Seasonal Adjustment in Times of Strong Economic Changes Jens - - PowerPoint PPT Presentation

*This presentation represents the author’s personal opinions and does not necessarily reflect the *views of the Deutsche Bundesbank or its staff.

Seasonal Adjustment in Times of Strong Economic Changes

Jens Mehrhoff* Deutsche Bundesbank 6th Eurostat Colloquium on Modern Tools for Business Cycle Analysis Luxembourg, 26-29 September 2010

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 2

• 1. General principles of seasonal

adjustment

❙ Seasonally adjusted data have proven helpful in monitoring the current economic situation, particularly in order to gain information about news and turning points at an early stage. ❙ However, abrupt sharp economic movements, such as the recent financial and economic crisis, also affect the calculation of seasonally adjusted data. ❙ The interplay of fact and diagnosis will be examined in this presentation using data on Argentine currency in circulation. ❙ The different revision policies contained in the ESS Guidelines on Seasonal Adjustment will be used with the X-12-ARIMA and TRAMO/SEATS methods.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 3

• 1. General principles of seasonal

adjustment

❙ According to these guidelines, the purpose of seasonal adjustment is to filter

• ut the usual seasonal fluctuations, i.e. those movements that recur with

similar intensity in the same season each year (item 0). ❙ Hence, this implies that unusual movements that are readily understandable in economic terms will continue to be visible. Therefore, it is necessary that these unusual movements are treated as outliers and thus not attributed to seasonal factors (item 1.4). ❙ Given that a crisis does not occur year after year and the persistence of the conditions responsible for seasonality, one could either use appropriate

• utlier variables or forecast seasonal/calendar factors in these times. Both

approaches guarantee that the full effects of the crisis remain visible in seasonally adjusted data (item 3.2).

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 4

• 2. The case of Argentine currency in

circulation

❙ The 2008/2009 financial and economic crisis is not long enough ago to allow the evaluation of its final effects on seasonality and seasonal adjustment. For this reason, the empirical example chosen here is that of Argentine currency in circulation, which was affected by the 2001/2002 Argentine crisis. ❙ Three different regimes can be identified in the time series. The first of these is the time before the Argentine crisis, in particular the period from January 1992 to November 2001. The height of the crisis covers the months from December 2001 to May 2002 – this view is supported by the outlier identification of RegARIMA modelling. Finally, from June 2002 to December 2007 a new development of the time series can be observed.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 5

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 6

• 3. Seasonal adjustment before the

crisis

❙ A more detailed inspection of unadjusted data reveals a certain pattern of seasonality in currency in circulation. There are peaks in December/January followed by troughs until July, when the value of currency in circulation peaks again before declining until December. Looking at reasons behind this pattern, it is the extra half-wage payments received in July and December/January which are the prime cause of the seasonality. ❙ As the time series is a monthly stock series of daily averages it shows working day effects. For the sake of parsimony and in order to avoid multicollinearity, two regressors are built. The “Monday” regressor counts the number of Mondays minus the number of Fridays. Tuesdays, Wednesdays and Thursdays are put together in the “weekday” regressor, with their number again being measured in deviation from the number of

• Fridays. Additionally, a special Christmas regressor is set up which counts

the number of working days within 15 calendar days before Christmas.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 7

• 3. Seasonal adjustment before the

crisis

❙ Prior to estimation, unadjusted data are transformed into natural logarithms. In addition to the aforementioned three calendar regressors, the baseline model (as of November 2001) consists of a constant term and a level shift for August 2001. The ARIMA model is seasonally and non-seasonally integrated with both a seasonal and non-seasonal AR term and a seasonal MA term, i.e. ln(1 1 0)(1 1 1)12. ❙ Following the Argentine practice, the monthly-specific seasonal smoothing filters in the seasonal adjustment core of X-12-ARIMA, hereinafter referred to as X-12, are set to 3 × 9. By contrast, the decomposition with SEATS is based solely on signal extraction from TRAMO's RegARIMA model. Estimation is performed with an experimental version of the hybrid program X-13ARIMA-SEATS.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 8

• 3. Seasonal adjustment before the

crisis

1.17 0.253 0.30 Christmas regressor –1.52 0.029 –0.04 “Weekday” regressor 1.90 0.091 0.17 “Monday” regressor t-value Standard error Parameter estimate Variable Table 1: Estimated semi-elasticities of the calendar regressors

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 9

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 10

• 4. Seasonal adjustment during the

crisis

❙ At the end of 2001, one of the worst political, economic and social crises in Argentina broke out. In order to avoid a bank run, the government de facto froze all bank accounts in December 2001. The most striking feature of currency in circulation is the change in the exchange rate regime beginning in January 2002. At the beginning of 2002, the national government declared that it would halt repayments on its national debt and strongly devalued the peso, adopting a managed float regime. ❙ How should one deal with such a situation in the context of seasonal adjustment? The ESS Guidelines suggest two viable alternatives for seasonally adjusting new data. Either partial concurrent adjustment where the extraordinary effects of the crisis are accounted for by the introduction of

• utlier variables or the use of forecast seasonal and calendar factors in

combination with internal checks, i.e. controlled current adjustment.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 11

• 4. Seasonal adjustment during the

crisis

❙ In what follows, revisions from partial concurrent adjustment are analysed with and without outliers in both RegARIMA modelling and – for X-12 only – the seasonal adjustment (SA) core. The model is identified based on November 2001 data. ❙ Provided RegARIMA outlier variables are introduced into the model, the following will be used which are identified both in real time and ex post (the corresponding t-values are at least about four in absolute terms throughout): ❙ a level shift in December 2001, ❙ an additive outlier in January 2002, ❙ an additive outlier along with a level shift in February 2002 and ❙ a level shift in May 2002.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 12

• 4. Seasonal adjustment during the

crisis

❙ While the model parameters are re-estimated, the model specification is kept constant. ❙ Revisions are calculated with the concurrent estimate as target, i.e. revision = later estimate/first estimate – 1, showing how much a given adjustment changes when adding more data. Old unadjusted data remain unchanged, i.e. revisions are not calculated in real time but much like the automatic History procedure, known from X-12-ARIMA.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 13

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 14

• 4. Seasonal adjustment during the

crisis

❙ When outliers are correctly specified in RegARIMA modelling, the revisions from the X-12 method and the SEATS method are similar, although X-12 produces somewhat lower revisions than SEATS. Conversely, if RegARIMA

• utliers are neglected, revisions skyrocket. The need for outlier modelling

becomes even more obvious when considering SEATS without RegARIMA

• utliers: the estimated model is of little use, if at all, and thus SEATS fails to

estimate the seasonal factor reliably. In some cases, SEATS is unable to admissibly decompose the model and has to replace it with a decomposable

• ne.

❙ As regards revisions, RegARIMA outlier modelling is more important than extreme value detection in X-12's seasonal adjustment core. Nonetheless, employment of seasonal adjustment core outliers lowers revisions, especially if RegARIMA outliers are disregarded. Eventually, there is strong evidence that seasonality is present even in times of crisis.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 15

• 5. Seasonal adjustment during the

crisis revisited

❙ After mid-2002 first signs of a recovery of the Argentine economy became visible. ❙ The RegARIMA model is reviewed after 12 months with the first release of November 2002 data and again four years later with the first November 2006 data release. While no additional outliers are identified after the crisis, the model is changed in 2002 to ln(0 1 2)(0 1 1)12 from ln(1 1 0)(1 1 1)12, i.e. the non-seasonal and seasonal AR terms, implying an MA(∞) structure, are replaced by two non-seasonal MA terms. In 2006, the model is again revisited to ln(1 1 1)(1 1 1)12 – a mixture of non-seasonal and seasonal AR and MA terms, now without a constant. The X-12/SEATS specification remains

• unchanged. However, if outliers are neglected right from the beginning, the

models read ln(0 1 2)(0 1 0)12 and ln(1 1 0)(0 1 1)12, respectively, both without the constant term.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 16

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 17

• 5. Seasonal adjustment during the

crisis revisited

❙ While the partial concurrent adjustment data vintages that make use of RegARIMA outlier modelling closely resemble the development of controlled current adjustment figures, those estimated without RegARIMA outliers might send the wrong message during and shortly after the crisis. The gradients differ considerably between the vintages and to the forecast. This causes the above strong revisions. What is even more problematic is that, in addition, the direction of change is revised in several cases. ❙ However, a notable result is that, in the long run, the results from either adjustment are similar no matter whether outliers are modelled or not. In spite of this, seasonal adjustment without outlier modelling in the short run produces results that are remarkably different from the long run.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 18

• 5. Seasonal adjustment during the

crisis revisited

❙ There are at least two explanations for why the short-run results without

• utlier modelling are well apart from the final results, while those with outlier

modelling are fairly close even in the short run. ❙ Firstly, RegARIMA estimates and forecasts are biased as regards the seasonality of the series if outliers are not modelled in the short run. This has a greater impact on SEATS than on X-12 because the decomposition of SEATS is based on RegARIMA modelling alone. ❙ Secondly, in the short run, the effects of the crisis are partially allocated to the seasonal factors rather than to the trend-cycle or the irregular component, i.e. they do not remain fully visible in seasonally adjusted data. This misperception diminishes only in the long run when sufficient “normal” data before and after the crisis become available.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 19

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 20

• 5. Seasonal adjustment during the

crisis revisited

❙ What is striking is that revisions of partial concurrent adjustment with RegARIMA outliers and of controlled current adjustment for both X-12 and SEATS remain low shortly and long after the crisis. ❙ Diametrically, revisions of X-12 and SEATS without RegARIMA outliers increase shortly after the crisis from an already high level and, though decreasing, remain high long after the crisis. This effect can be attenuated

• nly if at least X-12's seasonal adjustment core outlier identification is

utilised.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 21

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 22

• 5. Seasonal adjustment during the

crisis revisited

❙ As the legislation governing the extra half-wage payments remained unchanged during the entire period under analysis, no significant changes should be observed in the seasonality pattern of the series. ❙ It will be examined whether the seasonality experienced changes after the crisis or whether, on the contrary, the seasonal pattern is similar to that

• bserved before the crisis, meaning that seasonal adjustment should be

carried out for the entire period. It turns out that seasonality extends beyond the crisis with the aforementioned peaks and troughs still visible after the crisis despite the strong growth of the series.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 23

• 6. Implications for seasonal adjustment

❙ The results show that, as long as the reasons for seasonality continue to exist, seasonal adjustment makes sense. Furthermore, it is important to identify outliers, as otherwise the estimate of seasonal factors is distorted, resulting in major revisions. It should be emphasised that outlier modelling at the current end of the time series is a crucial issue. The results point to largely different estimates in the short run between seasonal adjustment with and without outliers. But, as time passes, these differences fade away and the results become more similar in the longer run. In the end, the preliminary results obtained with outlier modelling during the crisis are close to the final results, no matter whether outliers are identified or not in the long run. Then again, seasonal adjustment in the short run without outlier treatment is remarkably distinct from the final results. This fact makes a good case for

• utlier modelling right from the beginning of a crisis.

Luxembourg, 26-29 September 2010 Seasonal Adjustment in Times of Strong Economic Changes 24

• 6. Implications for seasonal adjustment

❙ Revisions from both adjustment techniques are virtually the same if outliers are taken into account. Controlled current adjustment performs as well as partial concurrent adjustment. This justifies the classification of both methods as best alternatives in the ESS Guidelines. Hence, seasonal adjustment reveals important information for short-term business cycle analysis as it points out the news in the time series – even in times of strong economic changes.