Financial Times Series. Lecture 12

Transcription

1 Financial Times Series Lecture 12

2 Multivariate Volatility Models Here our aim is to generalize the previously presented univariate volatility models to their multivariate counterparts We assume that returns evolve according to r t = μ t + a t where μ t = E r t F t 1 and a t = a 1t,, a kt are the innovations Furthermore, we assume that μ t evolves according to p q μ t = Υx t + Φ i r t i Θ i a t i i=1 i=1 where x t is an m-dimensional vector of exogenous variables with x 1t = 1 and Υ is a k m matrix

3 Multivariate Volatility Models We also define the k k matrix Σ t = Cov a t F t 1 Our main interest is the evolution of Σ t It is not obvious how to generalize a univariate model, say a GARCH, to this context, since we now also have to consider (time-dependent) correlations/covariances between assets

4 EWMA However, the generalization of an EWMA is straight forward; Given F t 1 = a 1,, a t 1 we have Σ t = t 1 1 λ 1 λ t 1 λj 1 a j a j j=1 For a t so large that λ t 1 0, we may write Σ t = 1 λ a t 1 a t 1 + λσ t 1

5 Estimation For á priori estimates of λ and Σ t, Σ t can be computed recursively using the above formula If we assume that a t = r t μ t follows a multivariate normal distribution with mean zero and covariace matrix Σ t we may use the log-likelihood function 1 2 T Σ t 1 2 T a t Σ t 1 a t t=1 t=1 which may be evaluated recursively, replacing Σ t by Σ t

6 Example HS & N225 We will use log returns in percentages for the Hang Tsen and N225 indices from to (See Tsay 3rd ed p.507) Note the crisis in the fall of 2008!

7 Example HS & N225 Univariate estimated GARCH(1,1) models are, where 1 denotes HS and 2 denotes N225; r 1t = a 1t, a 1t = σ 1t ε 1t and σ 2 1t = r 2 1t σ 2 1t 1 r 1t = a 1t, a 1t = σ 1t ε 1t σ 2 1t = r 2 1t σ 2 1t 1

8 Example HS & N225 HS N225

9 Example HS & N225 We note that for both the models the sums of the estimated α and β are close to one, suggesting an IGARCH or EWMA model The large β estimates are probably due to the financial crisis

10 Example HS & N225 Using λ = 0.86 as suggested by the univariate GARCH fits and the recursive formula and the unconditional covariance of the data as our á priori estimate, we get

11 Multivariate GARCH The are many different multivariate generalizations of the univariate GARCH model Bollerslev, Engle and Wooldridge (1988) propose the Diagonal Vectorization (DVEC) Model, defined by m s Σ t = A 0 + A i a t i a t i + B j Σ t j i=1 j=1 where A i and B j are symmetric matrices and is elementwise multiplication (Hadamard product in French )

12 DVEC(1,1) Since the matrices A 1 and B 1 are symmetric we consider only the lower triangular part for which the elements are σ 2 11,t = A 11,0 + A 11,1 a 2 1,t 1 + B 11,1 σ 2 11,t 1 σ 2 21,t = A 21,0 + A 21,1 a 1,t 1 a 2,t 1 + B 21,1 σ 2 21,t 1 σ 2 22,t = A 22,0 + A 22,1 a 2 2,t 1 + B 22,1 σ 2 22,t 1

13 Example Using the monthly returns of the Pfizer and Merck stocks, respectively

14 Example Using the ML routines of Ledoit et al (found at we get Parameter A 11,0 A 12,0 A 22,0 Estimate A 11, A 12, A 22, B 11, B 12, B 22,

15 Example Here are the estimated volatilities and correlation

16 Cons of the DVEC It may produce a non-positive definite covariance matrix It does not allow for dynamic dependence between volatility series

17 BEKK This is a model will give a non-negative definite covariance matrix and allows for dynamic dependence between the volatility series; m s Σ t = AA + A i a t i a t i A i + B j Σ t j B j i=1 j=1 where A is a lower triangular matrix and A i and B j are k k matrices

18 Estimation for Pfizer and Merck Using bekk.m of the MFE toolbox, available at we get the estimates Parameter Estimate p-value A 11, A 12, A 22, A 11, A 12, A 21, A 22, B 11, B 12, B 21, B 22,

19 BEKK for Pfizer and Merck

20 BEKK cons It should be noted that, due to the definition of the model the parameter values have no direct interpretation We may also note that the number of parameters to be estimated grows like k 2 m + s + k k + 1 /2

21 CCC MV-GARCH Bollerslev (1990) proposed a multivariate GARCH with constant conditional correlation(s) The model may be written Σ t = α 0 + α 1 a 2 t 1 + β 1 Σ t 1 In the two-dimensional case we have σ 2 11,t σ 2 22,t = α 10 α 20 + α 11 α 12 α 21 α 22 a 2 11,t 1 a 2 22,t 1 + β 11 β 12 β 21 β 22 σ 2 11,t 1 σ 2 22,t 1

22 CCC MV-GARCH Letting η t = a 2 t Σ t we may rewrite the model as a 2 t = α 0 + α 1 + β 1 a 2 t 1 β 1 η t It follows (just as in the univariate case) that the model can be considered as an ARMA(1,1) for a 2 t

23 Properties If all the eigenvalues of α 1 + β 1 are between zero and one then the above described ARMA(1,1) is weakly stationary For the weakly stationary process we have that the unconditional variance of a t is I α 1 β 1 1 α 0

24 DCC models In real life correlations between time series tend to be non-constant over time To take this into account we may use Dymanic Conditional Correlation (DCC) models One such model was proposed by Engle (2002) in which where J t = diag Q t and ρ t = J t 1 2 Q t J t 1 2 Q t = 1 θ 1 θ 2 Q + θ 1 ε t ε t + θ 2 Q t 1 for scalars θ 1 + θ 2 < 1, Q the unconditional covariance of the standardized innovations ε t

25 Example (dcc.m from MFE)

26 Residuals check To check the different models abilities to capture dependence and volatility of the assets we may create (observed) standardized residuals as ε t = Σ t 1 2 at We may apply autocorr or Ljung-Box to the standard residuals and their squares to check model adequacy

27 AC (squared on the bottom) Pfizer data Merck data

28 Pfizer BEKK residuals Merck

29 Residuals check for DCC Pfizer Merck

30 DCC vs. BEKK It is hard to tell from just the plots which is better The BIC for BEKK is The BIC for DCC is But these are computed under assumption of normality

31 DCC vs. BEKK (DCC upper row, Pfizer left column)

32 DCC vs. BEKK Descriptives and JB-tests for residuals Model/data Skewness Kurtosis JB p-value DCC/Pfizer DCC/Merck BEKK/Pfizer BEKK/Merck So, it is not safe to assume normality

33 Improvements Just as for the univariate GARCH the multivariate variants may be extended to include a skewness part t- To be able to capture more kurtosis one may assume that the white noise has Student s distribution There are many other multivariate volatility models, see e.g.