Presentation is loading. Please wait.

Presentation is loading. Please wait.

TESTING AMI PERFORMANCE OF J-DEMETRA+ Turaç YAVUZ TURKSTAT

Similar presentations


Presentation on theme: "TESTING AMI PERFORMANCE OF J-DEMETRA+ Turaç YAVUZ TURKSTAT"— Presentation transcript:

1 TESTING AMI PERFORMANCE OF J-DEMETRA+ Turaç YAVUZ TURKSTAT

2 Contents Why to test ? Log / Level Test
Errors in detection of seasonality Average number of outliers Correct model identification Average number of stationary parameters Conclusion

3 Why to test ? Importance of AMI
An average user of a seasonal adjustment software might not be a modeling expert. It can get chaotic when one would have to deal with large number of series.

4 Testing Environment; Data used; CPU: Intel Core i5 3.20 Ghz RAM: 4 GB
Platform: Windows 7 64 bit Software: Jdemetra & 1.2.0 Data used; Simulated series by Agustin Maravall using Matlab

5 Data Structure 120 obs Log Level 240 obs

6 Data Structure For each category ≈500 series generated by 50 different type of ARIMA model; 16 Airline-type (0,1,q) (0,1,Q) 16 Non-seasonal 18 Other seasonal

7 Errors in Log/Level Test (in % of series) TRAMO-SEATS
Series in Logs Series in Levels 120 240 Airline-type 2.6 0.0 Non-seasonal 5.3 0.5 Other seasonal 3.3 0.2 0.1 TOTAL 3.7 TOTAL (TSW+) 1.0 0.4 0.2 0.0 Difference between version and is insignificant

8 Errors in Log/Level Test (in % of series) X13
Series in Logs Series in Levels 120 240 Airline-type 4.1 0.0 Non-seasonal 10.0 1.0 Other seasonal 3.7 0.2 TOTAL 5.8 0.4 0.1 TOTAL (TSW+) 1.0 0.4 0.2 0.0 Difference between version and is insignificant

9 Errors in Detection of Seasonality Test (%) TRAMO-SEATS
Number of Obs. 120 240 Airline-type 0.0 Non-seasonal 3.3 2.6 Other seasonal 0.1 TOTAL 1.1 0.8 TOTAL (TSW+) 0.7 0.8 Difference between version and is insignificant

10 Detection of Seasonality (frequencies for “Seasonality” variable in output) X13
120 obs. 240 obs. . 1 Airline-type 7 15593 3 15997 Non-seasonal 153 14846 80 14920 Other seasonal 16449 16500 For version, all values of seasonality variable in output file are “1”.

11 Detection of Seasonality Visual Spectral Test (%) TRAMO-SEATS
Bad Good Severe Airline-type 120 4.3 95.1 0.6 240 4.8 94.4 0.8 Non-seasonal 21.8 68.1 10.1 20.9 68.3 10.8 Other seasonal 5.4 93.6 1.0 5.6 93.3 1.1 TOTAL 10.2 85.9 3.9 Difference between version and is insignificant

12 Average Number of Outliers per Series TRAMO-SEATS
Series in Logs Series in Levels 120 240 Airline-type 0.18 0.11 Non-seasonal 0.15 0.10 0.09 Other seasonal 0.17 0.16 TOTAL 0.12 TOTAL (TSW+) 0.17 0.10 Difference between version and is insignificant

13 Average Number of Outliers per Series X13
Series in Logs Series in Levels 120 240 Airline-type 0.07 0.06 0.08 Non-seasonal Other seasonal 0.05 TOTAL TOTAL (TSW+) 0.17 0.10 Difference between version and is insignificant

14 Correct Model Identification (%) TRAMO-SEATS
Complete Model Orders Differencing 120 240 Airline-type 71.3 75.2 96.9 98.1 Non-seasonal 58.5 68.0 90.0 92.0 Other seasonal 45.0 70.6 90.6 96.1 TOTAL 58.1 71.4 92.5 95.5 TOTAL (TSW+) 64.5 78.9 94.0 97.5 Difference between version and is insignificant

15 Correct Model Identification (%) X13
Complete Model Orders Differencing 120 240 Airline-type 68.1 73.8 94.4 96.6 Non-seasonal 33.6 29.1 72.2 74.8 Other seasonal 34.9 47.2 86.7 87.6 TOTAL 45.7 50.4 84.7 86.6 TOTAL (TSW+) 64.5 78.9 94.0 97.5 Difference between version and is insignificant

16 Correct Model Identification By Model Order (Top 5) TRAMO-SEATS
Complete Model Orders Differencing (3,0,0) (0,0,0) 93.2 99.5 (2,0,0) (0,0,0) 99.3 (1,1,0) (0,0,0) 92.4 99.2 (1,0,0) (0,0,0) 91.7 96.6 (0,0,2) (0,0,0) 90.3 97.0 Difference between version and is insignificant

17 Correct Model Identification By Model Order (Bottom 5) TRAMO-SEATS
Complete Model Orders Differencing (0,0,0) (0,0,0) 0.0 90.5 (0,1,0) (0,0,0) 98.2 (2,0,1) (0,0,0) 2.9 47.1 (1,0,1) (1,0,0) 14.0 74.8 (1,0,1) (0,0,0) 23.2 72.4 Difference between version and is insignificant

18 Average Number of Stationary Parameters TRAMO-SEATS
120 240 In Simulation Model Airline-type 1.86 1.85 1.69 Non-seasonal 1.59 1.63 1.50 Other seasonal 2.36 2.58 2.64 TOTAL 1.95 2.03 1.96 TOTAL (TSW+) 1.93 1.97 1.94 Difference between version and is insignificant

19 Average Number of Stationary Parameters X13
120 240 In Simulation Model Airline-type 1.86 1.84 1.69 Non-seasonal 2.03 2.44 1.50 Other seasonal 2.42 2.75 2.64 TOTAL 2.11 2.35 1.96 TOTAL (TSW+) 1.93 1.97 1.94 Difference between version and is insignificant

20 Conclusions Run-time facts;
Jdemetra+ is remarkably faster than other SA softwares (TSW, TSW+) Average process time of 8000 series with 120 observation is ≈ 29 seconds for TRAMO-SEATS method (45 seconds for X13) 10 of series could’nt be processed with TRAMO-SEATS method (and 1 of for X13)

21 Conclusions I/O facts;
Results of some diagnostic tests are not included in the csv output file (Friedman, Kruskall-Wallis, Evolutive Seasonality etc.). Even if the combined seasonality test result is “Identifiable seasonality not present”, the value of “Seasonality” variable in csv output is “1”.

22 Conclusions ARIMA Model Identification;
Reliability level of X13 is not adequate, For TRAMO-SEATS method, +Jdemetra is more reliable than TSW but not TSW+.


Download ppt "TESTING AMI PERFORMANCE OF J-DEMETRA+ Turaç YAVUZ TURKSTAT"

Similar presentations


Ads by Google