11.06.2013

Sylwia Grudkowska, NBP

Demetra+ User Manual – the final version

This document summarises the last actions that have been taken to deliver the final version of the Demetra+ User Manual and next steps that are necessary to release the document. SASG members are invited to discuss it.

  1. An overview

During last SASG meeting (16.11.2012) it has been agreed that the final version of the Demetra+ User Manual should be released as soon as possible. For this reason, the review the English wording wasdone by Mark Boxall (December 2012). The remarksfromM. Boxall, except for one, are already incorporated in the document. M. Boxall suggests using the term "process" (noun) instead of "processing" (adverb), because"…this seems strange to a native speaker but maybe it is the correct terminology in this case". The term "processing" refers to the functionalities in Demetra+ GUI (e.g.: "single-processing"), therefore the change in the document cannot be made, unless the relevant modification in the application is performed.

Then, the US Census Bureau and Agustin Maravall have been asked for validation of the Manual (February 2013). US Census Bureau delivered its remarks (March 2013). No reply has been received from Agustin Maravall yet. Almost all comments and remarks from US Census Bureau has been dealt with. The few thathaven’t been handled are listed in the next section of this document.

Finally, during last SAUG meeting (16.04.2013) it has been requested to add to the Manual the explanation of the seasonality tests results. Some SAUG members stated that they need a guidance in situation when seasonality tests shows (seemingly) contradictory results. Therefore SASG members are kindly asked to discuss how to handle this request in the User Manual. The exemplary screens that could be explained in the Manual are presented in the third part of this document.

  1. Feedback from US Census Bureau

Remark no. 1

US Census Bureau points that "…the official name of the program is X-12-ARIMA, not X12. When referring to the X-12-ARIMA program distributed by the Census Bureau, you should use "X-12-ARIMA". When referring to the constructs within Demetra+, you can use "X12". Similarly, it is X-11, not X11…".

I suggest to add the statement (in bold) after the following extract from User Manual (p. 9):

Demetra+ uses notation "X12","Arima" and "TramoSeats" instead of "X-12-ARIMA", "ARIMA" and "TRAMO/SEATS" respectively. This notation is also used in this Manual. However, when the references to the original programs are made, the notation "TRAMO/SEATS", "X-12-ARIMA " and "X-11" is used.

Remark no. 2

US Census Bureau recommends having an experienced technical editor go over the document to make the necessary corrections.

Remark no. 3

US Census Bureau suggests to add the following explanation for S-I ratios: "changes from seasonal increases to seasonal decreases within a calendar month that is not very seasonal (small seasonal factors) would not be alarming, but such changes in several calendar months with substantial seasonality certainly would be (unless there was a compelling economic explanation)".

I suggest the following changes (in bold) to the Manual (p.122):

Changes in seasonality over time are acceptable unless there is a noticeable change from below to above the overall mean(or vice versa),especially when they concern several calendar periods with substantial seasonality and no compelling economic explanation for this exists. However, in case of a calendar period that is not very seasonal (small seasonal factors) such changes would not be alarming.

Remark no. 4

US Census Bureau suggest to add the following text: "These problems affect all methods, including model-based methods. Model based methods implicitly forecast the missing data for symmetric filters, but the result is an asymmetric filter."after the following extract from the Manual (p. 235):

"…Moving averages, which are weighted averages of a moving span of a time series (see hereafter), have two important deficiencies:

  • They are not resistant and might be deeply impacted by outliers;
  • Unless forecasts and backcasts are available, different moving averages are needed for the ends of the series.

These drawbacks adversely affect the X11 output and stimulated the development of this method. The United States Census Bureau program X12 includes all the capabilities of X11 and a pre-processing module that fits RegArima models[1]. Therefore, X12 is divided in two parts (see Figure 1) "…

Remark no. 5

US Census Bureau suggests using the name"moving seasonality test" instead of "evolutive seasonality test". This change in the documentation requires the modification of the software.

Remark no. 6

US Census Bureau suggests that "in the Visual spectral analysis text, it should be mentioned that is k/120 except at trading day frequencies. (i.e. different from the Fourier frequencies of the periodogram of Demetra+)".

Below there is an extract form the Manual that concerns Visual spectral analysis (p.259):

"Criterion of "visual significance" is based on the range of the values,

where:

;

;

– value of autoregressive spectrum estimator.

The particular value is considered to be visually significant if at a trading day or at a seasonal frequency (other than the seasonal frequency = 0.5) is above the median of the plotted values of and is larger than both neighbouring values and by at least 6/52 times the range .

For a given series , which may contain missing values, the periodogram is computed as follows:

In the first step, the series is standardized:

,

In a second step we compute at the so-called Fourier frequencies , which are the values of the periodogram:

,

where is the number of non-missing values.

Under the white noise assumption, the values of the periodogram should be asymptotically distributed as a Chi-square with 2 degrees of freedom.

The default frequency for trading days is computed as follows (for series of quarterly series):

,

where:

.

Other frequencies correspond to trading days frequencies:

  • For monthly series, 2.714 ( default = 2.188);
  • For quarterly series, 1.292, 1.850, 2.128 (default = 0.280) ".
  1. Proposal for interpretation of the seasonality tests’ results from Demetra+

Demetra+ displays the results of seasonality tests, namely, Friedman test, Kruskal-Wallis test, test for presence of identifiable seasonality, combined seasonality test and residual seasonality test, for both X-12-ARIMA and Tramo/Seats. Originally, these tests have been used by X-12-ARIMA method. Therefore, the question is if the interpretation of these tests should be the same for both methods.

Below there are extracts from two documents concerning the existence of seasonality in time series.

  1. From the ONS Guide to Seasonal Adjustment with X-12-ARIMA:

“(…) the most relevant test that checks the presence of seasonality is "Combined seasonality test for the presence of identifiable seasonality". In particular, one of the following statements will always appear:

1. IDENTIFIABLE SEASONALITY PRESENT

2. IDENTIFIABLE SEASONALITY PROBABLY NOT PRESENT

3. IDENTIFIABLE SEASONALITY NOT PRESENT

It is recommended that a series is adjusted in the first two cases and not adjusted in the last one. However there are two cases where one might need to deviate from this practice:

Marginal seasonality - When a series is only marginally seasonal, the criterion might swing from "non seasonal" to "seasonal" as new data become available. It might however not be desirable to change the policy of seasonally adjusting or not all the time, as this causes revisions to the published seasonally adjusted series and revisions back (i.e., when a series is deemed non-seasonal its seasonally adjusted version is same as the unadjusted, but when one actually adjusts the series the adjusted version is a different series)…

Indirect Seasonal Adjustment - For many data sets it is quite common to do the seasonal adjustment at a low level of aggregation an add the seasonally adjusted series together to get the seasonally adjusted version of an aggregate series. The problem with this approach is that there is often residual seasonality in the aggregate series. If one attempts to adjust the (indirectly) seasonally adjusted aggregate series, the IDS test often shows that seasonality is present. The cause of this problem is that the more volatile the irregular component is, the more difficult for any seasonality in the series to be detected. As the irregular component is more volatile for the lower level series, some of them are deemed non-seasonal. Although this is acceptable for the individual series, when added together their suppressed seasonal components become significant.(…) The decision as to which component series should be seasonally adjusted can be based on IDS test. Whereas the general IDS test is based on three different tests (which are listed just after table D8):

  • Test for seasonality assuming stability;
  • Nonparametric test for the presence of seasonality assuming stability;
  • Moving seasonality test;

the alternative test consists of seasonally adjusting a series if just one of the first two tests passes.(…)

  1. From A. Maravall (2011),Seasonality tests and Automatic Model Identification in Tramo-Seats:

(…) the final decision concerning the presence of seasonality is made by the AMI algorithm, namely, by the ARIMA model it yields and by how this model is decomposed by SEATS. When this decomposition yields a seasonal component, the series will be seasonally adjusted; otherwise not.(…)

Screens:

Screen no. 1 – parametric an non-parametric stable seasonality tests passes, moving seasonality test passes, combined seasonality test fails.

Screen no. 2 – parametric stable seasonality test fails at 0,05 level, non-parametric stable seasonality test passes on the 0,01 level.

Screen no 3 - The presence of residual seasonality.

From the Guidelines on SA: (…)The presence of residual seasonality, as well as over-smoothing, is concrete risks which could negatively affect the interpretation of seasonally adjusted data (…).

1

[1]The current official seasonal adjustment program for the U. S. Census Bureau is X-13ARIMA-SEARS that performs Seats seasonal adjustments. For more information see Monsell(2007) and the X-13ARIMA-SEATS Reference Manual (2013).