TURKISH STATISTICAL INSTITUTE Turk Stat Problematic time series

TURKISH STATISTICAL INSTITUTE Turk. Stat Problematic time series and how to treat them N. Alpay KOÇAK Turkish Statistical Institute UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 1

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment of short time series • For some series that are too short to be seasonally adjusted using either TRAMO-SEATS or X-12 -ARIMA, it is possible to adjust them using alternative, less standard, procedures. For series that are long enough to run X-12 -ARIMA or TRAMO-SEATS but remain quite short (3 -7 years), some instability problems can appear. Several empirical comparisons have been done to investigate the relative performance of X-12 -ARIMA and TRAMOSEATS on short time series. • As a general rule when the series are shorter than seven years, the specification of the parameters used for pretreatment and seasonal adjustment has to be checked more often (e. g. twice a year in order to deal with the higher degree of instability of such series). UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 2

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment of short time series • Series shorter than 3 years should not be seasonally adjusted. Seasonal adjustment of short time series (3 -7 years) should be performed, whenever possible, with standard tools. • Moreover back-recalculated time series (even non-official) should be used to extend the sample and stabilise seasonal adjustment when they are reliable for estimating a seasonal component. • Simulations on relative performances of the existing standard tools for adjustment of short series should be carried out. Users should be informed of the greater instability of seasonally adjusted data for short series, and of methods used. • Clear publication policy rules should be defined. The settings and parameters for seasonal adjustment should be checked more than once per year. UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 3

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment of short time series (Proposal) • Use TRAMO&SEATS for the series longer than 3 years! • 1) Find the best model at the start of the year, 2) fix the model specification and parameters values, and position of outliers and type of the calendar effects, 3) Check parameter stability twice a year for entire series! • Do not forget that Airline model (0, 1, 1) may be the most representative model for a series! UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 4

Turk. Stat TURKISH STATISTICAL INSTITUTE Estimation Span Level-Log Th(1) Bth(1) Weekdays Effect with Mongolia Fixed Days [I-2000 : IV-2003] Log -0. 53 0. 47 -0. 0075 0. 0687 [I-2000 : I-2004] Log -0. 57 0. 38 -0. 0061 0. 0684 [I-2000 : II-2004] Log -0. 62 0. 58 -0. 0075 0. 0621 [I-2000 : III-2004] Log -0. 57 0. 19 -0. 0053 0. 0708 [I-2000 : IV-2004] Log -0. 57 0. 15 -0. 0048 0. 0722 [I-2000 : I-2005] Log -0. 70 0. 28 -0. 0091 0. 0428 [I-2000 : II-2005] Log -0. 63 0. 27 -0. 0090 0. 0465 [I-2000 : III-2005] Log -0. 62 0. 29 -0. 0090 0. 0474 [I-2000 : IV-2005] Log -0. 79 0. 43 -0. 0096 0. 0425 [I-2000 : I-2006] Log -0. 20 -0. 13 -0. 0055 0. 0419 [I-2000 : II-2006] Log -0. 92 0. 26 -0. 0087 0. 0096 [I-2000 : III-2006] Log -0. 97 0. 24 -0. 0078 0. 0101 [I-2000 : IV-2006] Log -0. 93 0. 35 -0. 0086 0. 0094 [I-2000 : I-2007] Log -0. 99 0. 24 -0. 0073 0. 0003 [I-2000 : II-2007] Log -0. 89 0. 14 -0. 0060 0. 0055 [I-2000 : III-2007] Log -0. 89 0. 06 -0. 0054 0. 0085 [I-2000 : IV-2007] Log -0. 90 0. 03 -0. 0054 0. 0089 Airline UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey Moving Holid ay Regr essor Effect Mongolia Example Intitial Assumptions: Model is Airline, Weekdays (1 regressor) and Moving Holidays (1 regressor) are in the model. Log/Level selectioin is automatic. Airline parameters, and calendar effect parameters are estimated, every quarter. Orange : Statistically significant at 10% Yellow : Statistically significant at 5% Green : Statistically significant at 1% 11/26/2020 5

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment in times of economic crises • In case of economic crises or turmoil, economic time series shows sudden changes in tendency, (rapidly up or down) • Given that a time series is represented by an ARIMA modelin the process of seasonal adjustment, the changes in ARIMA model can be seen mostly. – Specification of ARIMA model, – Parameters of ARIMA model, – Types, dates and parameters of outliers – Parameters of calendar effects • When new data become available at the end of the data, huge revision of old published data probably can be seen if any of properties above mentioned has been changed. UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 6

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment in times of economic crises • Actually, this issue is a matter of fact that which revision policy is chosen for the series. • Current seasonal adjustment → More revision • Concurrent seasonal adjustment → Less revision • Semi-concurrent revision → Mild revision UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 7

TURKISH STATISTICAL INSTITUTE Turk. Stat Seasonal adjustment in times of economic crises • Do not take into account an last observation as an outlier. – You can not judge this point as an AO or LS UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 8

TURKISH STATISTICAL INSTITUTE Turk. Stat Effect of 2009 Q 1 Economic Crisis on Seasonal Adjusted Series UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 9

Turk. Stat TURKISH STATISTICAL INSTITUTE Focused! UNECE Workshop on Seasonal Adjustment 20 – 23 February 2012, Ankara, Turkey 11/26/2020 10