Econ 240C Lecture 16 1 May 29, 2007 I. Autoregressive Conditional Heteroskedastic(ARCH) Models Many economic time series are nonstationary in mean and variance. Other features that some economic time series exhibit are episodes of unusually high variance which may persist for awhile. One way of modeling these features is to model the variance as well as the series. In forecasting an economic time series, we have seen the importance of using conditional forecasts, for example, one period ahead forecasts conditional on all current and past knowledge. In the same way, if the variance is not constant, conditional forecasts of the variance can be important to the forecaster, especially in situations where risk is important. An example is portfolio analysis where forecasts of the mean return for the holding period as well as the variance for the holding period are critical to the decision maker. A. Autoregressive Error Variance Model Suppose, for example, that the time series is an AR(1): y(t) = a0 + a1 y(t-1) + e(t) where the error has mean zero and, ê2(t) = ê2(t-1) + ê2(t-2) + ... + WN(t). If the parameters etc. are zero then the expected estimated variance is constant or homoskedastic: Et-1[ê2(t)] = Engle Multiplicative ARCH Model Suppose the error process, e(t) has a multiplicative structure: Econ 240C Lecture 16 2 May 29, 2007 e(t) = WN(t)√[e2(t-1)] where the mean of the white noise series is zero and its variance is one, the white noise and lagged error, e(t-1), are independent, and is greater than zero and lies between zero and one. The mean of the error process, e(t), conditional or unconditional will be zero. The error process will not be serially correlated, and its unconditional variance will be constant. However, the conditional variance of the error process will be autoregressive of order one, i.e. ARCH(1). Since WN(t) and e(t-1) are independent, their joint density, f{WN(t), e(t-1)}, will be the product of the marginal densities: f{WN(t),e(t-1)} = g{WN(t)}h{e(t-1)}. 1. The unconditional expectation of e(t) is: E[e(t)] = ∫ WN(t)√[e2(t-1)]f{WN(t),e(t-1)} = ∫ WN(t)g{WN(t) ∫ √[e2(t-1)]h{e(t-1)} = E[WN(t)] E{ √[e2(t-1)]} = 0 since white noise has mean zero. 2. The covariance of e(t) and e(t-1) will be zero since E[e(t)e(t-1)] =E{WN(t)√[e2(t-1)]WN(t-1)√[e2(t2)]} and using independence of WN(t) and e(t-1), = E{WN(t)WN(t-1)} E{√[e2(t-1)]√[e2(t-2)]} =0 since E{WN(t)WN(t-1)} is zero. Econ 240C 3. Lecture 16 3 May 29, 2007 The unconditional variance of the error is: E[e(t)]2 = E{[WN(t)]2[e2(t-1)]} and by independence, E[e(t)]2 = E{[WN(t)]2 E[e2(t-1)]} = 1 [e2(t-1)]], and since for the unconditional variance it is true that: E[e(t)]2 = E[e(t-1)]2 , the unconditional variance is constant: E[e(t)]2 = The conditional mean of the error is, using independence of WN(t) and e(t-1) : Et-1[e(t)] = Et-1[WN(t)] Et-1[√[e2(t-1)] = 0 since the conditional expected value of white noise is zero. 5. The conditional variance of the error is Et-1[e(t)]2 = Et-1{[WN(t)]2[e2(t-1)]} and by independence: = Et-1[WN(t)]2 Et-1[e2(t-1)] = 1 [e2(t-1)], so the one period ahead forecast of the variance is autoregressive of order one and will be persistent. A shock to the system, e(t-1), will increase the one period ahead forecast of the variance and this larger variance will tend to persist since the variance is autoregressive. The closer is to one the longer the episode of high variance will tend to persist. II. Simulation of an ARCH(1) Time Series Econ 240C Lecture 16 4 May 29, 2007 Many economic time series are autoregressive of the first order. Suppose we have an autoregressive time series: y(t) = 0.9 y(t-1) + e(t) where the conditional variance of e(t) is also autoregressive e(t) = WN(t)√[1 + 0.7e2(t-1)] . A sample of 100 observations of white noise can be computer generated and used to generate 100 observations of the error, e(t), where e(0) is set to zero to initiate the error series, and y(0) is set to zero to initiate the ARCH(1) time series. Plots of the white noise, error, the autocorrelation and partial autocorrelation of the error squared, and the ARCH(1) series follow. __________________________________________________________ One Hundred Observations of Simulat ed Whit e Noise 4 3 2 1 0 -1 -2 -3 20 40 60 80 100 WN Econ 240C Lecture 16 5 May 29, 2007 One Hundred Simulated Observations of an Error Series with an Autoregressive Variance 15 10 5 0 -5 -10 -15 20 40 60 80 100 ERROR IDENT ERRORSQ SMPL range: 1 - 100 Number of observations: 100 ____________________________________________________________ Autocorrelations ac Partial Autocorrelations pac ____________________________________________________________ . ° ******** | . ° ******** | 1 0.629 | . ** . | 2 0.280 - | . ° . | 3 0.106 | . * . | 4 -0.018 - 0.629 . ° **** 0.190 . ° *. 0.027 . ° 0.096 . Econ 240C . * Lecture 16 . | . 6 May 29, 2007 ° *. | 5 -0.048 0.039 ____________________________________________________________ Q-Statistic (5 lags) Correlations 48.755 S.E. of 0.100 ____________________________________________________________ One Hundred Observations of a Simulated ARCH(1) Time Series: y(t) = 0.9 y(t-1) +error 15 10 5 0 -5 -10 -15 -20 20 40 60 80 100 Y _____________________________________________________ Note from the figures above that although we start with white noise, the error has a conditional variance which is heteroskedastic and autoregressive of the first order and the autoregressive time series itself has some episodes of high variance, especially relative to the first thirty observations. It is especially important to note that the error looks like white noise, as illustrated by the following Econ 240C Lecture 16 7 May 29, 2007 autocorrelation and partial autocorrelation functions. It is the error squared that is autoregressive. IDENT ERROR SMPL range: 1 - 100 Number of observations: 100 ____________________________________________________________ Autocorrelations ac Partial Autocorrelations pac ____________________________________________________________ .*** . | .*** . | 1 -0.204 - ° *. | . ° . | 2 . | . * . | 3 -0.087 - ° *. | . ° *. | 4 | . ° | 5 -0.002 0.204 . 0.040 - 0.002 . * 0.083 . 0.097 0.066 . ° . . 0.034 ____________________________________________________________ Q-Statistic (5 lags) Correlations 6.003 S.E. of 0.100 ____________________________________________________________ Econ 240C Lecture 16 8 May 29, 2007 III. Generalized Autoregressive Conditional Heteroskedastic (GARCH) Models Bollerslev generalized these models to allow for an ARMA structure in the error variance, capturing the benefits of parsimony in the number of parameters needed to fit the error structure. This was especially important because of the persistence in the error variance of time series such as the inflation rate, requiring a distributed lag of past error variances. In the Bollerslev model: e(t) = WN(t)√h(t) where h(t) = ie2(t-i) + ih(t-i) A. The Unconditional Mean Presuming independence between WN(t) and e(t-i), i≥1 then E[e(t)] = E[WN(t)} E[√h(t)] = 0 since white noise has mean function equal to zero. B. The Conditional Mean Once again assuming independence, Et-1[e(t)] = Et-1[WN(t)] Et-1[√h(t)] = 0 since the one period ahead forecast for white noise is zero. C. The Conditional Variance Once again assuming independence Et-1[e(t)]2 = Et-1[WN(t)]2 Et-1[h(t)] = Et-1[h(t)] = h(t) Econ 240C Lecture 16 9 May 29, 2007 = ie2(t-i) + ih(t-i) since the one period ahead forecast for the variance of white noise is one, and the period ahead forecast of h(t) depends on information known at time t-1. The empirical procedure for GARCH models is (1) to estimate the appropriate model for the economic time series, (2) identify the residuals to make sure they are white, and (3) identify the square of the residuals to see if there is a GARCH error structure. IV. Maximun Likelihood Estimation of ARCH Models It is possinle to use EViews/TSP to estimate an ARMA model and if the error is ARCH, estimate an AR model for the error variance. The estimated ARMA model can be used to forecast the time series and the estimated AR model of the error squared can be used to forecast the error variance for construction of confidence intervals. This procedure estimates the ARMA model separately from the AR model for the error square. Consequently, more efficient estimators can be obtained by using maximum likelihood methods to estimate these relationships jointly. A. Homoskedastic Variance Suppose (1) y(t) = b y(t-1) + e(t) (2) e(t) = WN(t)√a0 In this case, the errors are normal, the unconditional mean of e(t) is zero and the unconditional variance is a0. Since the errors are independent from one another, i.e. they are Econ 240C Lecture 16 10 May 29, 2007 orthogonal, the density function for the sample of errors, i.e. the unconditional likelihood function for the sample, L, is the product of the normal densities for the errors: L = 1,T (2)-1/2(a0)-1/2 exp[-1/2{(e(i) - 0)/√a0}2 L = (2)-T/2(a0)-T/2 exp[-(1/2)(1/a0)1,T[e(i)]2 and the logarithm of the likelihood function is: ln L= -(T/2)ln(2)-(T/2)ln(a0)-(1/2)(1/a0)1,T[y(i)-b y(i-1)]2 The derivative with respect to b is: ∂(ln L)/∂b = -(1/2)(1/a0)1,T[y(i)-b y(i-1)][-y(i-1)] = 0 and the estimator for b can be calculated by formula: b = {1,T[y(i)y(i-1)}/{1,T[y(i-1)y(i-1)} The derivative with respect to a0 is: ∂(ln L)/∂ a0 = -(T/2) (a0)-1 +(1/2)( (a0)-2 1,T[e(i)]2 = 0 and the estimator for a0 can be calculated by formula as well: â0 = {1,T[e(i)]2}/T These are the usual regression estimators. A. Heteroskedastic Variance Suppose (1) y(t) = b y(t-1) + e(t) (2) e(t) = WN(t)√h(t) (3) h(t) = [a0a1e2(t-1)] For initial values of the parameters, b*, a0*, a1* the errors, e(t), for the sample can be calculated from (1). The conditional errors, Et-1[e(t)] = WN(t)√ [a0* a1* e2(t-1)] are a known quantity, and for any given t, a constant, Econ 240C Lecture 16 11 May 29, 2007 [e2(t-1)], times white noise. Hence these conditional errors are normal and the conditional likelihood finction, CL, for the sample is: CL = 1,T (2)-1/2(h(t))-1/2 exp[-1/2{(e(i) - 0)/√h(t)}2 =1,T(2)-1/2(a0*+a1*e2(t-1))-1/2exp[-1/2{(e(i)-0)/√a0*+a1*e2(t1)}2 =(2)-T/2(a0*+a1*e2(t-1))-T/2 exp[-(1/2){a0*+a1*e2(t-1)}1,T[e(i)]2 and the logarithm of the conditional likelihood 1 function is: ln CL= -(T/2)ln(2)-(T/2)ln(a0*+a1*e2(t-1))(1/2){a0*+a1*e2(t-1)}-11,T[y(i)-b* y(i-1)]2 For the initial values of the parameters, b*, a0*, a1* we will obtain a value for the log conditional likelihood function and then will have to search this three dimensional parameter space for second iteration values, b** a0**, a1** that will raise the conditional likelihood, and continue to iterate until we find the maximum. There are algorithms that perform these calculations. One is available on Regression Analysis of Time Series(RATS), version 3.1 and later. One hundred observations of the simulated variable y with ARCH(1) errors was written to an ASCII file and copied to the RATS subdirectory of my Model 80. The 386 version RATS 3.10 for DOS was opened using the command RATS386. I had typed the ASCII batch file ARCH.COD, which was in the subdirectory as well. > RATS386 Econ 240C Lecture 16 <ALT> F 12 May 29, 2007 these keys struck simultaneously open the file menu source is selected from the file menu ARCH.COD is typed in file name box OK(button) runs the batch file The batch program and RATS output follow: ........................................................... allocate 0 100 open data simarc.dat data(org=obs) / y smpl 2 100 set u = 0.0 set v = 0.0 nonlin b0 b1 a0 a1 frml regresid = y(t)-b0-b1*y(t-1) frml archvar = a0+a1*u(t-1)**2 frml archlogl = -0.5*(log(v(t)=archvar(t))+(u(t)=regresid(t))**2/v(t)) eval b0 = 0.0; eval b1 = 0.9 eval a0 = 1; eval a1 = 0.7 maximize(method=bhhh,recursive,iterations=80) archlogl ....................................................................................................................................... NOTE: the smpl statement is critical since [e(t-1)]2 is required and will not be available if the sample starts with observation # 1. ** ALGORITHM DID NOT CONVERGE IN 80 STEPS ** ON THE LAST ITERATION THE CRITERION WAS 0.1249819E-01 NON-LINEAR MAXIMIZATION - ALGORITHM BHHH TOTAL OBSERVATIONS 99 SKIPPED/MISSING 0 USABLE OBSERVATIONS 99 DEGREES OF FREEDOM 95 FINAL FUNCTION VALUE -0.20000000E+51 NO. LABEL VAR LAG COEFFICIENT STAND. ERROR T-STATISTIC *** ******* *** *** ************ ************ ************ 1 B0 1 0 0.8221595E-02 0.8467668E-01 0.9709397E-01 2 B1 2 0 0.8997189 0.2401051E-01 37.47188 3 A0 3 0 1.029126 0.1932176 5.326254 4 A1 4 0 0.7332557 0.1714210 4.277515 ............................................................................................................................................... Note: the coefficients are close to the values used in the simulation. Good initial values are important, otherwise many iterations may be required. Econ 240C Lecture 16 13 May 29, 2007 V.ARCH-M Models The ARCH-M model relates return to risk. As in the capital asset model, the net return to an asset varies with an expected risk premium and an asset specific shock: y(t) = u(t) + e(t) where, taking conditional expectations: Et-1 y(t) = u(t) + 0 and the expected risk premium varies with the conditional variance of the shock: u(t) = h(t) and the shock to the asset is ARCH, h(t) = e2(t-1) + .. i.e. the error, e(t) has the properties: e(t) = WN(t) √h(t) Et-1[e(t)] = Et-1[WN(t)] Et-1[√h(t)] = 0 Et-1[e(t)]2 = Et-1[WN(t)]2 Et-1[√h(t)]2 = h(t)