9.

Correlation and Covariation

Alexander Chapter 4, section 4.5, Chapter 6, 7 Volatility is important but the analysis of the link between returns, namely covariation and correlation, is also important => covariance appears all the time. Correlation between series r1t and r2t Cov(r1t , r2t ) 12 p = 12 22 1 2

= corr(r1t , r2t ) = with

1 1 => Does not need to be annualized Minimization of portfolio risk and the covariance matrix Return of a portfolio of k returns rp = w1 r1 + w2 r2 + . . . + wk rk

with Variance of the portfolio

k X wi = 1 i=1

2 2 2 V (rp ) = w1 V (r1 ) + w2 V (r2 ) + . . . + wk V (rk )

+2w1 w2 Cov(r1 , r2 ) + 2w1 w3 Cov(r1 , r3 ) + ... +2wk1 wk Cov(rk1 , rk ) Matrix notation rp = w0 r 59

and V (rp ) = w0 Vr w with w = = w1 w2 . . . wk , r= r1 r2 . . . rk ,

Vr

V (r1 ) Cov(r1 , r2 ) . . . Cov(r1 , rk ) Cov(r1 , r2 ) V (r2 ) . . ... . . . . ... V (rk ) Cov(r1 , rk )

Risk management: min(w0 Vr w)

and the minimum is reached by taking the rst derivative of this quadratic form with respect to w and by solving the system w0 Vr w 0 = (Vr + Vr )w = 2Vr w = 0 w Value at risk of a portfolio Z (p0 Vr p)1/2 where p is the amount of money invested in dierent assets. => ESTIMATION OF COVARIANCES AND CORRELATIONS, POSSIBLY TIME-VARYING

60

 Whole sample correlation = s HV HV t =

(r1t r1 ) (r2t r2 )
i=1 T P

i=1

(r1t r1 )2

(r2t r2 )2

12 12

(r1ti r1 ) (r2ti r2 )
i=1 n P

(r1ti r1 )2

(r2ti r2 )2

12t 1t 2t

EWMA EW MA = t with
2 2 = (1 )r1t1 + 2 1t 1t1 2 2 = (1 )r2t1 + 2 2t 2t1

12t 1t 2t

12t = (1 )r1t1 r2t1 + 12t1 Same , easy to compute (Excel, GAUSS...) multivariate GARCH (M-GARCH): hawful to estimate. Example with two stocks r1t = 1 + 1t r2t = 2 + 2t 1t = u1t 2t p h11t p = u2t h22t

61

we cannot only consider h11t = 1 + 1 2 + 1 h11t1 1t1 h22t = 2 + 2 2 + 2 h22t1 2t1 but Ht = h11t h12t h21t h22t !

with the elements (symmetric matrix so h12t = h21t ) and where h12t = E(1t 2t |It1 ) but h11t = 1 + 11 2 + 12 2 + 13 1t1 2t1 + 11 h11t1 + 12 h22t1 + 13 h12t1 1t1 2t1 h12t = 2 + 21 2 + 22 2 + 23 1t1 2t1 + 21 h11t1 + 22 h22t1 + 23 h12t1 1t1 2t1 h22t = 3 + 31 2 + 32 2 + 33 1t1 2t1 + 31 h11t1 + 32 h22t1 + 33 h12t1 1t1 2t1 Multivariate models imply a lot of parameters are dicult to make converged (see Kroener webpage for routines with M-GARCH) => we need to simplify the problem (a lot of dierent specication: most simplest ones here) Diagonal model h11t = 1 + 11 2 + 11 h11t1 1t1 h12t = 2 + 23 1t1 2t1 + 23 h12t1 h22t = 3 + 32 2 + 32 h22t1 2t1 needs multivariate estimation by maximum likelihood. Very restrictive. Yesterday s news in stock 1 does not inuence stock 2. additional constraints to ensure positive deniteness of Ht

62

 The constant GARCH correlation (Bollerslev, 1990) Ht = Dt CDt where h1t 0 ... 0 p 0 h2t 0 Dt = diag( hit ) = . ... . . hkt 0 ... 1 12 . . . 1k 21 1 2k . ... . . k1 . . . 1

and

C=

Preferably in one shot by maximum likelihood but a two step procedure is possible: Estimate a univariate GARCH for each returns Obtain the standardized residuals, namely for r1t = 1 + 1t p 1t = u1t h11t C= with = u1t u2t . . . ukt u0 u T

ut

,t = 1...T

u = ( 1 , u2 , . . . uT )0 u 63

 Engle (1999) Ht = Dt C Dt where h1t 0 ... 0 p 0 h2t 0 Dt = diag( hit ) = . ... . . hkt 0 ... 1 12t . . . 1kt 1 2kt ... 1

and

C =C

EW MA

but with the same otherwise the matrix may be not positive deniteness Alexander orthogonal GARCH (need PCA)

21t = . . . k1t . . .

9.4

Principal component analysis (PCA)

The role of PCA is to reduce the dimension of the system so that only the most important sources of information are used. PCA assumes: Stationary variables: returns and not prices normalized variables x1t = r1t 1 1

otherwise the rst principal component is determined by the more volatile variable

64

 => the variance-covariance matrix = is also the correlation matrix where x1t x2t . for t = 1 . . . T Xt = . . xkt X0 X T

and

X = (X1, X2 , . . . XT )0

PCA analyze the eigenvalue and the eigenvector of the symmetric matrix . You can do it in Eviews 4 or in GAUSS such as = W W 0 where 1 0 0 2 = diag(i ) = 0 with 1 > 2 > . . . > k and the associated W = (w1 : w2 : . . . wk ) Each principal component is a linear combination of the columns of X and the weight are given by the eigenvectors The rst PC explains the greatest amount of the total variation of the X, P1 = Xw1 or more generally P = XW 0 0 k

65

 The second PC the greatest amount for the component orthogonal to the rst PC, .... If possible only k0 < k orthogonal combination are necessary, in the sense that they pick up enough variation within the data. This is measured by
k0 P

V ark = Notice that because

i=1

W 0 = W 1 We have a second identity P X which gives of one of the series Xi = w1i P 1 + w2i P 2 + . . . + wki P k and
ri = i + w1i P 1 + w2i P 2 + . . . + wki P k

= XW = P W 1 = P W 0

where
w1i = w1i . i

One hopes
ri = i + w1i P 1 + w2i P 2 + . . . + wmi P m + et

66

with m<k It is possible to test for the number of signicant eigenvalues but I wouldnt recommend it here because of volatility (heteroskadasticity in the data) Stock indexes AEX, CAC, FTSE and DAX 03/01/1996 - 06/10/2000 (T=1207)
800 700 600 500 400 300 200 250 500 750 AEX 1000 7000 6000 5000 4000 3000 2000 1000 250 500 750 CAC 1000

9000 8000 7000 6000 5000 4000 3000 2000 250 500 750 DAX 1000

7000 6500 6000 5500 5000 4500 4000 3500 250 500 750 FTSE 1000

Indices in level

67

.06 .04 .02 .00 -.02 -.04 -.06 -.08 250 500 750 DAEX 1000

.08 .06 .04 .02 .00 -.02 -.04 -.06 250 500 750 DCAC 1000

.06 .04 .02 .00 -.02 -.04 -.06 -.08 250 500 750 DDAX 1000

.05 .04 .03 .02 .01 .00 -.01 -.02 -.03 -.04 250 500 750 DFTSE 1000

Returns - ln(P t) - ln(P t-1)

lnAEX Mean 0.000900 Median 0.001162 Stdv 0.013441 A 21.33% Skewness -0.349501 Kurtosis 5.412 BJ 317.09

lnCAC 0.000975 0.001083 0.013258 21.04% -0.253973 4.797 175.27

lnDAX 0.000887 0.001337 0.014209 22.55% -0.378950 4.982 226.41

lnFTSE 0.000450 0.000508 0.010637 16.88% -0.126865 4.208 76.64

Descriptive Statistics lnAEX lnAEX 1 lnCAC 0.065 lnDAX -0.027 lnFTSE 0.178 lnCAC 0.065 1 0.597 0.158 lnDAX -0.027 0.597 1 0.098 lnFTSE 0.178 0.158 0.098 1

Correlation Matrix 68

w1 Eigenvalue 1.656 Variance prop 0.414 Cumulative prop 0.414

w2 1.142 0.285 0.699

w3 0.805 0.201 0.901

w4 0.394 0.098 1

Eigenvalues and Variance Proportion I would t take the rst principal component alone to build an European market index w1 lnAEX 0.120 lnCAC 0.681 lnDAX 0.658 lnFTSE 0.295 w2 0.742 -0.128 -0.275 0.599 w3 -0.651 -0.125 -0.08 0.743 w4 0.095 -0.709 0.696 0.043

Eigenvectors

Gauss program new; cls; load x[1207,4] = c:\gauss\indices.prn; lnx = ln(x); dx = lnx - lagn(lnx,1); dx = dx[2:rows(dx),.]; Cor = corrx(dx); /* computes correlation matrix */ {l,v} = eigrs2(cor); /* eigenvalues and eigenvector of a symetric matrix */ (gauss) l 0.39471570 0.80587760 1.1425244 1.6568823 (gauss) v 0.095587520 -0.65199227 0.74253179 0.12006517 -0.70933593 -0.12500478 -0.12867072 0.68165988 69

0.12 0.08 0.04 0.00 -0.04 -0.08 200 400 600 P1 800 1000 1200

0.06 0.04 0.02 0.00 -0.02 -0.04 -0.06 -0.08 200 400 600 P2 800 1000 1200

0.06 0.04 0.02

0.04 0.02 0.00

0.00 -0.02 -0.02 -0.04 -0.06 200 400 600 P3 800 1000 1200 -0.04 -0.06 200 400 600 P4 800 1000 1200

Figure 1: The four principal components

70

0.69697447 -0.084665107 -0.27056991 0.65867316 0.043956268 0.74304220 0.59906786 0.29508276 (gauss) Same but smallest eigenvalues and eigenvector rst, namely 1 < ... < n Next lines illustrate the second identity X=PW mdx = meanc(dx); stdx = stdc(dx); dxsd = dx - ones(rows(dx),1).*.mdx; dxsd = dxsd ./stdx; P= Dxsd*v; P1 = P[.,4];P2 = P[.,3];P3 = P[.,2];P4 = P[.,1]; X1 = v[1,4]*p1 + v[1,3]*p2 + v[1,2]*p3 + v[1,1]*p4; x1~dxsd[.,1];

9.5

Orthogonal GARCH and EWMA

Standardize the series take most important PC: P1, P2, Pm m<k and the associated A=W with = 1 2 . . . k

and estimate GARCH on each of these series Then Ht = AGt A

71

with Gt = V (P 1) 0 ... 0 0 V (P 2) 0 . . ... . . . . 0 ... V (P m) = hP C 0 ... 0 1t PC 0 h2t 0 . . ... . . . . PC 0 ... hmt

Orthogonal EWMA also possible such that V (P 1) 0 0 V (P 2) Gt = 0 0 0

2,P C 0 1t 0 2,P C 2t = V (P m) 0

0 0 2,P C mt

72

Exercise 11 (Final evaluation) Take your three series and estimate (for week 8: give it to me or put it in my mailbox, my email box is full): the best multivariate model for the conditional mean (VAR, VECM, choice of p, ...) compute the trivariate variance-covariance matrix of the residuals for the following methods: Constant Correlation and choose your model (GARCH, TARCH,..) V1t = Dt CDt EWMA correlation V2t = Dt C Dt Full sample variance covariance matrix et 0 et V3 = T Compute the 5% 1-day and 10-days Value at Risk for a total investment of 10,000 Euros in each of your assets. Was it the optimal combination to allocate 30,000 equally in each of the three assets? Justify

73