FINANCIAL ECONOMETRICS SPRING 2013 WEEK VII MULTIVARIATE MODELLING OF VOLATILITY Prof. Dr. Burç ÜLENGİN MULTIVARIATE VOLATILITY There may be interactions among the conditional variance of the return series. Also covariance of the return series may change over the time. Therefore the full perspective of volatility modelling requires the treatment of variances and covariances together- simultaneously. When the variances and covariances are modelled it means that correlations are modelled too. MOVING CORRELATION OF THE RETURNS OF TWO FINANCIAL ASSETS 1.0 0.8 0.7 0.6 0.5 0.4 0.3 cor50 cor100 985 944 903 862 821 780 739 698 657 616 575 534 493 452 411 370 329 288 247 206 165 124 83 42 0.2 1 CORRELATION 0.9 MULTIVARIATE GARCH In multivariate GARCH models, yt is a vector of the conditional means (Nx1), the conditional variance of yt is an matrix H (NxN). The diagonal elements of H are the variance terms hii, and the off-diagonal elements are the covariance terms hij. h11h12 .......h1N h h ......h 21 22 2N H .................... hN 1hN 2 ...hNN MULTIVARIATE GARCH There are numerous different representations of the multivariate GARCH model. The main representations are: VECH Diagonal BEKK- Baba, Engle, Kraft, Kroner Constant correlation representation Principle component representation VECH REPRESANTATION Full treatment of the matrix H In the VECH model, the number of parameters can be exteremely large. Estimating a large number of parameters is not in theory a problem as long as there is large enough sample size. The parameters of VECH are estimated by maximum likelihood and the obtaining convergence of the typical optimization algorithm employed in practice be very difficult when a large number of parameters are involved. Also estimated variances must be positive and it requires the additional restrictions on parameters VECH REPRESANTATION 2 Variable Case 2 b b b h11 h11t a110 a11 a12 a13 1t 1 11 12 13 t 1 0 2 2 h12t a12 a 21 a22 a23 1t 1 2t 1 b 21 b22 b23 h12t 1 a 0 a a a 2 b b b h h 31 32 33 22t 1 22t 22 31 32 33 2t 1 h11t a110 a 11 12t 1 a12 1t 1 2t 1 a13 22t 1 b11 h11t 1 b12 h12t 1 b13 h22t 1 h12t a120 a 21 12t 1 a22 1t 1 2t 1 a23 22t 1 b 21 h11t 1 b22 h12t 1 b23 h22t 1 0 h22t a22 a 31 12t 1 a32 1t 1 2t 1 a33 22t 1 b 31 h11t 1 b32 h12t 1 b33 h22t 1 A and B are {Nx(N+1)/2 , Nx(N+1)/2} matrices . In the case of 2 variables, 3 equations and 21 parameters. 5 variables, 20 equations and 820 parameters. 10 variables, 55 equations and 4025 parameters. DIAGONAL REPRESENTATION The diagonal representation is based on the assumption that the individual conditional variances and conditional covariances are functions of only lagged values of themselves and lagged squared residuals. Bollerslev, Engle and Woodridge (1988) proposed In the case of 2 variables, this representation reduces the number of parameters to be estimated from 21 to 9. At the expense of losing information on certain interrelationships, such as the relationship between the individual conditional variances and the conditional covariances. Also estimated variances must be positive and it requires the additional restrictions on parameters DIAGONAL REPRESENTATION 2 Variable Case 2 b 0 0 h11 h11t a110 a11 0 0 1t 1 t 1 11 0 2 2 h12t a12 0 a22 0 1t 1 2t 1 0 b22 0 h12t 1 0 0 b a 0 0 0 a 2 h h 33 2 33 22t 22 22t 1 t 1 h11t a110 a 11 12t 1 b11 h11t 1 h12t a120 a 22 1t 1 2t 1 b22 h12t 1 0 h22t a 22 a33 22t 1 b33 h22t 1 H A ' t 1 t 1 BH(1) DIAGONAL REPRESENTATION OIL & NATURAL GAS PRICES ROIL 40 40 30 30 20 20 10 10 0 0 -10 -10 -20 -20 -30 -30 -40 -40 -50 RGAZ -50 97 98 99 00 01 02 03 04 05 06 07 97 98 99 00 rgast 1 1t roilt 2 2t h11t a110 a11 12t 1 b11 h11t 1 h12t a120 a22 1t 1 2t 1 b22 h12t 1 0 h22t a22 a33 22t 1 b33 h22t 1 01 02 03 04 05 06 07 DIAGONAL REPRESENTATION ESTIMATION OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: Diagonal VECH Sample: 1997M02 2007M01 Included observations: 120 Total system (balanced) observations 240 Disturbance assumption: Student's t distribution Convergence achieved after 198 iterations rgast 1 1t C(1) C(2) Coefficient 1.259 2.438 Std. Error z-Statistic 1.082 1.164 0.733 3.328 Prob. 0.245 0.001 C(3) C(4) C(5) C(6) C(7) C(8) C(9) C(10) C(11) Variance Equation Coefficients 63.635 41.552 1.531 0.837 1.852 0.452 9.479 2.224 4.261 0.224 0.179 1.256 -0.056 0.039 -1.421 -0.200 0.081 -2.467 0.352 0.358 0.986 0.858 0.065 13.109 1.067 0.043 24.956 0.126 0.651 0.000 0.209 0.155 0.014 0.324 0.000 0.000 C(12) t-Distribution (Degree of Freedom) 18.732 26.610 0.704 0.482 Log likelihood Avg. log likelihood Akaike info criterion -890.0937 Schwarz criterion -3.708724 Hannan-Quinn criter. 15.03489 15.31364 15.1481 roilt 2 2t 0 h11t a11 a 11 12t 1 b11 h11t 1 0 h12t a12 a22 1t 1 2t 1 b22 h12t 1 0 h22t a22 a33 22t 1 b33 h22t 1 DIAGONAL REPRESENTATION ESTIMATION OIL & NATURAL GAS PRICES Covariance specification: Diagonal VECH GARCH = M + A1.*RESID(-1)*RESID(-1)' + B1.*GARCH(-1) M is an indefinite matrix A1 is an indefinite matrix B1 is an indefinite matrix C(1) C(2) rgast 1 1t ROIL = C(2) h11t a110 a11 12t 1 b11 h11t 1 Substituted Coefficients: ===================== RGAZ = 1.258 h12t a120 a22 1t 1 2t 1 b22 h12t 1 roilt 2 2t 0 h22t a22 a33 22t 1 b33 h22t 1 ROIL = 2.438 Coefficient 1.259 2.438 Coefficient C(3) C(4) C(5) C(6) C(7) C(8) C(9) C(10) C(11) Estimated Equations: ===================== RGAZ = C(1) Variance Equation Coefficients 63.635 M(1,1) 63.63 0.837 M(1,2) 0.84 9.479 M(2,2) 9.48 0.224 A1(1,1) 0.22 -0.056 A1(1,2) -0.06 -0.200 A1(2,2) -0.20 0.352 B1(1,1) 0.35 0.858 B1(1,2) 0.86 1.067 B1(2,2) 1.07 Variance-Covariance Representation: ===================== GARCH = M + A1.*RESID(-1)*RESID(-1)' + B1.*GARCH(-1) Variance and Covariance Equations: ===================== GARCH1 = M(1,1) + A1(1,1)*RESID1(-1)^2 + B1(1,1)*GARCH1(-1) GARCH2 = M(2,2) + A1(2,2)*RESID2(-1)^2 + B1(2,2)*GARCH2(-1) COV1_2 = M(1,2) + A1(1,2)*RESID1(-1)*RESID2(-1) + B1(1,2)*COV1_2(-1) Substituted Coefficients: ===================== GARCH1 = 63.634 + 0.224*RESID1(-1)^2 + 0.352*GARCH1(-1) GARCH2 = 9.479 -0.199*RESID2(-1)^2 + 1.066*GARCH2(-1) COV1_2 = 0.837 -0.055*RESID1(-1)*RESID2(-1) + 0.857*COV1_2(-1) DIAGONAL REPRESENTATION VOLATILITY FORECAST OF OIL & NATURAL GAS PRICES Conditional Correlation Conditional Covariance Cor(RGAZ,ROIL) Var(RGAZ) 1.00 800 0.75 600 0.50 0.25 400 0.00 -0.25 200 -0.50 0 97 98 99 00 01 02 03 04 05 -0.75 06 -1.00 Cov(RGAZ,ROIL) 97 Var(ROIL) 300 98 240 200 200 160 100 120 80 0 40 -100 0 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 99 00 01 02 03 04 05 06 DIAGONAL REPRESENTATION REVISED MODEL ESTIMATION OIL & NATURAL GAS PRICES rgast 1 roilt 1 1t roilt 2 2t 0 h11t a11 a 11 12t 1 b11 h11t 1 h12t a a22 1t 1 2t 1 b22 h12t 1 0 12 h22t a a33 0 22 2 2t 1 b33 h22t 1 DIAGONAL REPRESENTATION REVISED MODEL ESTIMATION OIL & NATURAL GAS PRICES Covariance specification: Diagonal VECH GARCH = M + A1.*RESID(-1)*RESID(-1)' + B1.*GARCH(-1) M is a scalar A1 is a rank one matrix B1 is a rank one matrix Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: Diagonal VECH Sample: 1997M03 2007M01 Included observations: 119 Total system (balanced) observations 238 Disturbance assumption: Student's t distribution Convergence achieved after 26 iterations Tranformed Variance Coefficients Coefficient C(1) C(2) C(3) Coefficient 0.548 0.240 1.513 Std. Error 0.873 0.088 0.782 z-Statistic 0.628 2.730 1.934 Prob. 0.530 0.006 0.053 C(4) C(5) C(6) C(7) C(8) 1.974 0.474 -0.110 0.890 0.983 1.865 0.107 0.123 0.041 0.020 1.059 4.407 -0.894 21.543 48.854 0.290 0.000 0.371 0.000 0.000 t-Distribution (Degree of Freedom) C(9) Log likelihood Avg. log likelihood Akaike info criterion 6.928 2.791 -892.0641 Schwarz criterion -3.748168 Hannan-Quinn criter. 15.14393 2.482 0.013 15.35412 15.22928 M A1(1,1) A1(1,2) A1(2,2) B1(1,1) B1(1,2) B1(2,2) 1.974 0.224 -0.052 0.012 0.793 0.875 0.966 DIAGONAL REPRESENTATION REVISED MODEL ESTIMATION OIL & NATURAL GAS PRICES Covariance specification: Diagonal VECH GARCH = M + A1.*RESID(-1)*RESID(-1)' + B1.*GARCH(-1) M is a scalar A1 is a rank one matrix B1 is a rank one matrix C(1) C(2) C(3) C(4) C(5) C(6) C(7) C(8) ROIL = C(3) Substituted Coefficients: ===================== RGAZ = 0.548145947197+0.240483266384*ROIL(-1) Coefficient 0.548 0.240 1.513 ROIL = 1.51272216999 Variance-Covariance Representation: ===================== GARCH = M + A1.*RESID(-1)*RESID(-1)' + B1.*GARCH(-1) 1.974 0.474 -0.110 0.890 0.983 Tranformed Variance Coefficients Coefficient M A1(1,1) A1(1,2) A1(2,2) B1(1,1) B1(1,2) B1(2,2) Estimated Equations: ===================== RGAZ = C(1)+C(2)*ROIL(-1) 1.974 0.224 -0.052 0.012 0.793 0.875 0.966 Variance and Covariance Equations: ===================== GARCH1 = M + A1(1,1)*RESID1(-1)^2 + B1(1,1)*GARCH1(-1) GARCH2 = M + A1(2,2)*RESID2(-1)^2 + B1(2,2)*GARCH2(-1) COV1_2 = M + A1(1,2)*RESID1(-1)*RESID2(-1) + B1(1,2)*COV1_2(-1) Substituted Coefficients: ===================== GARCH1 = 1.974 + 0.224*RESID1(-1)^2 + 0.792*GARCH1(-1) GARCH2 = 1.974 + 0.012*RESID2(-1)^2 + 0.965*GARCH2(-1) COV1_2 = 1.974 -0.052*RESID1(-1)*RESID2(-1) + 0.875*COV1_2(-1) DIAGONAL REPRESENTATION VOLATILITY FORECAST OF OIL & NATURAL GAS PRICES Conditional Correlation Conditional Covariance Cor(RGAZ,ROIL) Var(RGAZ) .6 800 .5 .4 600 .3 400 .2 .1 200 .0 0 97 98 99 00 01 02 03 04 05 -.1 06 -.2 Cov(RGAZ,ROIL) 97 Var(ROIL) 120 95 80 90 40 85 0 80 -40 98 75 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 99 00 01 02 03 04 05 06 DIAGONAL REPRESENTATION TARCH MODEL ESTIMATION OIL & NATURAL GAS PRICES rgast 1 roilt 1 1t I1=1 if 1t-1<0 =0 otherwise roilt 2 2t I2=1 if 2t-1<0 =0 otherwise h11t a110 a11 12t 1 d11 I1 12t 1 b11 h11t 1 h12t a a22 1t 1 2t 1 d 21 ( I )(I 0 12 2 1 1t 1 2 2 2t 1 ) b22 h12t1 0 h22t a22 a33 22t 1 d 33 I 2 22t 1 b33 h22t 1 H A t 1 t'1 A D t 1I ( t 1 0)( t 1I ( t 1 0))D BH(1) DIAGONAL REPRESENTATION TARCH MODEL ESTIMATION OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: Diagonal VECH TARCH Date: 08/05/08 Time: 18:36 Sample: 1997M03 2007M01 Included observations: 119 Total system (balanced) observations 238 Disturbance assumption: Student's t distribution Presample covariance: backcast (parameter =0.5) Convergence achieved after 169 iterations Coefficient Std. Error z-Statistic C(1) C(2) C(3) 0.440 0.213 1.543 1.165 0.116 0.775 0.378 1.825 1.990 Prob. 0.706 0.068 0.047 Variance Equation Coefficients C(4) C(5) C(6) C(7) C(8) 1.465 -0.013 0.224 -0.234 0.971 1.116 0.028 0.089 0.056 0.016 1.313 -0.461 2.506 -4.155 61.584 0.189 0.645 0.012 0.000 0.000 t-Distribution (Degree of Freedom) C(9) 4.810 1.774 2.711 Log likelihood -898.265 Schwarz criterion Avg. log likelihood -3.77422 Hannan-Quinn criter. Akaike info criterion 15.24815 Covariance specification: Diagonal VECH GARCH = M + A1.*RESID(-1)*RESID(-1)' + D1.*(RESID(-1)*(RESID( -1)<0))*(RESID(-1)*(RESID(-1)<0))'D1.*(RESID(-1)*(RESID(-1)<0)) *(RESID(-1)*(RESID(-1)<0))' + B1.*GARCH(-1) M is a scalar A1 is a scalar D1 is a rank one matrix B1 is a scalar 0.007 15.45834 15.3335 Tranformed Variance Coefficients Coefficient M A1 D1(1,1) D1(1,2) D1(2,2) B1 1.465 -0.013 0.050 -0.052 0.055 0.971 DIAGONAL REPRESENTATION TARCH MODEL ESTIMATION OIL & NATURAL GAS PRICES Coefficient Estimated Equations: ===================== RGAZ = C(1)+C(2)*ROIL(-1) C(1) C(2) C(3) 0.440 0.213 1.543 Tranformed Variance Coefficients Coefficient Variance Equation Coefficients ROIL = C(3) Substituted Coefficients: ===================== RGAZ = 0.44021432107+0.212553531477*ROIL(-1) ROIL = 1.54264508993 C(4) C(5) C(6) C(7) C(8) 1.465 -0.013 0.224 -0.234 0.971 M A1 D1(1,1) D1(1,2) D1(2,2) B1 1.465 -0.013 0.050 -0.052 0.055 0.971 Variance-Covariance Representation: ===================== GARCH = M + A1.*RESID(-1)*RESID(-1)' + D1.*(RESID(-1)*(RESID(-1)<0))*(RESID(-1)*(RESID(-1)<0))'D1.*(RESID(-1)*(RESID(-1)<0))*(RESID(-1)*(RESID(-1)<0))' + B1.*GARCH Variance and Covariance Equations: ===================== GARCH1 = M + A1*RESID1(-1)^2 + D1(1,1)*RESID1(-1)^2*(RESID1(-1)<0) + B1*GARCH1(-1) GARCH2 = M + A1*RESID2(-1)^2 + D1(2,2)*RESID2(-1)^2*(RESID2(-1)<0) + B1*GARCH2(-1) COV1_2 = M + A1*RESID1(-1)*RESID2(-1) + D1(1,2)*RESID1(-1)*(RESID1(-1)<0)*RESID2(-1)*(RESID2(-1)<0) + B1*COV1_2(-1) Substituted Coefficients: ===================== GARCH1 = 1.464 - 0.013*RESID1(-1)^2 + 0.050*RESID1(-1)^2*(RESID1(-1)<0) + 0.971*GARCH1(-1) GARCH2 = 1.465 - 0.013*RESID2(-1)^2 + 0.054*RESID2(-1)^2*(RESID2(-1)<0) + 0.971*GARCH2(-1) COV1_2 = 1.464 - 0.013*RESID1(-1)*RESID2(-1) -0.052*RESID1(-1)*(RESID1(-1)<0)*RESID2(-1)*(RESID2(-1)<0) + 0.971*COV1_2(-1) DIAGONAL REPRESENTATION TARCH MODEL FORECAST OIL & NATURAL GAS PRICES VOLATILITY Conditional Correlation Conditional Covariance Cor(RGAZ,ROIL) Var(RGAZ) 600 .8 500 .6 400 .4 300 .2 200 .0 100 97 98 99 00 01 02 03 04 05 06 -.2 97 Cov(RGAZ,ROIL) 98 99 Var(ROIL) 160 140 120 120 80 100 40 80 0 -40 60 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 00 01 02 03 04 05 06 BEKK REPRESENTATION Engle and Kroner(1995) developed the Baba(1990) approach. BEKK representation of multivariate GARCH improves on both the VECH and diagonal representation, since H is almost guaranteed to be positive definite. BEKK representation require more parameters than Diagonal rep. but less parameters than VECH. It is more general than diagonal rep. as it allows for interaction effects that diagonal rep. does not. 2 12t 1 22t 1 a11 a12 b11 b12 h11t 1 h12 t 1 b11 b12 h11t h12 t a110 a120 a11 a12 1t 1 2 2 0 0 h h b b 2 h h a a a a b b 21 22 21 22 12 t 1 22 t 1 21 22 12 t 22 t a12 a22 21 22 1t 1 2t 1 2t 1 H A A BH(1)B ' t 1 t 1 BEKK REPRESENTATION OF OIL & NATURAL GAS PRICES rgast 1 1t roilt 2 2t h11t a 2 2 11 1t 1 b h 2 11 11t 1 h12t a11a22 1t 1 2t 1 b11b22 h12t 1 2 2 h22t a22 2t 1 b222 h22t1 BEKK ESTIMATION OF OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: BEKK Sample: 1997M02 2007M01 Included observations: 120 Total system (balanced) observations 240 Disturbance assumption: Student's t distribution Presample covariance: backcast (parameter =0.5) Convergence achieved after 25 iterations Coefficient C(1) C(2) 0.890 1.346 Std. Error z-Statistic Prob. 0.874 0.786 1.019 1.712 0.308 0.087 Variance Equation Coefficients C(3) C(4) C(5) C(6) C(7) 1.342 0.427 -0.079 0.908 0.987 1.367 0.093 0.127 0.032 0.014 0.982 4.605 -0.623 28.189 71.875 0.326 0.000 0.533 0.000 0.000 t-Distribution (Degree of Freedom) C(8) 6.729 Covariance specification: BEKK GARCH = M + A1*RESID(-1)*RESID(-1)'*A1 + B1*GARCH(-1)*B1 M is a scalar A1 is diagonal matrix B1 is diagonal matrix 2.788 2.414 Log likelihood -902.4941 Schwarz criterion Avg. log likelihood -3.760392 Hannan-Quinn criter. Akaike info criterion 15.1749 0.016 15.36073 15.25037 Tranformed Variance Coefficients Coefficient M A1(1,1) A1(2,2) B1(1,1) B1(2,2) 1.342 0.427 -0.079 0.908 0.987 BEKK ESTIMATION OF OIL & NATURAL GAS PRICES Estimated Equations: ===================== RGAZ = C(1) ROIL = C(2) Substituted Coefficients: ===================== RGAZ = 0.890273745039 ROIL = 1.34556553188 Variance-Covariance Representation: ===================== Coefficient C(1) C(2) 0.890 1.346 Variance Equation Coefficients C(3) C(4) C(5) C(6) C(7) 1.342 0.427 -0.079 0.908 0.987 GARCH = M + A1*RESID(-1)*RESID(-1)'*A1 + B1*GARCH(-1)*B1 Variance and Covariance Equations: ===================== GARCH1 = M + A1(1,1)^2*RESID1(-1)^2 + B1(1,1)^2*GARCH1(-1) Tranformed Variance Coefficients GARCH2 = M + A1(2,2)^2*RESID2(-1)^2 + B1(2,2)^2*GARCH2(-1) M A1(1,1) A1(2,2) B1(1,1) B1(2,2) COV1_2 = M + A1(1,1)*A1(2,2)*RESID1(-1)*RESID2(-1) + B1(1,1)*B1(2,2)*COV1_2(-1) Substituted Coefficients: ===================== GARCH1 = 1.342+0.182*RESID1(-1)^2+0.824*GARCH1(-1) GARCH2 = 1.342+0.0063*RESID2(-1)^2+0.974*GARCH2(-1) COV1_2 = 1.342 -0.034*RESID1(-1)*RESID2(-1) + 0.896*COV1_2(-1) Coefficient 1.342 0.427 -0.079 0.908 0.987 REVISED BEKK REPRESENTATION OF OIL & NATURAL GAS PRICES rgast 1 roilt 1 1t roilt 2 2t h11t a 2 2 11 1t 1 b h 2 11 11t 1 h12t a11a22 1t 1 2t 1 b11b22 h12t 1 2 2 h22t a22 2t 1 b222 h22t1 REVISED BEKK REPRESENTATION OF OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: BEKK Sample: 1997M03 2007M01 Included observations: 119 Total system (balanced) observations 238 Disturbance assumption: Student's t distribution Presample covariance: backcast (parameter =0.5) Convergence achieved after 26 iterations Coefficient Std. Error z-Statistic C(1) C(2) C(3) 0.548 0.240 1.513 0.873 0.088 0.782 0.628 2.730 1.934 1.974 0.474 -0.110 0.890 0.983 1.865 0.107 0.123 0.041 0.020 1.059 4.407 -0.894 21.543 48.854 0.530 0.006 0.053 0.290 0.000 0.371 0.000 0.000 t-Distribution (Degree of Freedom) C(9) 6.928 2.791 Tranformed Variance Coefficients Prob. Variance Equation Coefficients C(4) C(5) C(6) C(7) C(8) Covariance specification: BEKK GARCH = M + A1*RESID(-1)*RESID(-1)'*A1 + B1*GARCH(-1)*B1 M is a scalar A1 is diagonal matrix B1 is diagonal matrix 2.482 0.013 Log likelihood -892.0641 Schwarz criterion Avg. log likelihood -3.748168 Hannan-Quinn criter. Akaike info criterion 15.14393 15.35412 15.22928 Coefficient Std. Error z-Statistic M A1(1,1) A1(2,2) B1(1,1) B1(2,2) 1.974 0.474 -0.110 0.890 0.983 1.865 0.107 0.123 0.041 0.020 1.059 4.407 -0.894 21.543 48.854 Prob. 0.290 0.000 0.371 0.000 0.000 REVISED BEKK REPRESENTATION OF OIL & NATURAL GAS PRICES Estimated Equations: ===================== RGAZ = C(1)+C(2)*ROIL(-1) ROIL = C(3) Substituted Coefficients: ===================== RGAZ = 0.548145947197+0.240483266384*ROIL(-1) ROIL = 1.51272216999 Variance-Covariance Representation: ===================== GARCH = M + A1*RESID(-1)*RESID(-1)'*A1 + B1*GARCH(-1)*B1 Variance and Covariance Equations: ===================== GARCH1 = M + A1(1,1)^2*RESID1(-1)^2 + B1(1,1)^2*GARCH1(-1) GARCH2 = M + A1(2,2)^2*RESID2(-1)^2 + B1(2,2)^2*GARCH2(-1) COV1_2 = M + A1(1,1)*A1(2,2)*RESID1(-1)*RESID2(-1) + B1(1,1)*B1(2,2)*COV1_2(-1) Substituted Coefficients: ===================== GARCH1 = 1.974+0.224*RESID1(-1)^2+0.792*GARCH1(-1) GARCH2 = 1.974+0.012*RESID2(-1)^2+0.965*GARCH2(-1) COV1_2 = 1.974 -0.052*RESID1(-1)*RESID2(-1) + 0.875*COV1_2(-1) REVISED BEKK FORECASTING OF OIL & NATURAL Conditional Correlation GAS PRICES VOLATILITY Cor(RGAZ,ROIL) .6 Conditional Covariance .5 Var(RGAZ) 800 .4 .3 600 .2 400 .1 .0 200 -.1 0 97 98 99 00 01 02 03 04 05 -.2 06 97 Cov(RGAZ,ROIL) 98 Var(ROIL) 120 95 80 90 40 85 0 80 -40 75 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 99 00 01 02 03 04 05 06 CONSTANT CORRELATION REPRESENTATION Bollerslev(1990) employes the conditional corelation matrix R to derive a representation of the multivariate GARCH model. In his R matrix, Bollerslev restricts the conditional correlations to be equal to the correlation coefficients between variables, which are simply constants. Thus R is constant over time. This representation has the advantage that H will be positive definite. CONSTANT CORRELATION REPRESENTATION 12 13 .....1N 1 1 ..... 23 2N R 12 ................................ .......... ...... 1 N1 N 2 h11 0 0 .... ..0 t 0 h22 t 0 .......0 H . .................................. 0 hNN t 0 h11 0 0 .... ..0 t 1 ..... 12 13 1 N 12 1 23 ..... 2 N 0 h22 t 0 .......0 .......... .......... .......... .. . .................................. ................1 0 0 hNN t N 1 N 2 The individual variance terms hiit are taken to be individual GARCH processes CONSTANT CORRELATION REPRESENTATION OF OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: Constant Conditional Correlation Sample: 1997M02 2007M01 Included observations: 120 Total system (balanced) observations 240 Disturbance assumption: Student's t distribution Presample covariance: backcast (parameter =0.5) Convergence achieved after 21 iterations Coefficient Std. Error z-Statistic C(1) C(2) 0.938 2.078 1.037 0.605 0.904 3.435 Prob. 0.366 0.001 Variance Equation Coefficients C(3) C(4) C(5) C(6) C(7) C(8) C(9) 53.143 0.152 0.449 8.735 -0.145 1.033 0.154 50.532 0.148 0.454 1.906 0.041 0.020 0.102 1.052 1.027 0.989 4.583 -3.507 51.343 1.516 0.293 0.305 0.323 0.000 0.001 0.000 0.130 t-Distribution (Degree of Freedom) C(10) 10.97096 8.493965 1.291618 0.1965 Log likelihood -893.8184 Schwarz criterion Avg. log likelihood -3.724243 Hannan-Quinn criter. Akaike info criterion 15.06364 15.29593 15.15797 Covariance specification: Constant Conditional Correlation GARCH(i) = M(i) + A1(i)*RESID(i)(-1)^2 + B1(i)*GARCH(i)(-1) COV(i,j) = R(i,j)*@SQRT(GARCH(i)*GARCH(j)) Tranformed Variance Coefficients Coefficient Std. Error z-Statistic Prob. M(1) A1(1) B1(1) M(2) A1(2) B1(2) R(1,2) 53.143 0.152 0.449 8.735 -0.145 1.033 0.154 50.532 0.148 0.454 1.906 0.041 0.020 0.102 1.052 1.027 0.989 4.583 -3.507 51.343 1.516 0.293 0.305 0.323 0.000 0.001 0.000 0.130 CONSTANT CORRELATION REPRESENTATION OF OIL & NATURAL GAS PRICES Substituted Coefficients: ===================== RGAZ = 0.937503474235 ROIL = 2.07760000977 Variance and Covariance Representations: ===================== GARCH(i) = M(i) + A1(i)*RESID(i)(-1)^2 + B1(i)*GARCH(i)(-1) COV(i,j) = R(i,j)*@SQRT(GARCH(i)*GARCH(j)) Variance and Covariance Equations: ===================== GARCH1 = C(3) + C(4)*RESID1(-1)^2 + C(5)*GARCH1(-1) Variance Equation Coefficients C(3) C(4) C(5) C(6) C(7) C(8) C(9) 53.143 0.152 0.449 8.735 -0.145 1.033 0.154 Tranformed Variance Coefficients Coefficient GARCH2 = C(6) + C(7)*RESID2(-1)^2 + C(8)*GARCH2(-1) COV1_2 = C(9)*@SQRT(GARCH1*GARCH2) Substituted Coefficients: ===================== GARCH1 = 53.142 + 0.152*RESID1(-1)^2 + 0.448*GARCH1(-1) GARCH2 = 8.735 - 0.145*RESID2(-1)^2 + 1.033*GARCH2(-1) COV1_2 = 0.154*@SQRT(GARCH1*GARCH2) M(1) A1(1) B1(1) M(2) A1(2) B1(2) R(1,2) 53.143 0.152 0.449 8.735 -0.145 1.033 0.154 CONSTANT CORRELATION FORECAST OF OIL & NATURAL GAS PRICES VOLATILITY Conditional Correlation Conditional Covariance Cor(RGAZ,ROIL) Var(RGAZ) 600 .162 500 .160 .158 400 .156 300 .154 200 .152 100 .150 0 97 98 99 00 01 02 03 04 05 .148 06 .146 Cov(RGAZ,ROIL) 97 Var(ROIL) 50 200 40 150 30 100 20 50 10 0 0 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 98 99 00 01 02 03 04 05 06 REVISED CONSTANT CORRELATION REPRESENTATION OF OIL & NATURAL GAS PRICES Estimation Method: ARCH Maximum Likelihood (Marquardt) Covariance specification: Constant Conditional Correlation Sample: 1997M03 2007M01 Included observations: 119 Total system (balanced) observations 238 Disturbance assumption: Student's t distribution Presample covariance: backcast (parameter =0.5) Convergence achieved after 26 iterations Coefficient C(1) C(2) C(3) 0.781 0.199 2.233 Std. Error z-Statistic 1.079 0.109 0.573 0.724 1.834 3.899 Prob. 0.469 0.067 0.000 Variance Equation Coefficients C(4) C(5) C(6) C(7) C(8) C(9) C(10) 64.398 0.198 0.314 7.293 -0.135 1.053 0.120 49.848 0.170 0.432 1.229 0.033 0.020 0.103 1.292 1.163 0.727 5.936 -4.062 51.973 1.166 0.196 0.245 0.468 0.000 0.000 0.000 0.244 t-Distribution (Degree of Freedom) C(11) 13.509 11.166 1.210 Log likelihood -882.0092 Schwarz criterion Avg. log likelihood -3.705921 Hannan-Quinn criter. Akaike info criterion 15.00856 Covariance specification: Constant Conditional Correlation GARCH(i) = M(i) + A1(i)*RESID(i)(-1)^2 + B1(i)*GARCH(i)(-1) COV(i,j) = R(i,j)*@SQRT(GARCH(i)*GARCH(j)) 0.226 15.26545 15.11287 Tranformed Variance Coefficients Coefficient Std. Error M(1) A1(1) B1(1) M(2) A1(2) B1(2) R(1,2) 64.398 0.198 0.314 7.293 -0.135 1.053 0.120 49.848 0.170 0.432 1.229 0.033 0.020 0.103 z-Statistic Prob. 1.292 1.163 0.727 5.936 -4.062 51.973 1.166 0.196 0.245 0.468 0.000 0.000 0.000 0.244 REVISED CONSTANT CORRELATION REPRESENTATION OF OIL & NATURAL GAS PRICES Estimated Equations: ===================== RGAZ = C(1)+C(2)*ROIL(-1) Variance Equation Coefficients ROIL = C(3) Substituted Coefficients: ===================== RGAZ = 0.781034818457+0.199474480682*ROIL(-1) ROIL = 2.23319899793 Variance and Covariance Representations: ===================== GARCH(i) = M(i) + A1(i)*RESID(i)(-1)^2 + B1(i)*GARCH(i)(-1) COV(i,j) = R(i,j)*@SQRT(GARCH(i)*GARCH(j)) Variance and Covariance Equations: ===================== GARCH1 = C(4) + C(5)*RESID1(-1)^2 + C(6)*GARCH1(-1) GARCH2 = C(7) + C(8)*RESID2(-1)^2 + C(9)*GARCH2(-1) COV1_2 = C(10)*@SQRT(GARCH1*GARCH2) Substituted Coefficients: ===================== GARCH1 = 64.398 + 0.197*RESID1(-1)^2 + 0.314*GARCH1(-1) GARCH2 = 7.292 - 0.135*RESID2(-1)^2 + 1.052*GARCH2(-1) COV1_2 = 0.119*@SQRT(GARCH1*GARCH2) C(4) C(5) C(6) C(7) C(8) C(9) C(10) 64.398 0.198 0.314 7.293 -0.135 1.053 0.120 Tranformed Variance Coefficients Coefficient M(1) A1(1) B1(1) M(2) A1(2) B1(2) R(1,2) 64.398 0.198 0.314 7.293 -0.135 1.053 0.120 REVISED CONSTANT CORRELATION FORECAST OF OIL & NATURAL GAS PRICES VOLATILITY Conditional Correlation Conditional Covariance Cor(RGAZ,ROIL) .126 Var(RGAZ) 600 .124 500 .122 400 .120 300 .118 200 .116 100 .114 0 97 98 99 00 01 02 03 04 05 06 .112 97 Cov(RGAZ,ROIL) 98 Var(ROIL) 25 160 20 120 15 80 10 40 5 0 0 97 98 99 00 01 02 03 04 05 06 97 98 99 00 01 02 03 04 05 06 99 00 01 02 03 04 05 06 FACTOR VOLATILITY MODELS Principal Component Analysis Principal Component Analysis We collect 120 financial ratios in order to asses financial health of the firms. How can we reduce these ratios a few indices? The production control department collect several measures in order to control process. Can we develop some indices in order to summarize the process outcomes? In order to carry out efficient regression analysis we have to reduce multicollinearity among the explanatory variables if it exists. Can we generate some new indices in order to get orthogonal explanatory series that also contain most of the information of the original variables? Principal Component Analysis in Finance To reduce number of risk factors to a manageable dimension. For example, instead of 60 yields of different maturities as risk factors, we might use just 3 principal component. To identy the key sources of risk. Typically the most important risk factors are parallel shifts, changes in slope and changes in convexity of the curves. To facilitate the measurement of portfolio risk, for instance by introducing scenarios on the movements in the major risk factors. Basics & Background Ax x A is square matrix a11 a12 a13 x1 x1 a 21 a 22 a 23 x 2 x 2 a 31 a 32 a 33 x 3 x 3 n Basic properties A is a scalar quantityeigenvalue u normalized eigenvector n i Tr (A) i 1 i 1 u1' u 2 0 X is a column vector u1' u1 1 i Basics & Background IF matrix A composes of some observed x values Pricipal Component Scores yi1 u1' ( x i x ) yi 2 u '2 ( x i x ) .................... yin u 'n ( x i x ) MATHEMATICAL BACKGROUND n A nxn ' i u i u i i 1 ' ' A nxn 1u1u1 2u 2u 2 A 1 n i 1 A 1 A nxn square matrix ' ...... n u n u n 1 ' uiui i 1 ' 1 ' 1 ' u1u1 u 2u 2 ...... unun 1 2 n A BASIC EXAMPLE OF EIGENVALUES AND EIGENVECTORS Ax x a11 a12 x1 x1 a 21 a 22 x 2 x 2 2 1 1 5 1 5 3 4 3 15 3 2 1 1 1 1 1 3 4 1 1 1 - 0.70710678 - 0.31622777 , 0.70710678 - 0.94868330 Normalization 12 32 10 1 .3162 1 10 * 3 . 9486 12 (1)2 2 1 1 .7071 2 * 1 . 7071 MATHEMATICAL EXAMPLE 2 1 A 3 4 2 1 A I 0 3 4 8 < 8 < 8 <8< (2 ) * (4 ) 3 *1 0 l ® 1 , l ® 5 - 1, 1 , 1, 3 U1 U2 Basics & Background • Eigenvalue and Eigenvector: – Eigen originates in the German language and can be loosely translated as “of itself” – Thus an Eigenvalue of a matrix could be conceptualized as a “value of itself” – Eigenvalues and Eigenvectors are utilized in a wide range of applications (PCA, calculating a power of a matrix, finding solutions for a system of differential equations, and growth models) GEOMETRICAL APPROACH TO PRINCIPAL COMPONENT ANALYSIS x2 Mean corrected data x1 AXIS ROTATION x2 x1 AXIS ROTATION DIMEMSION REDUCTION x2’ x1’ x2 x1 AXIS ROTATION A x2 x1’ x2’ x1 X1’ = x1*cos + x2*sin X2’ = -x1*sin + x2*cos AXIS ROTATION = 0 Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x2 x1 8 16 10 12 6 13 2 11 8 10 -1 9 4 8 6 7 -3 5 -1 3 -3 2 0 0 3 8 23.091 21.091 Total Variance= Covariance15.083 Correlation 0.746 X1 ’ = x1Cor 8 4 5 3 2 1 0 -1 -3 -5 -6 -8 0 23.091 x2Cor 5 7 3 -1 5 -4 1 3 -6 -4 -6 -3 0 21.091 44.182 15.0833 0.746 0 x1' 8.000 4.000 5.000 3.000 2.000 1.000 0.000 -1.000 -3.000 -5.000 -6.000 -8.000 0.00 23.091 0 x2' 5.000 7.000 3.000 -1.000 5.000 -4.000 1.000 3.000 -6.000 -4.000 -6.000 -3.000 0.00 21.091 44.182 Total Variance= 52% Share of x1'= Correlation x1*cos 0 + x2*sin0 X2’ = -x1*sin 0 + x2*cos0 0.746 AXIS ROTATION = 10° Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x1 16 12 13 11 10 9 8 7 5 3 2 0 8 23.091 x2 8 10 6 2 8 -1 4 6 -3 -1 -3 0 3 21.091 Total Variance= Covariance15.0833 Correlation 0.746 X1’ = x1Cor 8 4 5 3 2 1 0 -1 -3 -5 -6 -8 0 23.091 x2Cor 5 7 3 -1 5 -4 1 3 -6 -4 -6 -3 0 21.091 44.182 15.0833 0.746 10 0.2 x1' 8.746 5.154 5.445 2.781 2.837 0.291 0.174 -0.464 -3.996 -5.618 -6.950 -8.399 0.00 28.656 Total Variance= x2' 3.536 6.200 2.087 -1.506 4.577 -4.113 0.985 3.128 -5.388 -3.071 -4.868 -1.566 0.00 15.526 44.182 Share of x1'= 65% Correlation x1*cos10 + x2*sin10 X2’ = -x1*sin10 + x2*cos10 0.717 AXIS ROTATION = 30° Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x1 x2 16 8 12 10 13 6 11 2 10 8 9 -1 8 4 7 6 5 -3 3 -1 2 -3 0 0 8 3 23.091 21.091 Total Variance= Covariance15.083 Correlation 0.746 X1’ = x1Cor 8 4 5 3 2 1 0 -1 -3 -5 -6 -8 0 23.091 x2Cor 5 7 3 -1 5 -4 1 3 -6 -4 -6 -3 0 21.091 44.182 15.0833 0.746 30 1 x1' 9.428 6.963 5.830 2.099 4.231 -1.133 0.500 0.633 -5.597 -6.330 -8.196 -8.429 0.00 36.837 Total Variance= x2' 0.333 4.064 0.100 -2.365 3.331 -3.964 0.866 3.098 -3.698 -0.966 -2.198 1.400 0.00 7.345 44.182 Share of x1'= 83% Correlation x1*cos30 + x2*sin30 X2’ = -x1*sin30 + x2*cos30 0.448 AXIS ROTATION = 40° Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x1 x2 16 8 12 10 13 6 11 2 10 8 9 -1 8 4 7 6 5 -3 3 -1 2 -3 0 0 8 3 23.091 21.091 Total Variance= x1Cor x2Cor 40 1 8 5 4 7 5 3 3 -1 2 5 1 -4 0 1 -1 3 -3 -6 -5 -4 -6 -6 -8 -3 0 0 23.091 21.091 44.182 Total Variance= Covariance 15.083 15.0833 Correlation0.746 0.746 X1’ = x1*cos40 + Share of x1'= x1' 9.343 7.563 5.759 1.656 4.745 -1.804 0.643 1.161 -6.154 -6.401 -8.453 -8.058 0.00 38.468 44.182 87% Correlation x2*sin40 X2’ = -x1*sin40 + x2*cos40 x2' -1.309 2.794 -0.914 -2.694 2.546 -3.708 0.766 2.941 -2.670 0.147 -0.743 2.841 0.00 5.714 0.127 AXIS ROTATION = 44° Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x1 16 12 13 11 10 9 8 7 5 3 2 0 8 23.091 x2 8 10 6 2 8 -1 4 6 -3 -1 -3 0 3 21.091 Total Variance= Covariance 15.08333 Correlation 0.746 X1 ’ = x1Cor 8 4 5 3 2 1 0 -1 -3 -5 -6 -8 0 23.091 x2Cor 5 7 3 -1 5 -4 1 3 -6 -4 -6 -3 0 21.091 44.182 15.08333 0.746 0.2404 0.755239 Total Variance= Share of x1'= x1*cos44 + x2*sin44 X2’ = -x1*sin44 + x2*cos44 x1' 9.252 7.711 5.697 1.499 4.884 -2.014 0.685 1.328 -6.297 -6.382 -8.481 -7.881 0.00 38.576 x2' -1.843 2.355 -1.243 -2.784 2.270 -3.598 0.728 2.870 -2.312 0.515 -0.256 3.299 0.00 5.606 44.182 87% Correlation 0.000 AXIS ROTATION = 70° Observation 1 2 3 4 5 6 7 8 9 10 11 12 Mean Variance x1 x2 16 8 12 10 13 6 11 2 10 8 9 -1 8 4 7 6 5 -3 3 -1 2 -3 0 0 8 3 23.091 21.091 Total Variance= Covariance 15.083 Correlation0.746 X1 ’ = x1Cor x2Cor 70 1 8 5 4 7 5 3 3 -1 2 5 1 -4 0 1 -1 3 -3 -6 -5 -4 -6 -6 -8 -3 0 0 23.091 21.091 44.182 Total Variance= 15.0833 0.746 Share of x1'= x1*cos70 + x2*sin70 X2’ = -x1*sin70 + x2*cos70 x1' 7.438 7.947 4.531 0.088 5.383 -3.415 0.939 2.476 -6.665 -5.471 -7.692 -5.559 0.00 31.918 44.182 72% Correlation x2' -5.803 -1.360 -3.670 -3.161 -0.166 -2.310 0.343 1.967 0.763 3.327 3.581 6.488 0.00 12.264 -0.669 FINDING OPTIMAL Portion of x1’ over total variance ° While x1 dimension explains 52% of the total variance, When ° = 40, new x1’ dimension explains 87% of total variance 87% * ° ANALYTICAL APPROACH TO PRINCIPAL COMPONENT ANALYSIS Assume there are p variables 1 = w11*x1+w12*x2+....+w1p*xp 2 = w21*x1+w22*x2+....+w2p*xp ........................................ p = wp1*x1+wp2*x2+....+wpp*xp s are principal components and wij is the weight of the jth variables on the ith principal component Var(1) > Var(2)> ... >Var(p) wi12 + wi22+....+wip2 = 1 i=1,2,...p wi1*wj1+wi2*wj2+....+wip*wjp = 0 for all ij ANALYTICAL APPROACH TO PRINCIPAL COMPONENT ANALYSIS • Assume there are p variables 1 = w11*x1+w12*x2+....+w1p*xp 2 = w21*x1+w22*x2+....+w2p*xp ........................................ p = wp1*x1+wp2*x2+....+wpp*xp – s are principal components and wij is the weight of the jth variables on the ith principal component X1’ = cos * x1 + sin * x2 X2’ = -sin * x1 + cos * x2 1 = w11*x1 + w12*x2 2 = w21*x1 + w22*x2 MATRIX ALGEBRA APPROACH TO PRINCIPAL COMPONENT ANALYSIS Assume there are p variables 1 = w11*x1+w12*x2+....+w1p*xp 2 = w21*x1+w22*x2+....+w2p*xp ........................................ p = wp1*x1+wp2*x2+....+wpp*xp MATRIX REPRESANTATION 1 = W1’X 2 = W2’X W2’*W2=1 W2’*W1=0 MATRIX ALGEBRA APPROACH TO PRINCIPAL COMPONENT ANALYSIS Var(1) = Var(W1’X) = W1’S W1 S= The Variance-Covariance matrix of original variables Max. Var(1) = W1’S st. W1 W2’*W2=1 If 1, 2, ... , p are the eigenvalues of S Sw1= 1w1 Var(1)= W1’S W1= W1’ 1w1= 1W1’w1=1 Variance explained by the first principal component = 1/Trace(S) Trace(S) = Sum of all is Variance explained by the first k principal components = (1+2+...+k)/Trace(S) VARIANCE EXTRACTION METHODS VARIANCE-COVARIANCE MATRIX THE SIZE EFFECT INCLUDED THE ANALYSIS CORRELATION MATRIX THE VARIABLES ARE STANDIZED FIRST, SO THAT MEAN= 0 & VARIANCE= 1 OF ALL VARIABLES THE VARIANCE COVARIANCE MATRIX OF THE STANDARDIZED VARIABLES IS THE CORRELATION MATRIX THE SIZE EFFECT EXCLUDED FROM THE ANALYSIS VARIANCE EXTRACTION METHODS A principal component represantation based on the variancecovariance matrix has the advantage of providing a linear factor model for the returns, and not a linear factor model for the standardized returns, as is the case when correlation matrix is used. Standardization makes each variable has common mean and variance, 0 and 1 respectively. A PCA on the covariance matrix captures all the movements in the variables, which may be dominated by the differing volatilities of individual variables. A PCA on the correlation matrix only captures the comovements in returns and ignores their individual volatilities. It is only when all variables have similiar volatilities that the PCA will have similar characteristics. STATISTICAL TESTS in PCA Sample Varianceof ˆi i 2 (n 1) CONFIDENCE INTERVAL FOR THE EIGENVALUES OF ˆi 1 Z[ / 2 ] 2 (n 1) i S ˆi 1 Z[ / 2] 2 (n 1) STATISTICAL TESTS in PCA H 0 : k 1 k 2 k 3 .... k m H1 : at least one of them X [(n k 1) (2m m 2) / 6m](m ln 2 2 k m ln ˆ ) j k 1 k m where X 2 ~ 2 ˆ ln j j k 1 m with dof (m 1)(m 2) / 2 If X2 > 2table Reject H0 j STATISTICAL TESTS in PCA If we want to test the last p-k roots H0 : k 1 k 2 k 3 .... k m H1 : at least one of them k X 2 [(n k 1) (2m 2 2m 2) / 6m ( j )] j 1 * (m ln k m ln ˆ ) j k 1 j k m where m p k and X 2 ~ 2 ˆ ln j j k 1 If X2 > 2table m with dof (m 1)(m 2) / 2 Reject H0 EXAMPLE 1= 5.7735 2= 0.9481 3= 0.3564 4= 0.1869 5= 0.1167 6= 0.0967 7= 0.0803 8= 0.0314 n = 292 H0: 2=3 H1: 23 hence k = 1 m = 2 2 * 22 2 2 X [(292 1 1) ](2 ln 0.6523 ln 0.9481 ln 0.3564) 6*2 X 2 289 (0.8547 0.05330 1.0317) 2 X 2 66.56 dof (m 1)(m 2) / 2 2 12 (2) 5.991 for 0.05 66.56 5.991 therefore Reject H 0 HOW MANY PRINCIPAL COMPONENTS? IF ALL THE INFORMATION IS EXTRACTED, P COMPONENTS SHOULD BE SELECTED RESEARCHERS CAN WANT TO ELIMINATE MARGINAL INFORMATION SO THAT ONLY MAIN INFO. IS UNDERLINED EXPLAIN RELATIVELY LARGE PERCENTAGE OF THE TOTAL VARIATION. 70-90% ARE USUALLY SUGGESTED FIGURES. EXCLUDE THOSE PRINCIPLE COMPONENTS WHOSE EIGENVALUES ARE LESS THAN AVERAGE EIGENVALUE. IF CORRELATION MATRIX IS USED, EXCLUDE PC WHOSE EIGENVALUES ARE LESS THAN 1. IF SAMPLE SIZE SMALL, THE CUT OFF POINT CAN BE LOWER, 0.7. USE SCREE PLOT TO CATCH THE “ELBOW” AND THE ELBOW POINTS THE NUMBER OF EIGENVALUES SHOULD BE EXCLUDED. Interpretation of PCs The first principle component captures a common trend in assets or interest rates. If the first PC changes at atime when the other components are fixed, then the returns(or changes in interest rates) all move by roughly the same amount. The second and higher order PC have no intuitive interpretation. But when we use factor rotation, some PC may represent a subgroups of the variables. A Numerical Example • Original data values & mean centered: 12 14 16 19 15 17 11 14.5 16.3 11.9 14 15 14.5 17 13.5 16.3 12.4 16.3 18 11.5 - 2.67 - 0.67 1.33 4.33 0.33 2.33 - 3.67 - 0.17 1.63 - 2.77 - 0.85 0.15 - 0.35 2.15 - 1.35 1.45 - 2.45 1.45 3.15 - 3.35 ORIGINAL & MEAN CENTERED DATA VALUES 20 6 18 4 16 14 2 12 10 0 8 -6 -4 6 -2 0 -2 4 -4 2 0 0 5 10 15 20 ORIGIN SHIFTING -6 2 4 6 A Numerical Example • Covariance Matrix results in: 6.3845555 4.1227777 4.1227777 4.2961111 • Which has Eigenvalues and Eigenvectors of: 0 1.08737 9.593297 0 0.6141953 - 0.7891540 - 0.7891540 - 0.6141953 A Numerical Example • Transformed values = EigenvectorsT x DataT u1= u2= -0.96912 2.62911 -0.52988 1.09308 0.96278 1.26804 0.28680 -0.32067 -1.24869 -1.48470 0.43660 -0.83461 -4.73756 0.56874 -2.72931 4.40097 -0.75643 -3.22104 Which of the two is the principle component? Check magnitude of Eigenvalues 0 1.08737 9.593297 0 0.94234 4.24351 ORTHOGONAL GARCH F1 a1 X1 a2 X 2 X1 a1F1 2 F2 F2 b1 X1 b2 X 2 X 2 1F1 2 F2 VAR[ X 1 ] VAR[a1F1 2 F2 ] 12Var[ F1 ] 22Var[ F2 ] VAR[ X 2 ] VAR[ 1F1 2 F2 ] 12Var[ F1 ] 22Var[ F2 ] COV [ X 1 , X 2 ] COV [(a1F1 2 F2 )(1F1 2 F2 )] COV [a1F11F1 ] COV [a1F1 2 F2 ] COV [ 2 F2 1F1 ] COV [ 2 F2 2 F2 ] a11COV [ F1F1 ] a1 2COV [ F1F2 ] 2 1COV [ F2 F1 ] 2 2COV [ F2 F2 ] COV[F1,F2]=0 a11COV [ F1F1 ] 2 2COV [ F2 F2 ] a11VAR[ F1 ] 2 2VAR[ F2 ] FACTOR GARCH 1. Calculate eigenvalues and eigenvectors 2. Determine the number of eigenvectors 3. Calculate the factor scores and keep the equations 4. Estimate a GARCH models for the each factor scores. 5. Using factor score equations estimate Variances and Covariances