Transcript An additive decomposition of revision to the UK `production`

An additive decomposition of revision
to the UK ‘production’ estimate of GDP
Introduction
• Significant user interest in understanding the
causes of revisions to UK GDP
• Much comment (/criticism) in UK press about
scale and extent of revisions to UK GDP
• Historically, UK has published ‘revisions triangles’
for some time
– these are the equivalent of the ‘real time databases’
relating to OECD’s MEI
• The idea was to extend this to include more detail
about the causes of each revision
Issues impeding analysis of cause of
revision
• The reasons for revisions are generally thought to be too
numerous to establish quantitatively where each revision
comes from
– e.g. late data for current periods will change history through the
process of seasonal adjustment
• Often many causes will underlie any individual revision,
even at a quite detailed level
– say, methods changes. benchmarking, changes to adjustments,
late data, etc.
• Untangling these effects can be very time consuming,
and is often subjective
• There are often so many small revisions, that it may be
impractical to count all of them
UK response
• A means of systemising as far as possible the
attribution of causes to individual revisions was
sought
• The GDP production team worked with a systems
development team over a period of 6 months to
set up systems to achieve this
• This is still work in progress, and new
‘modernised’ national accounts systems are
being built which incorporate and extend the
basic approach now used
GDP system
• The UK GDP team now produce a regular monthly report of
the causes of revisions
– needed monthly, because, although GDP is a quarterly series, it
is updated monthly
• The current system operates at the 2-digit SIC level
• All revision to growth in 2-digit indices are examined if the
impact of the revisions on GDP growth is greater than 0.02
percentage points
• For these series, the production system is ‘run’ with and
without each change since the last production run to quantify
the impact of each revision
• For example, if a series has had late data, changes to
‘coherence adjustments, and re-seasonal adjustment, these
are run sequentially, and the difference is then attributed to
each cause.
Example
SIC 74
Previously published growth rate
Latest estimate only taking account of late data
(applying previous seasonal factors)
% growth
3.0
3.6
impact of late data
0.6
Latest estimate only taking account of late data (
applying previous seasonal factors) and including
changes in coherence adjustments)
4.3
impact of changes in 'adjustments'
0.7
Latest estimate
4.0
impact of seasonal adjustment
-0.3
Example of results from the 2006 ‘Blue
Book’ run
Later
Proxy
5%
Industry Review
47%
Seasonal
Adjustment 3%
Seasonal
Adjustment
Review 3%
Annual
Coherence
Adjustment
20%
Absolute Quarterly
Revisions at BB2006
Data Quality
Adjust 9%
Weights 4%
Quarterly
Coherence
Adjustment
10%
GDP(O) annual revisions to divisions by cause
Positive
(58%)
7%
3% 4%
20%
Negative
(42%)
1%
36%
48%
10%
3%
44%
3%
4%
5%
12%
La t e r D e f l a t or
La t e r P r ox y
A nnua l C ohe r e nc e A dj ust m e nt s
A dj ust m e nt D a t a Qua l i t y
A dj ust m e nt Qua r t e r l y C ohe r e nc e
We i ght C ha nge s
I ndust r y R e v i e w
06 Q3 M3: absolute quarterly revisions to
divisions by cause 2005 and 2006 Q1 – Q3
2006 Q1-Q3
2005
10%
9%
19%
2%
10%
32%
3%
34%
30%
46%
5%
Nomenclature used to assign causes
The system identifies 15 different type of revision:
• 1
Forecast data for proxy series replaced by actual data
• 2
Forecast data for deflator series replaced by actual data
• 3
Firmer actual data for proxy series received from supplier
• 4
Firmer actual data for deflator series received from supplier
• 5
Seasonal adjustment (from later data)
• 6
Changes to 2-digit data quality adjustments (automatically assessed)
• 7
Changes to 2-digit quarterly coherence adjustments (automatically assessed)
• 8
Changes to MIDSS adjustments
• 9
Other
• 10
Changes to weights (automatically assessed)
• 11
Seasonal adjustment review
• 12
Methodological changes, i.e. Industry review
• 13
Changes to annual coherence adjustments (automatically assessed)
• 14
Errors - Source error
• 15
Errors - Processing error
• Some of these are ‘manually’ identified, but increasingly the process is becoming
automated.
Next steps
• Current system still quite labour intensive
• New systems being designed to systematise the
processes
• ‘Cut-off’ for deciding if revisions are ‘significant’ will
be reduced to zero
• Level of detail will be reduced from 2-digit to 4-digit
components.
Summary
• Current system identifies reasons c.90% of total revision
– Partially identified by system
– Remainder manually detected during normal quality assurance
procedure
• Causes of revision are recorded using standard coding
• Aim to have analytical output during the production round in
time for inclusion in briefing
– size of revision
– reason for revision
– which industry
• Also analysis over time
– e.g. between first estimate, and estimates at t+12 and t+24 etc