CORRELATIONS
Robert Engle
Stern School of Business
Correlation
Correlations for Life
What is the correlation between thunder and
rain?
What is the correlation between exercise and
health?
What is the correlation between happiness
and good food?
1
Correlations for Risk
Stock returns are correlated
Stocks in one country are correlated with stocks
in another
Bond returns on one firm or country or maturity
are generally correlated with returns on others
But stock and bond returns sometimes appear
uncorrelated
The risk of a portfolio is greater if all the assets
are highly correlated. It may go down (or up)
further, if they all move together.
QUOTATIONS
It is not the biggest, the brightest or the
best that will survive, but those who adapt
the quickest. Charles Darwin
The secret of life is to be interested in
one thing profoundly and a thousand
things well. Henry Walpole
Studies of high school graduates rarely
find any correlation between recognition in
high school and recognition thereafter.
2
ANTICIPATING CORRELATIONS
Can we anticipate future correlations?
3
4 4
3 3
2 2
1 1
Y_00
Y_50
0 0
-1 -1
-2 -2
-3 -3
-4 -4
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4
X X
4 4
3 3
2 2
1 1
Y__50
Y_90
0 0
-1 -1
-2 -2
-3 -3
-4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4
X X
.15
.10
LARGE CAP STOCKS
.05
AXP
.00
-.05
-.10
-.15
.15
10 YEARS OF
.10
.05
JPM
.00
-.05
-.10
-.15
.20
.15
.10
.05
INTC
.00
-.05
-.10
-.15
.2
AXP .1
MSFT
.0
JPM
-.1
INTC -.2
.12
MSFT .08
.04
MRK
.00
MRK -.04
-.08
-.12
-.15 -.10 -.05 .00 .05 .10 .15 -.15 -.10 -.05 .00 .05 .10 .15 -.15 -.10 -.05 .00 .05 .10 .15 .20 -.2 -.1 .0 .1 .2 -.15 -.10 -.05 .00 .05 .10
AXP JPM INTC MSFT MRK
4
DAILY CORRELATIONS
T3 T5 T20
MONTH YRRET YRRET CAN$ POUND$ AUS$ YEN$ SP500
5
WEEKLY EQUITY CORRELATIONS
1987-2002
US ITALY FRANCE JAPAN HONG
KONG
US 0.223 0.465 0.223 0.308
ITALY 0.537 0.237 0.269
FRANCE 0.340 0.347
JAPAN 0.229
WHY DO WE NEED
CORRELATIONS?
6
WHY DO WE NEED
CORRELATIONS?
CALCULATE PORTFOLIO RISK
DIVERSIFICATION
Diversified portfolios have lower variance
and risk because some assets go one
direction while others go the opposite.
There are many thousands of possible
stocks, bonds and other assets to invest
in. Can we reduce the risk to zero?
Clearly not. Assets are not uncorrelated.
7
PORTFOLIO RISK
Portfolio risk depends upon the
volatilities and correlations of all the
components.
For weights w and covariance matrix
Omega
P2 = w ' w
1
w=
' 1 0
8
ARE CORRELATIONS TIME VARYING?
YES
WHY?
Because the business practice of the
companies changes
Because shocks to the economy affect all
businesses
Because shocks to one part of the economy
will affect only some businesses
CONDITIONAL CORRELATIONS
DEFINE BOTH COVARIANCES AND
VARIANCES CONDITIONAL ON
CURRENT INFORMATION
Et 1 ( r1,t r2,t )
1,2,t =
Et 1 ( r1,2t ) Et 1 ( r22,t )
9
ESTIMATION
HISTORICAL CORRELATIONS
Use a rolling window of N observations for both
covariances and variances. We will use 100 days.
DYNAMIC CONDITIONAL CORRELATION or
DCC
Estimates conditional correlations by first adjusting
for differing variances and then updating correlations
as new information is received.
.8
.7
.6
.5
.4
.3
.2
.1
.0
94 95 96 97 98 99 00 01 02 03 04
C 100_AX P_G E
10
GENERAL ELECTRIC PROFITS
11
10.0
5.0
4.0
3.0
2.5
2.0
1.5
1.0
0.5
1990 1992 1994 1996 1998 2000 2002 2004
FNBEFORE HNBEFORE
5.0
3.0
2.0
1.5
1.0
0.5
1990 1992 1994 1996 1998 2000 2002 2004
FNB E FO R E FND O W N
HNB E FO R E HNUP
FNUP HND O W N
12
50.0
30.0
20.0
15.0
10.0
5.0
3.0
2.0
1.5
1.0
0.5
1990 1992 1994 1996 1998 2000 2002 2004
F NB E F O R E F ND O W N F NA F T E R
HNB E F O R E HNUP HNA F T E R
F NUP HND O W N
IMPLICATIONS
On Jan 1 2000 the market prices of Ford and
Honda reflected the best analysis of the
financial markets
What would happen to energy prices?
What would happen to the economy?
What choices would management make?
Five years later, Ford stock was down and
Honda was up.
The market rewarded the company that was
prepared for higher energy prices.
13
HISTORICAL CORRELATIONS
.6
.5
.4
.3
.2
.1
.0
-.1
-.2
-.3
1990 1992 1994 1996 1998 2000 2002 2004 2006
C 1 00 _ F O R D R _ H O N D AR
DYNAMIC CONDITIONAL
CORRELATION
14
MULTIVARIATE
MODELS
15
DYNAMIC CONDITIONAL
CORRELATION OR DCC
1. Estimate volatilities for each asset and
compute the standardized residuals or
volatility adjusted returns.
2. Estimate the time varying covariances
between these using a maximum likelihood
criterion and one of several models for the
correlations.
3. Form the correlation matrix and covariance
matrix. They are guaranteed to be positive
definite.
HOW IT WORKS
When two assets move in the same direction,
the correlation is increased slightly.
This effect may be stronger in down markets
(asymmetry in correlations).
When they move in the opposite direction it is
decreased.
The correlations often are assumed to only
temporarily deviate from a long run mean
UPDATING IS THE CENTRAL FEATURE
16
CORRELATIONS UPDATE LIKE
GARCH
Approximately,
t = + 1,t 1 2,t 1 + t 1
=
1
.8
.7
.6
.5
.4
.3
.2
.1
.0
95 96 97 98 99 00 01 02 03 04
C 9_AXP _G E
17
.9
.8
.7
.6
.5
.4
.3
.2
.1
.0
95 96 97 98 99 00 01 02 03 04
FACTOR MODELS
One or more factors influence all assets
Some assets are more affected by a
particular factor than others
Sometimes the factors have little volatility
and therefore have little influence
18
ONE FACTOR ARCH
One factor model such as CAPM
There is one market factor with fixed betas and
constant variance idiosyncratic errors
independent of the factor. The market has
some type of ARCH with variance m ,t.
2
ri ,t = i rm ,t + e i ,t
i ,i ,t = i2 m2 ,t + i
MARKET VOLATILITY
.030
.025
.020
.015
.010
.005
.000
94 95 96 97 98 99 00 01 02 03 04
C o n d it io n a l S t a n d a r d D e via t io n
19
CALCULATE DYNAMIC
CORRELATIONS
1 2 m2 ,t
t =
(
1
2 2
m ,t + 12 )( 22 m2 ,t + 22 )
.8
.7
.6
.5
.4
.3
.2
.1
94 95 96 97 98 99 00 01 02 03 04
C 4_AXP _G E
20
CORRELATION OF EXTREMES
How correlated are extreme returns?
Bankruptcy is an extreme event and
corresponds to an extremely large
negative stock return over a period of
time.
Are bankruptcies correlated?
21
ASYMMETRY IN MARKET RETURNS
22
HOW WELL DOES THIS WORK?
Examine 18 large cap stocks in the US.
Calculate correlations either historically or with
Dynamic Conditional Correlation (DCC)
Relate these correlations to the volatility of
S&P500.
Does High market volatility mean high
correlation?
RESULTS
23
PLOT
About 30 Correlations of these large cap
stocks on left axis
Estimated with DCC not using market
data
Compare with a GARCH of the S&P500
plotted on right axis
.024
.020
S&P volatility
.016
1.2 .012
0.8 .008
.004
0.4
0.0
Correlations
-0.4
94 95 96 97 98 99 00 01 02 03 04
24
MEAN CORRELATION AND MARKET
VOLATILITY
.032 .50
.028 .45
.024 .40
.020 .35
.016 .30
.012 .25
.008 .20
.004 .15
.000 .10
94 95 96 97 98 99 00 01 02 03 04
M E AN C O R 9F V9F _S P R E T
REGRESSION
Dependent Variable: MEANCOR9F
Method: Least Squares
Date: 09/10/06 Time: 20:00
Sample: 1/04/1994 12/31/2004
Included observations: 2770
25
REGRESSION IN DIFFERENCES
Dependent Variable: D(MEANCOR9F)
Method: Least Squares
Date: 09/09/06 Time: 11:37
Sample (adjusted): 1/06/1994 12/31/2004
FINDINGS
MARKET VOLATILITY IS PART OF THE
STORY
THE CURRENT DECLINE IN MARKET
VOLATILITY HAS NOT LEAD TO THE
EXPECTED DROP IN CORRELATIONS.
26
ANTICIPATING CORRELATIONS
FORECASTING FACTOR VOLATILITIES
IS PART OF THE ANSWER
HOW CAN WE MAKE THIS WORK
BETTER?
Research Agenda!
Build DCC models on the residuals
Build Factor DCC models
27
HOW DO WE FORECAST FACTOR
VOLATILITIES?
USE GARCH MODELS OR SIMILAR
MODELS FOR SHORT RUN
FORECASTS.
USE NEW MULTI-COUNTRY RESULTS
USING THE SPLINE GARCH FOR LONG
RUN MACRO BASED FORECASTS.
28
S&P500
1 .2
1 .0
0 .8
0 .6
0 .4
0 .2
0 .0
60 65 70 75 80 85 90 95 00
CVOL UVOL
MULTIPLE REGRESSIONS
All Countries
emerging 0.0376
( 0.0131 )** Time Effects
transition -0.0178
( 0.0171 )
log(mc) -0.0092 0.25
( 0.0055 )*
log(gdpus) 0.0273
( 0.0068 )** 0.2
nlc -1.8E-05
( 5.4E-06 )**
grgdp -0.1603 0.15
( 0.1930 )
gcpi 0.3976
( 0.1865 )** 0.1
vol_irate 0.0020
( 0.0008 )**
vol_gforex 0.0222 0.05
( 0.0844 )
vol_grgdp 0.8635
( 0.1399 )**
0
vol_gcpi 0.9981 1990 1994 1998 2002
( 0.3356 )**
29
ANTICIPATING CORRELATIONS
To forecast correlations, we must forecast the
volatility of the factors that influence the
companies.
When volatility is forecast to be high, then
correlations will be high.
Inflation, slow growth, macroeconomic instability
forecast high market volatility.
This does not work well when companies are
changing their business. May need to update
residual correlations using factor DCC.
30