eaafa938a81f922b805d7b394809987e.ppt

- Количество слайдов: 21

Automation in National Accounts: applied to SNA 2008 revision of time series for Supply and Use Tables Vincent Ohm National Accounts Statistics Netherlands (CBS) contribution for WPNA meeting OECD, Paris - 28 October 2011 Centraal Bureau voor de Statistiek

Outline • Objective • Business case for automation • Semi-automatic integration in National Accounts - recent developments at Statistics Netherlands (CBS) - basic outline of methodology (without technicalities) • Application : ‘Time series machine’ • Practical aspects Centraal Bureau voor de Statistiek

Objective Implementation of SNA 2008 changes in the NA ‘core systems’ (SUT & Sector Accounts) and satellite accounts for the complete time series Coverage: – Supply and Use Tables (and IOT) : 1969 -2009 – Sector Accounts : 1980 -2009 – Satellite Accounts : 1980 -2009 (mostly) Centraal Bureau voor de Statistiek

Business case for automation Need for efficiency - limited capacity for revision of time series • Staff reduction at CBS 2010 -2012 : NA department: approx. -20% in number of staff. Additional reductions up to 23% in budget expected for CBS for 2013 -2016 • Compiling time series for SNA 93 proved to be very time consuming and costly (8 working years!) New methodology and software - available since 2010 • developed within the framework of the ‘redesign chain of economic statistics’ programme at CBS Centraal Bureau voor de Statistiek

Recent developments at CBS • ‘Redesign chain of economic statistics’ programme at CBS: initiated in 2007 and now nearly completed (beginning 2012) • Aims to increase tranparency, quality and efficiency in the complete ‘chain’ of economic statistics : from ‘source’ (data collection) to statistical ‘end products’ (among which is National Accounts) • Within the National Accounts department : focus on redesign of the production of NA ‘core systems’ to implement the new CBS process architecture Centraal Bureau voor de Statistiek

CBS process architecture (selection) - applied to NA production Expert knowledge required Incident based Laborious and tedious ‘rest point’ manually automated Rule based source data for NA system Centraal Bureau voor de Statistiek without large discrepancies integrated and fully consistent system

Automated integration in National Accounts (1) • Since 2005 benchmarking of Quarterly Sector Accounts (QSA) to annual totals was achieved using a Matlab implementation of a multivariate Denton method which included lineair constraints and weights for the individual variables in the model • This implementation however proved to be inadequate for Quarterly Accounts (QNA) with respect to the maximum problem size and the relations between the variables: only lineair constraints available • Since 2008 an extended version of the multivariate Denton method was developed by colleagues of the R&D department and subsequently the new methodology was implemented by the IT department into software modules: ‘Quarterly machine’ Centraal Bureau voor de Statistiek

Automated integration in National Accounts (2) • Since 2009 additional methodology was developed for balancing NA systems for a single quarter or year based on Stone’s method. The methodology and implementation strongly resemble that of the Quarterly machine: ‘Balancing machine’ which will not be discussed here • Since 2010 the new ‘machines’ (methodology & software modules) have been used in QNA & SUT production process, and will also be implemented for QSA & Sector Accounts beginning 2012 (currently in testing phase) • In addtion new applications are being developed, for example for the Supply and Use Table time series discussed here Centraal Bureau voor de Statistiek

Outline of Quarterly machine methodology (1) • Input data: N ‘periods’ consisting of M ‘subperiods’, typically we would have N=3 years and M=4 quarters (or N=9 and M=1 for Time series) • Quarterly data for each variable (e. g. each ‘cell’ in the Supply and Use Table) is viewed in the time direction as a series for which Denton’s method (1971) is applied (i. e. quadratic minimization of first differences in the series: either proportionally or additively) • The benchmarking (or balancing) problem is formulated as a quadratic (convex) optimization problem which is solved numerically using the XPress MP package (no technical/mathematical details provided here!) • Each individual series can be assigned a weight for the ‘Denton terms’ in the optimization problem: these weight reflects the reliability of the source data of the series with respect to other series in the system Centraal Bureau voor de Statistiek

Outline of Quarterly machine methodology (2) We can define constraints on series and between different series: • ‘hard’ linear constraints: A + B (+…) = c (c = constant) exact • ‘soft’ linear constraints: A + B (+…) c approximately (weight) • ‘hard’ inequalities: A + B (+…) > c exact (also for < c) • ‘soft’ ratios: A / B r (r = constant) approximately (weight) • set upper and lower bounds for individual variables (subperiods) • set ‘exogeneous’ (fix) individual variables (subperiods) • impose periodical data on subperiodical data (exact or approximately) i. e. q 1 + q 2 + q 3 + q 4 = y or: q 1 + q 2 + q 3 + q 4 y (weight) Note: In the above A and B can be linear combinations of other variables. Centraal Bureau voor de Statistiek

Quarterly machine : implementation adds ‘rules’ to the data: model setup databases data . xls input module weight labels ‘object library’ accessed from VBA(*) VBA = Visual Basic for Applications Centraal Bureau voor de Statistiek . xls benchmark module . xls solution (rounded) benchmark problem (data + rules) ‘interface’ with XPress MP

Input Module : Example of a rule (VBA) Centraal Bureau voor de Statistiek

Application: Time series for SUT (1) • Complete SUT time series cover years 1969 - 2009 • We have a setup with N=1 and M=1 (subperiod = period; no quarters!) • General idea is to ‘divide’ the complete series up into a few subseries: – 2001 – 2009 – 1995 – 2001 – 1987 – 1995 – 1979 – 1987 (less detail) – 1969 – 1979 (less detail) • The bounding years: 2009, 2001, 1995, 1987 , 1979, 1969 are called interface years (in Dutch: steekjaren) which mark the original years where previous SNA revisions have been first implemented (in particular 1987, 1995, 2001 and now 2009). Best data available for these years! Centraal Bureau voor de Statistiek

Application: Time series for SUT (2) • SNA 2008 conceptual changes as well as incidental corrections of errors will be implemented manually by our NA experts for all interface years • Note: There is an extra step where we first balance the interface years using the balancing machine (not discussed here!) and some manual corrections where necessary: the result are fully balanced interface years once these are approved by our experts • The balanced interface years then serve as fixed boundaries between which we will interpolate the intermediate years using the ‘Time Series Machine’ (which is a special application of the Quarterly machine) Centraal Bureau voor de Statistiek

Application: Time series for SUT (3) • Interpolation by the Time Series Machine will be achieved for current and previous year prices simultaneously such that price and volume indices of the original series will be preserved as much as possible whereas new levels are given by the interface years • All necessary rules (relations between series as discussed before) as well as weights will be set to achieve an optimal result, which however will have to be approved of by our NA experts before it is accepted • Revision of the time series is therefore an iterative process, which however requires only limited capacity per iteration once the model setup has been programmed: however incidentally small adjustments in the rules or manual intervention (corrections) by experts may be necessary! Centraal Bureau voor de Statistiek

Application: Time series for SUT (4) • The result of one iteration of the Time series machine is a complete (sub)series of Supply and Use Tables which are internally fully consistent and which are matched to the new levels set by the interface years for which SNA 2008 concepts have been manually implemented • Once the results of the Time series machine have been carefully examined and approved by our experst for one (sub)series, the results will be loaded into the database and we will proceed with the next subseries, working back in time (so first 2001 -2009, then 1995 -2001, etc. ) • After the previous SNA revision (2001) the time series had to be manually adjusted and balanced year by year which was very laborious and tedious and which has cost 8 working years in capacity (klopt dit? ? ? ) Centraal Bureau voor de Statistiek

Application: Time series for SUT (5) • Although revision for the year 2009 and then for the interface years (2001, 1995, etc. ) will start by the end of this year, actual work on the time series (using the Time series machine) will start around June 2012 • Will this new semi-automated approach work? • Although we have not tested the Time series machine for SNA revision of time series before, we are confident the approach will be successful: • We have been using the Quartely machine for benchmarking quarterly data to annual totals for SUT for the last two years (2010 & 2011) and we have gained sufficent experience using the new software and methodology. And compiling the time series is a very similar problem. Centraal Bureau voor de Statistiek

Application: Time series for SUT (6) • Also we are currently using a (somewhat simplified) version of the Time series machine for implementing the new CPA classification into our SUT time series, where the approach is essentially the same: interpolate between manually corrected and balanced (fixed) interface years • This time series approach works quite well and the problem size is feasible. However for the SNA 2008 revisions we expect larger changes in the levels of the interface years, which may imply that more manual interventions and more iterations for each subseries will be needed to converge to an optimal end result • Lastly we have capable & helpful colleagues both on the technical/methodological aspects as well as on the National Accounts subject matter and data so we sleep soundly Centraal Bureau voor de Statistiek

Some practical aspects Centraal Bureau voor de Statistiek

Questions? Centraal Bureau voor de Statistiek

Thank you for listening For additional information about this subject or the automation methodology (‘machines’) we have developed at CBS, you are welcome to contact me at: vs. [email protected] nl Centraal Bureau voor de Statistiek