Seasonal Adjustment User Group meeting
held at Eurostat, in Luxemburg
on the 16 of April 2032 (10h00 to 17h30), Room A2/045
Minutes and to-do-list
Presidency: Dominique Ladiray (INSEE)
Secretariat: Dario Buono (ESTAT), Faiz Alsuhail (FI)
Logistics:
Web: Agenda and presentations available here
Participants: see bullet point 11
Document history:
Version / Release Date / Author / Reviewed By1.0 / 17/04/2013 / Faiz Alsuhail (FI)
2.0 / 24/24/2013 / Dario Buono (ESTAT)
Table of Content
Seasonal Adjustment User Group meeting 1
held at Eurostat, in Luxemburg 1
on the 16 of April 2032 (10h00 to 17h30), Room A2/045 1
Minutes and to-do-list 1
Table of Content 2
(1) Agenda approvaland review of previous minutes and to do listand calendar 3
(2) Release of JDemetra+ (version 1.2.1) 4
(3) Show and tell, (Sylwia Grudkowska, National Bank of Poland) 5
(3) Show and tell, (Enrico Enfante, Eurostat) 6
(4) Show and tell (Niccolo Massarelli, Eurostat) 6
(5) Show and tell (Veronique Elter, Statistics Luxembourg) 7
(6) Show and tell (Manca Golmajer, Statistics Slovakia) 8
(7) Show and tell (Faiz Alsuhail Statistics Finland) 8
(10) Future of methodology helpdesk, forum and FAQs 9
(11) Plug-ins for JDEMETRA+: econometric tools 9
(12) Quality ofdirect/indirect approach, (Marcus Scheiblecker, WIFO) 9
(13) Progress in preparation of J-Demetra+ User Manual (Sylwia Grudkowska, NBP) 10
(14) Contribution of SAUG members tofuture tasks forJDEMETRA+ 10
(15) AOB 11
(16) Next meetings to be held in 2013 11
(11) Participants 13
(1) Agenda approvaland review of previous minutes and to do listand calendar
Overview
Dominique Ladiray (INSEE) opened the meeting and welcomed everyone, in particular the new members of the SAUG, to Luxembourg.
The agenda was approved.
The previous list of tasks was gone through. The following table summarizes the status of the previously agreed tasks.
Task / Person responsible / Updated statusPresent a proposal about the list quality measures to be added as default options to J-Demetra+. / G. L. Mazzi (present), SAUG (comment) / Mazzi pointed out that the answer to the question can be found from the ESS Guidelines.
Deliver the cruncher and the user-defined regression variables to J-Demetra+. / J. Palate / The cruncher has been delivered. The user defined variables will be made available in the frozen version of J-Demetra+ by June.
Include the option of multivariate benchmarking in J-Demetra and provide it with a graphical interface. (To be agreed also at the SASG) / J. Palate / Closed.
Propose how to treat outliers in the direct vs. indirect seasonal adjustment. / M. Scheiblecker / Closed (see presentation in point 12).
Put the e-learning course to the new CROS/ESSnet-portal / P. Rey del Castillo / Still ongoing.
Inform the members of the TF on revisions of the ESS Guidelines on SA about the metadata J-Demetra+ produces automatically. / J. Palate / Still ongoing.
Inform the SASG about the need for 1) resources 2) governance and 3) a road-map for the future of J-Demetra+. / F. Alsuhail, D. Buono / Closed.
Prepare the User Manual for J-Demetra+. / S. Grudkowska, D. Ladiray, A. Kocak, and E. Infante, V.Elter / The task is ongoing (see point 13).
Co-ordinate the work between J-Demetra+ and Ecotrim, especially when it comes to univarite benchmarking. / D. Buono. / Still ongoing.
Test and document the use of X12 in J-Demetra+. / D. Ladiray / Still ongoing.
Test and document the use of calendar regressors in J-Demetra+. / D. Ladiray, R. Soares. / Closed. A presentation is available through CROS.
Test and document the use of J-Demetra+ with simulated data. / A. Kocak / Closed (see point 9).
Test and document the use of Tramo/Seats in J-Demetra+ and Demetra 2.2 / F. Alsuhail / Closed (see point 8).
Spread info on the progress of the group and promote knowledge of the software. / SAUG / Still ongoing.
Provide background on calendar adjustment and the treatment of half-days. / J. Palate, D. Ladiray. / Still ongoing.
Co-ordinate the work between the SAUG and the Ecotrim group. / D. Buono / Still ongoing.
(2) Release of JDemetra+ (version 1.2.1)
Overview
Jean Palate (NBB) presented the main improvements in the latest version of J-Demetra+ (1.2.1). Information that is more detailed can be found from his presentation.
He underlined, that there are very few differences between versions 1.2.0 and 1.2.1: mainly bugs have been corrected. Palate said the there are still many bugs in the software that have to be corrected.
Palate covered issues concerning data providers as well as the some methodological and IT questions related to the software.
In comparison with the previous version 1.1.0 the following methodological improvements have been introduced:
X11 diagnostics have been completed, the methodology of the calendar regressors has been documented and the standard deviations of the ARIMA parametes are now computed correctly..
Calendar and user defined variables are included in J-Demetra+ but not yet in the cruncher.
Palate pointed out that the estimation of the parameters of complex ARIMA models[1] may produce different results in comparison with other software. This is due to the fact that the likelihood function may have several local maxima. As J-Demetra+ uses a different optimisation algorithm one might not always find the same maximum. The solution is depending on the starting point of the optimisation procedure. This challenge is tackled by allowing several starting points in J-Demetra+ 1.2.1.
Palate gave an empirical example where the original Tramo/Seats and X12[2] and J-Demetra+ produce parameter estimates that are significantly different for a fixed model. There were also differences in the estimates of the other parameters (outliers, trading day regressors), yet smaller ones.
Discussion
Most of the discussion was around the ML estimation procedure and Palate noticed that different software produce different parameter estimates for given, fixed models. The differences in parameters may lead to very different seasonally adjusted results as Palate example illustrated. He underlined that only the optimization algorithm in the ML estimation procedure is different: otherwise, the methods are the same.
However, Palate pointed out, that for simple ARIMA-models (Airline etc.) you usually get similar parameter estimates and seasonally adjusted time series. The more complicated the models are, the more different the results. However, Palate underlined that the challenge of different results applies only for complex time series models. For about 80% of the time series the models are simple enough to produce similar results.
Some participants felt that the differences in parameter estimates is not a problem of J-Demetra+ but a more general issue of time series analysis and model selection.
It was felt that the software need further testing and that there should a frozen version of J-Demetra+. This version shall not include the latest improvements of Tramo/Seats. This software was agreed to be delivered by the end of May.
(3) Show and tell, (Sylwia Grudkowska, National Bank of Poland)
Overview
Sylwia Grudkowska presented her findings on the use of Demetra+. Her aim was to develop the semi-automatic algorithm for detecting Ramps that enhance the results from the Automatic Model Identification procedure of Tramo/Seats.
Grudkowskas main finding was that using ramp effect instead of level shift or transitory changes can improve the seasonal adjustment from a diagnostic perspective. She had a look at diagnostics of the residuals and the mean square error of the out of sample forecast.
Discussion
The SAUG felt that a tool to detect ramps can be helpful. The challenge, however, lies in the initial choice of the (timing of the) ramp effect.
Some members felt that the interpretation and properties of the seasonally adjusted time series, e.g.stability of growth rates and turning points, are more relevant to the users than model diagnostic. Hence, intuitive considerations are also needed when judging the performance of seasonal adjustment.
(3) Show and tell, (Enrico Infante, Eurostat)
Overview
Enrico Infante gave a short presentation which was based on his experience on seasonally adjusting the household saving rate and the household investment rate.
He told that Demetra (v. 2.2) was used until July 2012, since then some statistics have been seasonally adjusted with Demetra+ (.net version). J-Demetra+ has been tested and it is supposed to become the leading seasonal adjustment tool in the future.
Discussion
So far some testing of J-Demetra+ has been done and results have been compared to those obtained by Demetra+. Infante said, that there are some differences in the seasonally adjusted time series that can be confusing from the users’ point of view.
The switch of software will cause some revisions to the resulting time series. Hence, a proper time for the switch would be during the annual review of the seasonal adjustment models.
In the discussion it was underlined that J-Demetra+ and Demetra+ include different versions of Tramo/Seats and therefore results of these two software will differ.
Infante found the new tool useful but not fully ready for production.
(4) Show and tell (Nicola Massarelli, Eurostat)
Overview
Massarelli told that he is a user of seasonal adjustment but not a methodologist. Hence he will ask questions instead of providing answers.
His presentation was about his experiences on J-Demetra+, even though Demetra 2.2 is still used in several domains at Eurostat.
The following feedback was raised in his presentation: the use of seasonality test seemed to be confusing and more information on the interpretation of the results is needed. Some things about the automatic model identification seemed to be confusing for the users, e.g. how the software chooses between a seasonal and non-seasonal model[3] or how the log/non-log test works.
He has made some comparison between Demetra 2.2, Demetra+ and J-Demetra+. According to his findings, J-Demetra+ and Demetra+ give quality warning for results that are accepted by Demetra 2.2.
Another difference between different software, that Massarelli had paid attention to, was that Demetra 2.2 seems to give smoother seasonally adjusted time series than Demetra+ or J-Demetra+.
Discussion
Much discussion was around the smoothness and stability of the results. There were different views on whether a smooth seasonally adjusted time series is a good thing or not. From the users’ point of view, a smooth series is welcomed though by definition the seasonally adjusted series contains the irregular and the outliers.
(5) Show and tell (Veronique Elter, Statistics Luxembourg)
Overview and discussion
Veronique Elter shared the experiences she had gained from using X13 and J-Demetra+ to compile the SA of the QNA. A detailed discussion can be found from her presentation.
Her main findings were that the use of a weighted calendar didn’t give different results for the time series she had studied.
Another observation was that J-Demetra+ seemed to crash often. Palate said that it is because of insufficient memory available on the local machine. This problem has been identified and will be fixed in the new version.
Elter provided the SAUG with some user feedback: she noticed that when you refresh your current adjustment you lose the national calendar settings saved previously. Also the Excel output with the option of printing results by series didn’t to work.
Elter suggested to create a default folder for the output, where you could specify just once the path of the output file, the structure of the output and the series you want in the output; now you still need to redefine this for every output. The SAUG took note of the feedback.
(6) Show and tell (Manca Golmajer, Statistics Slovakia)
Overview
Manca Golmajer gave an overview of her experiences from using Demetra+ and J-Demetra+ in Statistics Slovakia. She also provided the SAUG with some feedback from using the software.
Golmajer finds Demetra+ very user-friendly: the user-interface is well organized and many colours are used. The tool also includes plenty of tests and diagnostics that increase the understanding of the seasonal adjustment process. For example, the number of trading day regressors can be modified with the help of the information provided by the joint F-test. In addition, the tool is very fast and efficient.
The chart scale in graphs is something Golmajer felt misleading. Moreover, in her opinion, it is not easy to compare different models for the same time series in Demetra+. It has also been criticised that the maximum order of seasonal autoregressive polynomial in SARIMA model is one in Demetra+ but in Demetra 2.04 it is two. SAUG took note of the feedback.
Discussion
Golmajer said the Statistics Slovakia has organized training and produced documentation related to the new software tool. However, the Eurostat help desk is and will be important in the future.
(7) Show and tell (Faiz Alsuhail Statistics Finland)
Overview and discussion
Alsuhail had made a test where he compared the output produced by Demetra 2.2 and J-Demetra 1.2.0. He used Tramo/Seats and identical model settings, including identical ARIMA models, TD regressors, outliers and series transformation, for a handful of series.
The study showed that despite using identical model setting the tow software produce different results. The levels of the seasonally adjusted data were close to each other whereas the growth rates were different.
(8) Show and tell (TURAÇ YAVUZ,TurkStat)
Overview and discussion
Yavuz had done a similar study as Augustin Maravall in 2012. The idea was to check the performance of the automatic model identification of J-Demetra+. A comparison between the performance of automatic model identification of J-Demetra+ and TSW+ was also made.
For the identification of the correct time series model for a seasonal model (other than Airline) J-Demetra+ finds the correct model in 45% of the cases when looking at series with 120 observations. When there are 240 observations the software finds the correct model in 71% of the cases. For X13 these percentages are 35 and 47 respectively.
The study showed that J-Demetra+ is remarkably faster than other SA softwares (TSW, TSW+): 95 000 series were processed by using the multiprocessing option in two days (16 hours). Average process time of 8000 series with 120 observation is about 29 seconds for TRAMO-SEATS method (45 seconds for X13).
(10) Future of methodology helpdesk, forum and FAQs
Overview and discussion
Dario Buono reminded the SAUG that there is an e-mail address where you can send questions about seasonal adjustment software (Demetra+ and J-Demetra+). After the question is sent, Buono does the filtering and either answers it or distributes the questions to other people. The help-desk will be also maintained in the future but it will be tried to be streamlined promoting the use of other type of support.