Forecasting for QNA estimates Geoffrey Brent & Alex Stuckey, - - PowerPoint PPT Presentation

Forecasting for QNA estimates

Geoffrey Brent & Alex Stuckey, Australian Bureau of Statistics

geoffrey.brent@abs.gov.au

Views expressed are those of the authors and do not necessarily represent those of the ABS. Where quoted or used, they should be attributed clearly to the authors.

Background

Want to produce a timely, accurate series of Quarterly National Accounts, based on 2 inputs:

• Annual “benchmark” series

– Accurate but not quarterly, not timely (~2 yrs lag)

• Quarterly “indicator” series

– Timely but less accurate

Back series

For older years where benchmark data is available, use indicator series to pro-rate BM between quarters

• Aim to preserve quarter-to-quarter movements

from indicator series as much as possible

• Several different methods available: e.g.

Denton-Cholette, Cholette-Dagum, Chow-Lin, Enhanced Denton (Di Fonzo-Marini), Litterman, Fernandez

Forward series

For recent years w/ benchmark not yet available, use indicator to estimate QNAs

• Estimates are based on observed

relationships between benchmark and indicator

– e.g. B/I ratio from last benchmarked quarter

• Typically estimated within benchmarking

method, but can fit separately

Revisions

• Estimates revised each

year as new benchmark data arrives

• B/I relationship may

change gradually or abruptly

• If forward-series

estimation doesn’t get BI relationship right, bias and/or large revisions could result

Forecast methods

• Previous work by Y. Poorun, R. Mathews, J. Chien, P.

Gould suggested forecasting BI ratios separately might improve quality of forward estimates

• We used R to test a range of benchmarking and forecast

methods on ABS series (13 Industry, 44 Public Capital)

• FC methods included random walk, random walk with

drift, exponential smoothing, ARIMA

• Evaluated results on bias (initial relative to final

estimates) and magnitude/timeliness of revisions

Example forecast methods

Results: Industry

• Average across 13

series shown

• Denton/Cholette/

Dagum methods do well on revisions

• RWD reduces bias

noticeably, small increase in revs

• No outliers

Results: Public Capital

• Average outcomes

for 44 series

• Similar to Industry:

D/C/D best on revisions, RWD minimises bias

• Chow-Lin revisions

affected by outliers (~+0.2 to average)

Summary

• Explicitly forecasting BI ratios can reduce

bias (especially RWD method)

• Complex forecasting methods e.g. ETS,

Holt, ARIMA didn’t perform as well as RWD (possibly due to short data-fitting period?)

• RWD slightly increases revisions – but

may be worthwhile for bias reduction?