 3 40
1
2
3
4Four-step
Q-->SA
M--
>S
A--
>Q
Results on real data: q-o-q GRRMSR - (GRt|t + step i − GRt|t)
Circle size is proportional to the sectorial weight (the biggest circle represents the total industrial index)
Luxembourg, 6 March 2012 18/19
Final remarks
Results from the simulation exercise M SA Q outperforms Q SA in terms
of revisions on both SA data and growth rates, when airline model is considered with negative parameters (true sign)
This result is more clear-cut when• Both regular and seasonal MA parameters are
near the non-invertibility region• Time series are not very long (results not reported
in this presentation)
Luxembourg, 6 March 2012 19/19
Final remarks
Results from empirical analysis on IPI M SA Q slightly outperforms Q SA in
terms of revisions on both SA data and growth rates, supporting evidence from simulation
Further analysis More ARIMA models for simulations
• (1,1,0)(0,1,1)• (2,1,0)(0,1,1)
Different sample lengths Applications on other domains (e.g. foreign
trade)