How to select regressors and specifications in Demetra ?

1
How to select regressors and specifications in
Demetra?

• N. Alpay KOÇAK
• Turkish Statistical Institute

2
An observation mechanism on TramoSeats in
Demetra

• The main issue is whether the series has a
  significant seasonal and/or calendar affect
  component?
• To check seasonality,
• seasonal chart,
• seasonal periodogram or AR spectrum
• seasonality tests part of Demetra!
• To check calendar effects, look at spectrum
  graphics of original series and residuals!

3
Checking Seasonality and Calendar Effects
4
Checking Seasonality and Calendar Effects
5
An observation mechanism on TramoSeats in
Demetra

• SEATS may change the ARIMA model identifed by
  TRAMO. This is because TRAMO aims to fit better
  to series for forecasting, but SEATS aims to
  decompose the series underlying components using
  by proper ARIMA model. Since our aims to find
  best seasonally adjusted series, this change has
  no negative effect on seasonal adjustment
  process.

6
An observation mechanism on TramoSeats

• For example, lets assume that TRAMO selected an
  ARIMA model (0,1,1)(1,0,0)12 for a monthly time
  series. This model may be evaoluated as a good
  model to forecast the series. But, SEATS can not
  design a decomposition scheme using this ARIMA
  model. Then, SEATS changes this model to
  (0,1,1)(0,1,1).
• Not only this situation, but also there are
  several case that SEATS changes the model.

7
SEATS may change the model!
8
An observation mechanism on TramoSeats

1. Significancy of calendar effects is very
   important. It may affect directly the
   specification of the model, outliers and
   estimated parameters. Default calendar variables
   (TD1, TD2, TD6, TD7), Easter, National holidays
   or user-created calendar regressors will be
   tested, and if it is not significant then removed.

9
An observation mechanism on TramoSeats

• For example, lets assume that TRAMO selected an
  ARIMA model (3,1,0)(0,1,1)12 for a monthly time
  series without TD6 calendar effect. When TD6 is
  added to the model (assume that it
  issignificant), the model is changed to
  (0,1,1)(0,1,1)12 with TD6 calendar effect.

10
An observation mechanism on TramoSeats

• TRAMO can be evaluated with some statistical
  diagnostics since it is a model-based approach.
• Significancy of ARIMA parameters estimated
• Independence of the residuals
• Normality of the residuals
• Randomness of the residuals
• Linearity of the residuals (Less importance)

11
An observation mechanism on TramoSeats

• Problems and possible solutions of those are
  given below
• Insignificant parameters ? Change ARIMA model
  specification
• Independence of the residuals (significant
  autocorrelations at first, second and seasonal
  lag) ? Check the number outliers, if it does no
  work, increase MA order
• Normality of the residuals (excessive kurtosis or
  skewness) ? Check the number outliers or check
  data span consistency
• Randomness of the residuals (too much positive or
  negative residuals) It is a sign of nonlinearity
  ? Check the number outliers or check data span
  consistency
• Linearity of the residuals (Less importance) If
  the only problem is this, leave it ! It does not
  effect the value of estimated parameters to be
  used in filter design of SEATS

12
An observation mechanism on TramoSeats

• SEATS can also be evaluated with some statistical
  diagnostics since it is a model-based approach.
• Variance and autocorrelation fuctions of the
  components
• Under or over adjustment judgement from variances
• Significant autocorrelations in seasonal lags
• Cross-correlations of components
• There is not any exact solution of those
  problem. It can depends on time span, models,
  calendar etc. But, If any of them has problem,
  try Airline model (0,1,1)(0,1,1) which is most
  useful model for a number of series.

13
An observation mechanism on TramoSeats

• Since the aim of the process is to extract all
  seasonal variations, it has to be checked that
  remaining components has no seasonality, anymore.
• Residual seasonality test results
• Spectral analysis

14
An observation mechanism on TramoSeats

• Revision policy has to be determined to publish
  the data officially.
• Revision history
• Sliding spans