Professional Documents
Culture Documents
EUROSTAT
Luxembourg, 2006
www.kof.ethz.ch
www.idp.zhwin.ch
Real-Time Signalextraction : Efficient
Concurrent Level- and Turning-Point-Filters
Marc Wildi
wia@zhwin.ch
www.kof.ethz.ch
www.idp.zhwin.ch
Introduction
Framework
Statistical problem
www.kof.ethz.ch
www.idp.zhwin.ch
Framework: revision of the KOF economic
barometer
Important leading indicator for Swiss GDP
New `KOF-Konjunkturbarometer´ released in April
2006
Extended data set (KOF):
Business survey data, consumer confidence,
indicators in finance and construction sectors
New Filter (idp)
New filter technique (DFA)
Concurrent filter output published only
No revisions
Standardized Growthrates
3 3
2 2
GDP
1 1
0 0
-1 -1
-2 -2
KOF-Barometer
-3 -3
84 86 88 90 92 94 96 98 00 02 04
Statistical Problems
Trend
Eliminate (damp) `all´ high frequency components
Concurrent (real time) filter
Emphasizes the last observation
Concurrent filter is causal (completely asymmetric)
Statistical Problems
1. Fast and reliable detection of Turning Points
2. Minimization filter error variance (level-filter)
Both requirements are incongruent
-1.5
2001 2002 2003
www.kof.ethz.ch
www.idp.zhwin.ch
Introduction
Methodological approaches
Real-time (concurrent) filters
Model Based Approach, MBA (X-12, X-11 ARIMA
or TRAMO/SEATS)
Identify a time series model for the DGP
Forecast the future
Apply the symmetric filter to the extended time
series
Direct Filter Approach, DFA (Wildi,2004)
Minimize an efficient estimate of the filter error
variance
Criteria:
MBA: min ( X t 1 Xˆ t 1 ) 2
DFA:
Concurrent DFA- filter is computed `directly´
(explicitly) wypiwyg
Direct control of amplitude and time delay
Key to TP-detection
Target (symmetric filter) is known
www.kof.ethz.ch
www.idp.zhwin.ch
Introduction
www.kof.ethz.ch
www.idp.zhwin.ch
Series 3 (Business Survey, linearized)
Salient features
Seasonal spectral peaks
(Potential) location known
height and width must be accounted for
6*2=12 degrees of freedom
Low frequency (trend) peak
Location, height and width
3 parameters
Conclusions
www.kof.ethz.ch
www.idp.zhwin.ch
Boundary filter: the series
The Data
www.kof.ethz.ch
www.idp.zhwin.ch
Boundary filter: performance measure
TRAMO
In sample Out of sample
Mean 79% 78%
Median 79% 79%
TRAMO
Mean Median
in sample 41% 41%
out of sample (1 year) 59% 58%
out of sample (2 years) 58% 55%
out of sample (3 years) 63% 60%
out of sample (7 years) 70% 68%
Examples based on
particular time series
www.kof.ethz.ch
www.idp.zhwin.ch
Series 4 Business Survey: Ideal Trend
(1 B) X t (1 0.2063) t
www.kof.ethz.ch
www.idp.zhwin.ch
Problem
0.8
0.6
0.4
0.2
0.0
N /2
min ˆ | (k )
( ) |2 I ( )
k NX k
k 1
2 [T / 2]
T k 0 W ( ) 2
| ( ) ˆ ( ) |2 I ( )
A
k k k NX k
min ˆ
2 [T / 2] 2W ( ) 2 A( ) Aˆ ( )(1 cos( ˆ ( )))I ( )
T k 0 k k k k NX k
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
1.5
-1.5
2001 2002 2003
DFA -1 0 1 2
Error type 1 3 1 2
Type 2 (between TP's)
Type 2 (at TP's)
Total false diagnostics 6
Without Anticipations 3
Baro unfiltered
1 Level Filter
TP-Filter
0
1994
1994
1995
1995
1996
1997
1997
1998
1998
1999
1999
2000
2001
2001
2002
2002
2003
2004
2004
2005
-1
-2
4.00
3.00
2.00
1.00
Old Baro
New Baro
0.00
1991 M 1
1991 M 12
1992 M 11
1993 M 10
1994 M 9
1995 M 8
1996 M 7
1997 M 6
1998 M 5
1999 M 4
2000 M 3
2001 M 2
2002 M 1
2002 M 12
2003 M 11
2004 M 10
2005 M 9
-1.00
-2.00
-3.00
Filter characteristics
Analysis of filter errors
www.kof.ethz.ch
www.idp.zhwin.ch
Series 33, ideal trend:
Filter Characteristics
Diagnostic:
overshooting of amplitude function
Corresponding spectral peak in filter errors
Solution: restrict transferfunction so that amplitude
cannot overshoot (A(ω)<1.1 for all ω)
Very easy (1 line of code) because filter is optimized
directly
www.kof.ethz.ch
www.idp.zhwin.ch
Results available in Excel
Email to: wia@zhwin.ch
lag CHIMT_DEX12_F1A_S00
CHIMT_DEX12_F1A_S00
CHIMT_DEX12_F1B_S00
CHIMT_DEX12_F1B_S00
CHIMT_DEX12_F2A_S00
CHIMT_DEX12_F2A_S00
CHIMT_DEX12_F2B_S00
AR MA AR MA AR MA AR
0 1 0.18251845 1 0.31541543 1 0.2914562 1
1 -0.85140615 0.04659423 -1.46708376 -0.19845817 -1.38151327 -0.21745491 -3.36915574
2 -0.62738702 -0.0979975 1.32884811 0.25008302 -1.27816326 -0.49948554 5.26453485
3 1.097489533 0.04350718 -1.50275891 -0.34005661 3.30640999 0.65706077 -4.54329994
4 -0.54478372 0.01116092 0.7923393 0.05620913 -0.50853241 0.24175743 1.95563106
5 0.071559525 0.01116092 -0.24387793 -0.1137235 -3.20087087 -0.79455828 -0.01255998
6 0.318119421 0.01116092 0.34025138 0.08870741 2.21039355 0.16903169 -0.30171467
7 -0.25004938 0.01116092 -0.12113046 0.02593908 1.43406135 0.54303915 0.05273983
8 -0.04935325 0.01116092 -0.05889456 -0.02144223 -2.16750486 -0.31343054 0.0053496
9 0.250000492 0.01116092 0.00202535 -0.00241562 0.23241833 -0.13225003 -0.0017537
10 -0.05142859 0.01116092 0.0008351 0.0006807 0.59435117 0.06191141 0.00011557
11 -0.30472177 0.01116092 -6.84E-05 -3.30E-05 -0.29594568 -0.01214883 -8.71E-05
12 -0.20334143 -0.17135753 -2.66E-06 -1.57E-06 0.18811676 0.05742151 -6.18E-06
13 0.341975756 -0.03543331 -1.26E-08 -7.10E-09 -0.07787042 -0.0219154 -2.90E-07
14 0.25735004 0.10915842 -6.91E-10 -1.07E-10 -0.01230402 0.00620898 -5.69E-09
15 -0.27954077 -0.03234626 -9.64E-11 -5.83E-11 -0.00559811 -0.00690134 -4.54E-11
scale 1 1 1 1
intercept 0 0 0 0
Mean series (the
-1.50464644
means are computed in sample
29.1801432
from ) -8.96015021 -31.13819