Title: Short-term forecasting of the GDP growth rate using the BTS in industry and in services: An out-of-sample analysis
1Short-term forecasting of the GDP growth rate
using the BTS in industry and in services An
out-of-sample analysis
2The two issues addressed in the paper
- 1) BTS are widely used for the short-term
forecasting of economic activity. - However, to our knowledge, there has been no
recent attempt to establish the significance of
their contribution to the quality of forecasts
(using modern econometric techniques).
- 2) The service survey contains a specific piece
of information on GDP growth with respect to the
industry survey (Bouton and Erkel-Rousse, 2003). - However, it remains to be shown that this
specific piece of information permits one to
significantly improve the quality of short-term
GDP forecasts with respect to models involving
variables from the industry survey exclusively.
3The study consists in
- - estimating several VAR models and univariate
calibration models of the GDP quarterly growth
rate encompassing miscellaneous variables
derived from the industry and service surveys
as well as GDP lags - - estimating competing models with no service
variable or, even, no BTS variable (simple
AR models of GDP) on several time periods
(real-time analysis - RTA) - - simulating each model up to a four-quarter
forecast horizon, and deriving series of
forecasting errors (RTA) - - comparing the predictive accuracy of the
different models using Clark-West or Harvey,
Leybourne and Newbold tests (depending on the
models)
- Data used quarterly survey data and, to a lesser
extent, monthly survey data.
4Structure of the Talk
- Data
- Methodology
- Results
- Conclusion
5Data (1) General characteristic features
- - 1962 creation of the INSEE industry survey
- - 1988 creation of the INSEE service survey (on
a quarterly basis) - - June 2000 creation of the monthly service
survey, adding of a few questions - - Since the 1990s progressive extension of the
sector coverage of the service survey (in the
data used business services (2/3), household
services (gt 20), real estate activities
(gt10)) - - January 2004 several changes in the wording of
the questionnaires of the two surveys and
adding of a few questions in the service survey
for European harmonisation purpose - - January 2004 the two surveys become compulsory
6Data (2) The data used in the study
- Macro variable to be forecasted
- ? the quarterly GDP growth rate from the
French Quarterly Accounts (hereafter
GDP) - Industry survey data
- ? the 6 main monthly balances and 2
quarterly ones (past and expected demand) - ? 2 static common factors in industry a
monthly one (derived from the 6 main
monthly balances published each month
by INSEE) and a quarterly one
(encompassing the 6 monthly balances the 2
quarterly ones) - Service survey data
- ? the 3 main monthly balances and 3 main
quarterly ones ? the dynamic common factor in
services (introduced by Cornec and
Deperraz, 2007, derived from the 6 balances
and published each month by INSEE)
7Methodology (1)
- We define several forecasting models of GDP for
each month in a given quarter, so as to be able
to up-date the short-term forecasts of GDP each
month on the basis of the most recent piece of
information given by the BTS. - Notation
- Ind_m1 variable Ind derived from the industry
survey relating to month 1 - Ser_m4 variable Ser derived from the industry
survey relating to month 4 - with m1 (m2,m3, resp.) 1st (2nd, 3rd, resp.)
month in the current
quarter and m4 1st month in the following
quarter. - Examples
- - In the second quarter of year 2000, m1
April, m2 May, m3 June, m4 July, of year
2000. - - In the last quarter of year 2004, m1
October, m2 November, m3
December 2004, m4 January 2005. -
8Methodology (2) VAR models
- Comparison of the predictive accuracy of 3 VAR
models of 2 kinds (non restricted, restricted)
per couple of survey variables used - - VAR with 3 variables GDP, Ind_mi,
Ser_mi - - VAR with 2 variables GDP, Ind_mi
- - simple AR model of GDP (basic
benchmark) - with Ind_mi (Ser_mi, resp.) a survey variable
from the industry (services resp.) survey
relating to month mi, i 1 to 4. - Non restricted VAR VAR with 2 lags estimated
using OLS Restricted VAR VAR with 4
lags with exclusion restrictions,
estimated using SURE - Tests of equal predictive accuracy in nested
models Clark-West tests, with - 1) Scilab Grocer (correcting autocorrelation
within forecast error series using Newey-West
variances, for forecast horizons 2 to 4
quarters) - 2) SAS procedure autoreg, options nlag4 and
backstep (testing for
autocorrelation up to 4 lags and correcting it
when necessary using Yule-Walker
estimates)
9Methodology (3) Univariate calibration models (1)
- Intuition when the length of both estimation
series and forecast error series is short
(months 2 and 3, especially 2 no survey in
August), univariate calibration models might be
better adapted because more parsimonious than VAR
models. - Kinds of models estimated one set per month mi
(i 1 to 4). At a given month when BTS are
available up to the 1st forecast horizon - Models used for the forecasting of GDP at a 1
quarter horizon - GDP function of the current and
lagged values of survey
variables - Models used for the forecasting of GDP at a 2
quarter horizon - GDP function of the lagged values
of survey variables - Models used for the forecasting of GDP at a 3
quarter horizon - GDP function of the lagged values
of survey variables to the
exclusion of the first lags -
10Methodology (4) Univariate calibration models (2)
- Comparison of the predictive accuracies of
several univariate calibration models of GDP - - some including explanatory variables
from the two BTS - - some including explanatory variables
from the industry survey only - - AR model of GDP (basic benchmark)
- - the best VAR3 models
- with, again, different sets of models depending
on the month in (or just after) the quarter (mi,
i 1 to 4). - Depending on the models whose predictive
accuracies are compared, we performed - - either Clark-West tests (for the comparison
of nested models) - - or Harvey, Leybourne and Newbold tests (for
the comparison of non-nested models) - Software used Scilab
Grocer.
11Results 1) VAR models (1)
- Choice of survey variables
- - Most leading balances
- Industry survey expected production
(monthly) Service survey expected operating
profit (quarterly) - - Most correlated common factors (with GDP) at
each month mi - Industry survey
- for m1 and m4 the quarterly common
factor - for m2 and m3 the monthly common factor
- Service survey the dynamic common factor
12Results 1) VAR models (2)
- Contribution of BTS to the short-term forecasting
of GDP (with respect to benchmark AR models) - - Very significant at the 1 and 2 quarter
horizons in most cases (very small P-values) - - Significant at the 5 or 10 thresholds for
several models relating to quarterly months
(m1 and m4) at the 3 and, even, 4 quarter
horizons (even though the quality of their
forecasts remains poor) - - Often more significant to forecast the first
release of GDP than the last available release
(at the 3 and 4 quarter horizons and, to a
lesser extent at the 2 quarter horizon)
13Results 1) VAR models (3)
- Contribution of the service survey to the
short-term forecasting of GDP (with respect to
VAR models with 2 variables GDP and Ind_mi) - 1) In the case of quarterly months (m1 and
m4), for which fairly long time series in
services are available - - Significant, especially when the service
variable is the dynamic common factor - - Higher contribution for the 2 and 3
quarter forecast horizons (non significant
at the 4 quarter horizon) -
2) In the case of non quarterly months (m2 and
m3), for which only short time series for
services are available - Less high
contribution of the service survey at this stage
but some encouraging significant results for
month m2 - Important remark The methodology
used creates a serious bias against the
service survey (either linear interpolations are
used for data before June 2000 or the last
available quarterly value of a balance
that in m1- is used for months m2 and m3
while more recent monthly industry data are
used)
14Results 1) VAR models (4)
- Example Results from the non restricted VAR
model with the quarterly common factor in
industry and the dynamic common factor in
services (m4)
15Results 2) Univariate calibration models (1)
- We are currently working on these models.
- Two kinds of models have been estimated, using
the Scilab-Grocer software - Models whose optimal specifications are
automatically determined by the software - Models whose specifications are determined by the
authors so that every explanatory variable has an
impact of the expected sign on GDP. - Our first preliminary results suggest that these
kinds of models might enable one to lead to
clearly positive results as concerns the
contribution of the service survey, notably in
month m3 (especially models whose specifications
are automatically determined). - However (at least on the basis of our
preliminary results), these kinds of models do
not appear to always lead to significantly better
GDP forecasts than the VAR models.
16Results 2) Univariate calibration models (2)
Example Univariate models relating to a
three-quarter forecast horizon
17Conclusion
- 1) The study clearly confirms the predictive
power of BTS for GDP growth at short-term
horizons - 2) It also shows that the quarterly service
survey has predictive power alongside with
the manufacturing survey - 3) As far as monthly service data are concerned,
it is definitely too early to have firm
conclusions. Our results derive from
methodologies that generate negative biases
to the detriment of the service survey (use of
either linear interpolated data before June
2000 or less up-to-date data than industry
ones). Nonetheless, some of the results
suggest that monthly service data might also
permit one to improve the short-term
forecasting of GDP.
To be confirmed in 6 or 7 years when long enough
monthly service series are available!