Ronald van der Stegen Reducing statistical discrepancy between direct and indirect GDP
Introduction: direct and indirect – Direct seasonal adjustment of GDP: seasonal adjustment of GDP – Indirect seasonal adjustment of GDP: sum of seasonally adjusted components of GDP – 2013 first quarter: Indirect GDP q-to-q growth -0.4% Direct GDP q-to-q growth +0.1% 2
Direct and indirect GDP 3
Introduction: GDP 4 Before SA After SA Gross Domestic product (GDP)S≠0S’≈0 +ImportS≠0S’≈0 -ExportS≠0S’≈0 -Consumption of householdsS≠0S’≈0 -Consumption of governmentS≠0S’≈0 -Gross fixed capital formationS≠0S’≈0 -Changes in stocksS≠0S’≈0 =Statistical discrepancy (SD) from index formula (≠0: constant prices) S≠0S’>S
Project: – Minimize: (SD(t)-SD(t-1))/GDP(t-1) Achieved by: -Idea 1: Optimize X12-Arima -Idea 2: Multivariate pretreatment -Idea 3: Rebasing with multivariate Denton Tested on data of
Quality of seasonally adjusted results 1.Standard quality measures of X12-Arima 2.Fluctuations in the statistical discrepancies 3.Revisions of published results 6
Idea 1: improve settings of X12Arima – Numerous settings tried: ‐ Series are very volatile today ‐ Small reduction in fluctuation of SD possible by harmonizing X12Arima setups – Important sources for discrepancy are: ‐ Outliers ‐ Regression effects ‐ Extrapolation 7
Idea 2: multivariate pretreatment – Based on a structural time series model (STM) – Consistency constraints over ‐ Additive outliers ‐ Level shifts ‐ Time dependent regressors ‐ Near future: time dependent seasonal factors – STM removes above effects – Seasonal components of STM too volatile to use for seasonal adjustment therefore seasonal adjustment by X11 8
Idea 3: rebasing – First idea 2 than Multivariate Denton technique – All series are balanced to same order of magnitude – Equal weights for the series 9
Results: Quality measures X12Arima – GDP: acceptable reduction in quality 10 M1M2M3M4M5M6M7M8M9M10M11Q Curre nt Idea Idea
Results: Statistical discrepancy – Significant reduction 11
Results: Revisions 12 – Similar revisions of the GDP
Conclusions – More uniformity in seasonal adjustment results in less statistical discrepancy without significant reduction of the quality of the results – Increased uniformity is established with multivariate pretreatment 13
Contact: Ronald van der Stegen 14