The ABS has previously communicated on the potential use of State Space Models, in particular Seemingly Unrelated Time Series (SUTSE) models, for measuring impacts introduced into statistical outputs by its transformation program. Changes to the design of the questionnaire forms and data collection procedure are two potential causes of statistical impact during transformation.
While the use of SUTSE models so far has been primarily focussed on impact measurement, these models could be used more widely as part of the quality assurance processes that happen during regular statistical production.
As such, Methodology has started developing time series models to produce forecasts of key economic statistics, and are working with the Quarterly Economy Wide Surveys (QEWS) area and the National Accounts compilation team to trial incorporating model forecasts as part of quality assurance. The idea is that analysis of the model outputs will be incorporated into the regular production process, and they will be used before, during and after process changes happen through transformation.
An initial case study has explored forecasting Private New Capital Expenditure on Equipment, Plant and Machinery. The Capital Expenditure (Capex) survey collects not only actual expenditure in the reference quarter, but also expectations of future expenditure. Analysis has shown that the series on expectations have considerable correlation with the actual expenditure series, through their long-term trends, with a leading lag. An historical analysis assessed the performance of the forecasts produced by alternative models in recent years. SUTSE models, which exploit this correlation between expected and actual expenditure, had lower average prediction errors compared to a time series model based on only the time series of actual expenditure.
The models that have been built provide forecasts and 95% prediction intervals, which can be used as an objective way to assess the extent to which an observed estimate is consistent with both the historical behaviour of the time series itself, and the related series. The outputs provide an additional source of information to complement existing processes used for validation. Use of these outputs for quality assurance purposes was trialled in real-time for the Capex survey for the first time for June quarter 2019 estimates.
Next steps of this work are to review how the forecasts were used as part of the regular production process, identify any improvements to the model and presentation of the model outputs, and to extend to other key economic time series.
The author would welcome any feedback, particularly on methods for identifying related variables which can improve forecasts for key economic indicators, and the use of forecasts as part of quality assurance.
For more information or to provide feedback, please contact Jennie Davies Methodology@abs.gov.au