Transcript December 14, 2009

INTERNATIONAL SEMINAR ON EARLY WARNING AND
BUSINESS CYCLE INDICATORS
14-16 December 2009, Scheveningen, The Netherlands
10 years experience
in forecasting turning points
Jacques Anas
Director, Economic Indicators and Statistical Models
[email protected]
Introduction

In 1996, The Chamber of Commerce of Paris decided to investigate the development
of a leading indicator for the French economy. This has been the starting point of this
detecting system I will present : methodology and real time results.

The concept of turning point must be clearly defined : A hidden event (local extrema,
change of regime), quite difficult to predict.

Expectations of users (Government, Central Banks, corporate sector, economists for
their prediction) are not identical.
If the goal is clear (anticipate an economic upturn or downturn), the
communication is an issue (probability, threshold and signal, risk, quantitative
implication). There is a need to simplify the messages and to link the signal to an
economic analysis.

A tool to be used in complement of other tools (modelisation, short-term analysis by
economists). It must be considered as a safety net.

Validation is an issue (ex-post dating to have a reference chronology)
10 years experience in forecasting turning points
December 14, 2009
The concept of detection
1. We consider here the follow-up of cyclical movements through
cyclical indicators with the purpose to be timely or even leading
the movements;
2. It is different than using warning indicators, more structural
(imbalances for ex) and potentially leading to crises.
Past TP’s
 Dating (chronology)
Present TP’s  Real time detection (implicitly predicting the future
also!)
Future TP’s  Predicting turning points (there is an interesting
topic: what is a false alarm? It may happen that the signal is
invalidated by an external shock or by a quick policy-mix answer.
www.coe.ccip.fr
10 years experience in forecasting turning points
December 14, 2009
Turning point detection system

A conceptual framework: the ABCD approach

3 indicators:
• IARC: Indicateur avancé de retournement conjoncturel (to predict A
and D)
• IESR: Indicateur d’entrée-sortie de récession (to detect B and C)
• IRC: Indicateur de croissance sous-jacente
To validate the signals of the IARC and IESR indicators
10 years experience in forecasting turning points
December 14, 2009
The framework: the ABCD approach
What kind of a cycle?
•
Classical business cycle  Level
•
Growth cycle
 Deviation to trend
•
Growth rate cycle
 Variation
Vocabulary issue:
phases of BC : recession and upturn
phases of GC: slowdown and rebound
phases of AC : deceleration and re-acceleration
www.coe.ccip.fr
10 years experience in forecasting turning points
December 14, 2009
ABCD approach
Three possible representations
•
•
•
Classical cycle (in level)
Growth cycle (deviation from trend)
Growth rate cycle
10 years experience in forecasting turning points
December 14, 2009
10 years experience in forecasting turning points
December 14, 2009
Three possible sequences of downturn
movements
Pure acceleration cycle (« trou d’air »)
• αβ
Growth cycle:
• αA(β)D
• αA(βαβ)D with an acceleration cycle during the slowdown
β is the first optimistic signal)
• αAB(β)CD
• αAB(βαβ)CD with an acceleration cycle during the recession
Business cycle: (
• αAB(βCαBβ)CD case of a double-dip recession
10 years experience in forecasting turning points
December 14, 2009
United States
Les fluct uat ions de la product ion
B u sin e sc yc le (b ilo n s$ 2 000,sa )
14000
A
Cycle d es affaires
..
B
A
13000
B
. ..

12000
11000
A
D
C
B
C D
.
10000

niveau
tendance
9000
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
Cycle des
affaires ou cycle
classique
(business cycle),
représenté par
le niveau de la
production.
2010
Grow th cycl e i n %
Cycle d e croissance
3
.
A
A
A
2
1
0
0
-1
.
écart à tendance
-2
D
D
phase de ralentissement
phase de récession
-3
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
Cycle de
croissance
(growth cycle
ou deviation
cycle),
représenté par
l'écart du
niveau de la
production à sa
tendance.
2010
Grow th rate cycl e i n %

6
.


4
Cycle d u t aux d e croissance
2
0

-2
0
-4
Tr end gr owt h r at e
GDP gr owt h ( m a3)
-6
1999
2000
2001
2002
.
B
taux de croissance instantané
2003
2004
2005
2006
2007
2008
2009
2010
.

.
C
Cycle du taux
de croissance
ou cycle
d'accélération
(accelation
cycle),
représenté par
le taux de
croissance
instantané de la
production.
© Coe- Rexecode
2
10 years experience in forecasting turning points
December 14, 2009
Euro area
Les fluct uat ions de la product ion
B u sin e sc yc le b ilo n so fe u ro s2 000,sa )
Cycle d es affaires
2000
A
..
B
A
1900
B
. ..

C D
1800
A
D
.
1700

niveau
tendance
1600
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
Cycle des
affaires ou cycle
classique
(business cycle),
représenté par
le niveau de la
production.
2010
Grow th cycl e i n %
Cycle d e croissance
4
.
A
A
3
A
2
1
0
0
.
écart à tendance
-1
D
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
D
phase de ralentissement
phase de récession
-2
2010
Cycle de
croissance
(growth cycle
ou deviation
cycle),
représenté par
l'écart du
niveau de la
production à sa
tendance.
Grow th rate cycl e i n %
6

4
0




2
.

Cycle d u t aux d e croissance


-2
0
Tr end gr owt h r at e
GDP gr owt h ( m a3)
-4

-6
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
.
B
taux de croissance instantané
2009
2010
.

.
C
Cycle du taux
de croissance
ou cycle
d'accélération
(accelation
cycle),
représenté par
le taux de
croissance
instantané de la
production.
© Coe- Rexecode
4
10 years experience in forecasting turning points
December 14, 2009
A couple of probabilistic indicators
applied to France, the euro area and
United States
Importance of non linear processes in business cycle analysis
A leading indicator of the growth cycle (IARC)
• Based on Neftçi’s a posteriori probability method applied to
leading series (Neftçi, 1982, 1984).
A coincident indicator of the business cycle (IESR)
• Based on the univariate Markov switching model applied to
coincident series (Hamilton, 1989, 1990).
An important step is the selection of indicators on which the
model is applied. An aggregation method has been developed
to combine the probabilities
10 years experience in forecasting turning points
December 14, 2009
1. IARC indicator predicting A and D

Computed since 1997
Recursive bayesian algorythm of Neftçi (1982)

Pt 1  (1  Pt 1 )Tt  f 0 ( xt )
Pt 
Pt 1  (1  Pt 1 )Tt  f 0 ( xt )  (1  Pt 1 )(1  Tt ) f 1 ( xt )

Need of a preliminary dating of the growth cycle turning points to
select the candidate leading indicators

Multiple criteria (statistical, practical and economic) to select the
leading indicators (real, financial and surveys)
10 years experience in forecasting turning points
December 14, 2009
Prediction of a global turning point
based on one component
Each « k » component allows for the computation of the probability of
a coming economic turning point (Rt=1)
P k ( Rt  1)  P( Rt  1 Stk  1) P( Stk  1)  P( Rt  1 Stk  0) P( Stk  0)
 tk  P ( Rt  0 S tk  1)
 tk  P ( Rt  1 Stk  0)
Pk ( Rt  1)   k  (1   k   k ) Pt k
10 years experience in forecasting turning points
December 14, 2009
Components of US IARC
Stock Market Index
(MA3)
Interest Rate Differential
(MA3)
Chief Executives
Polls
10 years experience in forecasting turning points
Consumers
anticipations
Manufacturing Inventories
(inverse)
New Construction
Permits
December 14, 2009
Weighted aggregation of the «a
posteriori probabilities» (a priori
weights)
1
N
N
1
Pk ( Rt  1) 

N
k 1
N
  
k 1
1  
N
k
k
k
k


(
1




)
P

t
k 1

k
Pt k
N
k
10 years experience in forecasting turning points
December 14, 2009
2. IESR indicator (coincident for B and C)

Modelling coincident variables by Markov Switching models
Yt   ( St ) 
p
  j (St )Yt  j   t
j 1

(St)t K-state first order Markov chain with transition probabilities pij
where p11+ p12+ …+ p1K =1

Parameters estimation: maximum likelihood in conjunction with
Expectation-Maximization

Choice of K: comparison of 2 and 3 regimes in terms of cycle
replication measured by:
1 T
QPS   ( Rt  Pt ) 2
T t 1
10 years experience in forecasting turning points
December 14, 2009
Decision rules for communication
Empirical thresholds for aggregate probabilities
• IARC
• 60% and 80% for the aggregate IARC
• SERI
• 50% for filtered probabilities and the aggregate probability
www.coe.ccip.fr
10 years experience in forecasting turning points
December 14, 2009
3. IRC indicator
(temporal disaggregation)
Goal : we need to validate quicly a signal
Answer: by estimating a monthly GDP underlying growth rate ( few
revisions, increased timeliness, reduced volatility)
General form of ADL models (Autoregressive Distributive Lag)-
 yt   yt 1  m  gt  xt0  xt11   t
  1
t  2,..., n

2


NID
(0,

)
t

xt is the matrix of k endogenous series of size (n x k)
0
1 are vectors of size (n x 1)
m and
g are constant
10 years experience in forecasting turning points
December 14, 2009
Selection process of the best model
• Univariate selection of the best series for estimating the last three years
period of definitive GDP (2004-2006)
• Selection of the best three series (in level or in difference)
The choice :
- Estimate and extrapolate the model using all past quarterly GDP
available at the cost of revising the estimate
or
- Estimate and extrapolate using only »final» data at the cost of
missing strong movements which can’t be captured with surveys
(present case)
10 years experience in forecasting turning points
December 14, 2009
The selected model for the United
States
- Housing Index
- manufacturing ISM
- Household confidence index (Conference Board)
X t  (V1 ,V2 ,V3 )t
Yt
: monthly variation of GDP
 0,98(1,70) 
 0,74( 1,22) 




Yt  0,82Yt 1  2285(67)  X t  2,12(2,30)   X t 1  1,39( 1,60)    t
 0,09

 0,13

0,30) 
(0,42) 


10 years experience in forecasting turning points
December 14, 2009
Performance analysis
IARC of France (1997-2006)
10 years experience in forecasting turning points
December 14, 2009
4 recessions in France since 1970
1900
Milliards d'euros aux prix de l'année précédente chainés (taux annuel)
1700
1500
1300
1100
900
700
500
1970
1975
1980
1985
10 years experience in forecasting turning points
1990
1995
2000
2005
2010
December 14, 2009
9 slowdowns (4 are related to a recession)
3
Ecart à la tendance
2
1
0
-1
-2
-3
1970
1975
1980
1985
10 years experience in forecasting turning points
1990
1995
2000
2005
2010
December 14, 2009
Slowdowns happen when growth is below
trend growth rate
8
Croissance du PIB et de la tendance
(variation sur un trimestre en %)
Tendance
PIB (mm3)
6
4
2
1
3
0
2
-2
-4
1970
1975
1980
1985
10 years experience in forecasting turning points
1990
1995
2000
2005
2010
December 14, 2009
Asian Crisis
Signal for the related slowdown
At the end of May we could say : “the economic slowdown may have been of short duration,
probably three quarters, and also of low amplitude
Recherche du prochain pic
100
oct.98
88.2
forte probabilité
de retournement
80
possibilité de
retournement
60
40
20
0
J FMA MJ J A S ON D J F MA MJ J A SO N D
1997
1998
10 years experience in forecasting turning points
December 14, 2009
Slowdown of 2001
As early as September 2000, the 80 threshold was reached in the
October 2000 IARC indicator for the euro area published on November 23,
euro area
2000
10 years experience in forecasting turning points
December 14, 2009
Signal of slowdown exit in July-August 2003
A signal was given however in February 2002 but in October 2002, we could say that the rebound
was aborting
August 2003 IARC indicator for France published on October 3, 2003
10 years experience in forecasting turning points
December 14, 2009
Acceleration cycle in 2005
A slowdown was anticpated in June but the signal was weak and
cacelled out as soon as September 2005”
10 years experience in forecasting turning points
December 14, 2009
Performance of the indicators IARC, IESR
and IRC during the current cycle
“How they work together”
10 years experience in forecasting turning points
December 14, 2009
EURO area
-Signal of slowdown with June 2007 IARC
-Signal of recession with August 2008 IESR (published 14th of October)
-Signal of exit of slowdown with May 2009 IARC
-Signal of exit of recession with September 2009 IESR
Zone euro
Indicat eur d’ent rée et sort ie de récession (IESR)
1.0
récession
0.5
non récession
0
2001
2002
2003
2004
2005
2006
2007
2008
2009
© Coe-Rexecode
10/ 12/ 2009
10 years experience in forecasting turning points
63
December 14, 2009
IARC, IRC and GDP for France
IARC of August 2007 >80  Peak of the growth cycle in
the coming 3 months
5
September 18,2007 signal !
4
3
2
1
0
-1
-2
Start of the recession according
to IRC!
-3
m jj jj aa sss ooo nnn ddd jjj
j f m a m j j a s o n d j ff m
m aa m
m aa m
m jj jj aa ss oo nn dd jj ff m
2006
2007
2007
10 years experience in forecasting turning points
2008
2008
2009
2009
2009
27-déc-06
27-déc-06
27-déc-06
27-déc-06
27-déc-06
10-janv-07
10-janv-07
27-déc-06
10-janv-07
10-janv-07
27-déc-06
21-févr-07
10-janv-07
21-févr-07
27-déc-06
21-févr-07
10-janv-07
21-févr-07
27-déc-06
10-janv-07
21-févr-07
10-janv-07
05-mars-07
10-janv-07
27-déc-06
05-mars-07
21-févr-07
05-mars-07
05-mars-07
21-févr-07
10-janv-07
27-déc-06
05-mars-07
21-févr-07
04-avr-07
05-mars-07
04-avr-07
21-févr-07
04-avr-07
10-janv-07
05-mars-07
04-avr-07
21-févr-07
05-mars-07
21-févr-07
04-mai-07
10-janv-07
04-avr-07
04-mai-07
10-janv-07
05-mars-07
04-avr-07
21-févr-07
04-mai-07
04-mai-07
27-déc-06
05-mars-07
05-juin-07
04-avr-07
10-janv-07
04-mai-07
05-juin-07
21-févr-07
05-mars-07
04-avr-07
05-juin-07
04-mai-07
05-juin-07
05-mars-07
24-août-07
04-avr-07
05-juin-07
10-janv-07
24-août-07
04-mai-07
21-févr-07
05-juin-07
04-mai-07
04-avr-07
24-août-07
05-mars-07
01-oct-07
05-juin-07
24-août-07
04-mai-07
21-févr-07
01-oct-07
24-août-07
04-avr-07
21-févr-07
01-oct-07
04-mai-07
05-juin-07
24-août-07
09-nov-07
05-mars-07
10-janv-07
05-juin-07
01-oct-07
04-avr-07
24-août-07
09-nov-07
04-mai-07
01-oct-07
05-juin-07
09-nov-07
17-déc-07
21-févr-07
01-oct-07
04-avr-07
24-août-07
09-nov-07
04-mai-07
05-mars-07
17-déc-07
05-juin-07
24-août-07
01-oct-07
09-nov-07
17-déc-07
24-janv-08
04-mai-07
05-juin-07
24-août-07
09-nov-07
24-janv-08
17-déc-07
05-mars-07
10-janv-07
01-oct-07
04-avr-07
27-déc-06
17-déc-07
21-févr-07
05-mars-07
24-janv-08
25-févr-08
09-nov-07
01-oct-07
24-août-07
05-juin-07
25-févr-08
04-mai-07
24-janv-08
17-déc-07
01-oct-07
07-mars-08
09-nov-07
25-févr-08
24-janv-08
24-août-07
17-déc-07
04-avr-07
05-juin-07
07-mars-08
09-nov-07
25-févr-08
24-janv-08
05-mars-07
07-avr-08
01-oct-07
07-mars-08
04-mai-07
25-févr-08
24-août-07
17-déc-07
09-nov-07
07-avr-08
24-janv-08
07-mars-08
21-févr-07
07-mai-08
04-avr-07
05-juin-07
01-oct-07
17-déc-07
25-févr-08
07-avr-08
04-avr-07
07-mars-08
07-mai-08
09-nov-07
24-août-07
24-janv-08
09-juin-08
07-avr-08
25-févr-08
17-déc-07
07-mai-08
04-mai-07
01-oct-07
07-mars-08
24-janv-08
09-juin-08
05-mars-07
07-avr-08
08-juil-08
09-nov-07
05-juin-07
07-mai-08
25-févr-08
17-déc-07
07-mars-08
24-août-07
09-juin-08
08-juil-08
07-avr-08
10-janv-07
24-janv-08
04-avr-07
01-oct-07
03-sept-08
07-mai-08
25-févr-08
09-juin-08
04-mai-07
09-nov-07
08-juil-08
04-mai-07
07-mars-08
17-déc-07
07-avr-08
03-sept-08
25-sept-08
07-mai-08
24-janv-08
09-juin-08
05-juin-07
25-févr-08
08-juil-08
21-févr-07
03-sept-08
07-mars-08
24-août-07
25-sept-08
01-oct-07
29-oct-08
07-mai-08
09-nov-07
07-avr-08
09-juin-08
08-juil-08
17-déc-07
03-sept-08
24-janv-08
25-sept-08
05-mars-07
29-oct-08
25-févr-08
25-nov-09
07-mars-08
07-avr-08
04-avr-07
09-juin-08
08-juil-08
07-mai-08
03-sept-08
05-juin-07
25-sept-08
04-mai-07
29-oct-08
25-nov-09
23-déc-08
05-juin-07
09-nov-07
24-août-07
17-déc-07
01-oct-07
24-janv-08
25-févr-08
07-mars-08
07-avr-08
07-mai-08
08-juil-08
03-sept-08
25-sept-08
29-oct-08
25-nov-09
23-déc-08
janv-09
09-juin-08
December 14, 2009
IARC, IRC and GDP for France
(real time analysis , even for GDP)
IARC of May 2009 <-80  Trough of the growth cycle in
the coming 3 months
4
3
June 5, 2009 signal !
2
1
0
-1
End of recession according
to IRC
-2
-3
-4
-5
-6
-7
j f m a m j j a s o n d j f m
ma m
m j jj a ss o nn dd jj ff m
m aa m jj j aa ss oo nn dd
2008
2009
2009
10 years experience in forecasting turning points
2010
2010
07-mars-08
07-mars-08
07-mars-08
07-mars-08
07-mars-08
07-mars-08
07-mars-08
07-avr-08
07-mars-08
07-avr-08
07-avr-08
07-avr-08
07-mars-08
07-avr-08
07-avr-08
07-avr-08
07-mars-08
07-avr-08
07-mai-08
07-avr-08
07-mai-08
07-mai-08
07-avr-08
07-mai-08
07-mars-08
07-avr-08
07-mai-08
09-juin-08
07-mai-08
09-juin-08
07-avr-08
07-mai-08
09-juin-08
09-juin-08
07-mai-08
07-avr-08
09-juin-08
08-juil-08
07-mai-08
09-juin-08
08-juil-08
07-mars-08
07-avr-08
07-mai-08
09-juin-08
08-juil-08
08-juil-08
09-juin-08
03-sept-08
07-mai-08
08-juil-08
03-sept-08
09-juin-08
07-avr-08
08-juil-08
03-sept-08
07-mai-08
09-juin-08
08-juil-08
03-sept-08
25-sept-08
25-sept-08
08-juil-08
03-sept-08
09-juin-08
07-mai-08
25-sept-08
03-sept-08
07-avr-08
08-juil-08
29-oct-08
25-sept-08
09-juin-08
03-sept-08
29-oct-08
25-sept-08
08-juil-08
07-mai-08
29-oct-08
03-sept-08
25-nov-08
25-sept-08
09-juin-08
29-oct-08
25-nov-08
08-juil-08
03-sept-08
25-sept-08
29-oct-08
25-nov-08
23-déc-08
07-mars-08
07-avr-08
07-mai-08
09-juin-08
08-juil-08
03-sept-08
25-sept-08
29-oct-08
25-nov-08
23-déc-08
28-janv-09
23-déc-08
25-nov-08
29-oct-08
25-sept-08
03-sept-08
28-janv-09
23-déc-08
08-juil-08
25-nov-08
29-oct-08
25-févr-09
09-juin-08
28-janv-09
25-sept-08
23-déc-08
25-févr-09
25-nov-08
03-sept-08
28-janv-09
29-oct-08
25-mars-09
07-mai-08
23-déc-08
25-févr-09
08-juil-08
25-sept-08
25-nov-08
28-janv-09
25-mars-09
25-févr-09
27-avr-09
23-déc-08
29-oct-08
03-sept-08
25-mars-09
09-juin-08
28-janv-09
25-nov-08
27-avr-09
25-févr-09
25-sept-08
27-mai-09
23-déc-08
25-mars-09
08-juil-08
29-oct-08
28-janv-09
27-avr-09
07-avr-08
27-mai-09
25-févr-09
25-nov-08
23-juin-09
25-mars-09
03-sept-08
23-déc-08
27-avr-09
27-mai-09
28-janv-09
25-sept-08
23-juin-09
25-févr-09
24-juil.-09
07-mai-08
29-oct-08
25-mars-09
27-avr-09
25-nov-08
27-mai-09
23-juin-09
23-déc-08
24-juil.-09
09-juin-08
28-janv-09
25-sept-09
25-févr-09
25-mars-09
08-juil-08
27-avr-09
27-mai-09
23-juin-09
03-sept-08
24-juil.-09
25-sept-09
25-sept-08
25-oct.-09
29-oct-08
25-nov-08
23-déc-08
28-janv-09
25-févr-09
25-mars-09
27-avr-09
27-mai-09
23-juin-09
24-juil.-09
25-sept-09
25-oct.-09
24-nov.-09
December 14, 2009
IARC, IRC, IESR and GDP in the United States
(real time analysis , except for GDP)
IARC of May 2009 <-80  Trough of the growth cycle in
the coming 3 months
4
June 3, 2009 signal !
2
IESR = 0.65 in October
still no sign of exit of the
recession
0
-2
August 28, 2009 statement :
End of recession according to a
markovian model based on initial
claims for unemployment (date : June)
-4
-6
End of recession according to IRC
-8
j f mam j j a s o n d j f ma m j j a so n d j f mam j j a s o n d
2008
2009
10 years experience in forecasting turning points
2010
déc-08
déc-08
déc-08
déc-08
déc-08
déc-08
déc-08
janv-09
janv-09
déc-08
janv-09
janv-09
janv-09
févr-09
janv-09
févr-09
déc-08
déc-08
févr-09
janv-09
janv-09
févr-09
mars-09
mars-09
févr-09
janv-09
mars-09
avr-09
févr-09
mars-09
avr-09
févr-09
janv-09
mars-09
avr-09
mai-09
déc-08
févr-09
mai-09
avr-09
mars-09
juin-09
mai-09
févr-09
avr-09
juin-09
juil-09
mai-09
mars-09
juin-09
avr-09
juil-09
janv-09
août-09
mai-09
juin-09
juil-09
févr-09
août-09
sept-09
mars-09
avr-09
mai-09
juin-09
juil-09
août-09
sept-09
nov-09
December 14, 2009
Improvements for the future

The results of the indicators were quite useful to complement the short-term economic analysis. When the
indicator is not going in the same direction of an analysis, it stimulates the discussion and imposes some
cautiousness in assertions. At times, it may correct the prediction. But there is a real difficulty to communicate.

The communication with the sequential Netçi’s approach is not easy because it is a recursive process with a
discontinuity (semi-parametric approach). A times series approach would be preferable. We may also consider a
multivariate approach. Models with more than two regimes may help to consider anticipating A,B, C and D at the
same time.

It is rather difficult to make a difference between anticipating a deceleration and anticipating a slowdown. We
have not been successful in making this difference but at least we could quickly detect the re-acceleration (which
may be an abortion of a rebound for exemple). A particular example is the signal of economic upturn in United
States given during the summer 2001 but which was invalidated by September 11.

A route to improve the pertinence of the signal is to increase the threshold from 80 to 90 for example, but we may
lose on lead. There is a trade-off between timeliness and minimizing first-type errors.
10 years experience in forecasting turning points
December 14, 2009