Title: HIGH DIMENSION DYNAMIC CORRELATIONS
1HIGH DIMENSION DYNAMIC CORRELATIONS
- By
- Robert Engle
- Stern School of Business
2FINANCIAL FORECASTING
- Forecasting Risk
- In high dimension systems this means forecasting
volatilities and correlations - Predictive Failure?
- Black Swans?
3WHY DO WE NEED CORRELATIONS?
- CALCULATE PORTFOLIO RISK
- FORM OPTIMAL PORTFOLIOS
- PRICE, HEDGE, AND TRADE DERIVATIVES
4ARE CORRELATIONS TIME VARYING?
- Yes Correlations are time varying
- Derivative prices of correlation sensitive
products imply changes. - Derivatives on correlation now are traded.
- Time series estimates change. There are many
varieties.
5(No Transcript)
6GENERAL ELECTRIC PROFITS
7CHANGING EXTERNAL EVENTS
- CONSIDER BOEING AND GENERAL MOTORS
- CORRELATIONS MAY HAVE CHANGED BECAUSE OF CHANGING
ENERGY PRICES. ENERGY PRICE VOLATILITY WILL MAKE
THE RETURNS MORE CORRELATED.
8(No Transcript)
9Dynamic Conditional Correlation
- A simple approach to estimating and forecasting
large covariance matrices - Engle(2002), (200?)
- Based on standardized returns s
10TWO SPECIFICATIONS
- Where in the integrated model
- And in the mean reverting model
11CORRELATION TARGETING FOR THE INTERCEPT
12ESTIMATION
- Two steps
- GARCH on returns
- MLE on standardized returns
13OOps
- This is not a good solution for very large
systems - Inversion of R many many times is slow
- Downward bias in alpha for large systems- see
Engle and Sheppard - Possibly need more structure
14THE MACGYVER METHOD
15(No Transcript)
16A SOLUTION
- Estimate all bivariate models. Each should be
consistent for the same parameters. - Combine these estimates
- Calculate all the correlations
17MONTE CARLO
- Generate 1000 observations on n dimensional
systems which are gaussian DCC with unit
conditional variance. - Construct with the following parameters
- N3,5,10,20,30,50
18MLE ESTIMATION
- Bivariate Log likelihood function
- Although the parameters should be positive and
sum to less than one, these restrictions are not
imposed.
19RESTRICTED MLE ESTIMATION
- Log likelihood function
- To constrain the sum of alpha and beta to be less
than one a penalty is assigned whenever they
exceed one.
20ESTIMATORS
- Mean, Median and trimmed Mean where the 5
largest and smallest values are deleted before
taking the mean. - When n3, median and trimmed mean are the same.
- The estimators are computed for both restricted
and unconstrained Bivariate MLE.
21EXPERIMENTS
22RMS ERRORS ON BETA
23RMS ERRORS ON ALPHA
24BIAS ON BETA
25BIAS ON ALPHA
26BIAS OF MEDIAN ESTIMATOR
27ADVANTAGES OF THIS APPROACH
- Requires only bivariate estimation
- Does not require the same number of observations
on all assets - Robust to occasional non-convergence and
diverging estimates - Potential to investigate validity of DCC
assumptions relative to block DCC, GDCC - Possible to estimate only a subset of pairs and
apply parameters to all.
28DYNAMIC EQUICORRELATION
- Robert Engle and Bryan Kelly
29Dynamic EquiCOrrelation (DECO)
- Suppose all pairs of assets have the same
correlation, but that this changes over time. It
is like the average correlation. - Can you estimate this directly?
30PLAUSIBILITY?
- Elton and Gruber use this model for Asset
Allocation. See their textbook. - Derivatives are priced with a single average
correlation - Credit Risk often assumes homogeneity
- In many cases this can be interpreted as an
average correlation - A block equicorrelation model allows more
flexibility.
31ADVANTAGES
- The likelihood is very simple
- It can be estimated about as fast as GARCH no
matter how many assets there are. - It is easy to interpret and analyze.
32Estimation
33Estimation
- Assuming Normality and examining the second step
as in DCC
34UPDATING
- Each period, new information becomes available on
the equicorrelation. We specify this in several
ways. - Where U can be various measures. i.e.
35FACTOR DCC
- And its special cases
- FACTOR ARCH
- FACTOR DOUBLE ARCH
36FACTOR ARCH
- Introduced by Engle Ng and Rothschild(1990)(1992)
- Or, assuming normality
37ESTIMATION OF FACTOR ARCH
- MLE is estimation by OLS to get the betas
- GARCH estimation of the factor
- Correlations are calculated from
38FACTOR DOUBLE ARCH
- Idiosyncrasies are also GARCH models
-
- or
39ESTIMATION OF FACTOR DOUBLE ARCH
- Regress returns on the factor assuming the errors
are GARCH. This is MLE by weak exogeneity
argument. - Model the factor volatility with GARCH
- Calculate correlations
40FACTOR DCC
- Model the residuals of the FACTOR DOUBLE ARCH and
the residuals of the factor with DCC - Two step method is not MLE
- If residual correlations are all zero then there
is no change from FACTOR DOUBLE ARCH
41COVARIANCE MATRIX
- Now four terms in covariance matrix
- and correlations are given by
42FEATURES
- This model has time varying correlations between
the idiosyncrasies. - This is what would be observed if there are
multiple factors with time varying volatilities - This model has time varying covariances with the
factor that is time varying betas.
43PERFORMANCE
44DATA
- 18 LARGE CAP STOCKS AND SP500
- aa, axp, ba, cat, dd, dis, ge, gm, ibm, ip, jnj,
jpm, ko, mcd, mmm, mo, mrk, msft - 1994-2004 DAILY FOR 2771 Obs
45MacGyver
- Use TARCH or GJR-GARCH for all 18
- Estimate 153 bivariate models without
restrictions - Median alpha0.0157
- Median beta. 9755
46Bivariate Alphas
47Bivariate Betas
48RESULTS
- 153 TIME SERIES
- LOOK AT AVERAGE CORRELATION BY DCC AND BY 100 DAY
HISTORICAL OR MOVING AVERAGE
49(No Transcript)
50(No Transcript)
51(No Transcript)
52(No Transcript)
53(No Transcript)
54FACTOR DCC
- Estimate DCC on residuals from FACTOR DOUBLE ARCH
- Use MacGyver method again
- Median alpha.009
- Median beta .925
- Less persistent and less responsive to shocks
than for DCC.
55(No Transcript)
56(No Transcript)
57(No Transcript)
58(No Transcript)
59(No Transcript)
60UPDATING
- USING PARAMETER VALUES ESTIMATED FROM 1994-2004,
CALCULATE CORRELATIONS THROUGH SEPTEMBER 14,2007 - WHAT DO THE CORRELATIONS SHOW THROUGH THE SUMMER
OF 2007 WITH THE MARKET TURBULENCE?
61(No Transcript)
62(No Transcript)
63WHICH CORRELATION ESTIMATE IS MOST ACCURATE?
- CONSIDER AN ASSET ALLOCATION SOLUTION.
- FOR EACH PAIR OF STOCKS FORM THE MINIMUM VARIANCE
PORTFOLIO AND SEE WHICH COVARIANCE ESTIMATOR
ACHIEVES THE SMALLEST VARIANCE.
64WHICH HEDGE GIVES THE BEST LONG - SHORT PORTFOLIO?
- For each pair of assets hold one long and take a
short position in the other to minimize variance.
65Average standard deviation of the minimum
variance portfolio of all pairs of stocks
66Average standard deviation of the minimum
variance long - short portfolio of all pairs of
stocks
67WINNING FRACTIONS MV
68WINNING FRACTIONS HEDGE
69(No Transcript)