Engle’s ARCH Model: Time-Varying Volatility Revisited in Finance “Statistical properties of the ARCH(1) Model: “Given an ARCH(1) model for ¥; and let's denote /, = {¥, Y;-1, ..., ¥%} denote the information set at time t¢ (conditioning set of random variables). +The ARCH(1) model (4) - (8) has the following seven statistical properties: 1) ECE 1) = 0,8(€,) = 0. 2) var(¥e | Ie) = ECE? he) = 02 3) var(¥,) = E(e?) = E(o?) = w/(1- a). +4) {¥,} isan uncorrelated process: cov(¥;,¥%-)) = E(€¢ée-j) = 0 for j > 0. 5) The distribution of ¥; conditional on /,.. is normal with mean jt and variance 02. 6) The unconditional (marginal) distribution of ¥, is not normal and kurt(¥,) > 3. 7) {¥2) and {¢?} have a covariance stationary AR (1) model representation. The persistence of the autocorrelations is measured by a.Engle’s ARCH Model: Time-Varying Volatility Revisited in Finance “Statistical properties of the ARCH(1) Model: 6) The unconditional (marginal) distribution of ¥, is not normal and kurt(¥,) = 3. “% Proof: With ¥; = + o¢2¢, then ¥; = + (w + yo? 4z? 4)/2z—. “Even though z, ~ iid N(0,1), ¥ is a complicated nonlinear function of z; and past values of z? and so ¥; cannot be a normally distributed random variable. Next, to see that kurt(¥,) 2 3 first note that kurt(¥,) = an (9) “Now by iterated expectations and the fact that E(z?) = 3 E(e) = E(E(osz# | e-1)) = E(oP E(z? | te-1)) = 3E(?) of =?) = (+a)? Next, usin; 8 = w? + 2waye?_, + are,Engle’s ARCH Model: Time-Varying Volatility Revisited in Finance tatistical properties of the ARCH(1) Model: 6) The unconditional (marginal) distribution of ¥, is not normal and kurt(¥;) > 3. 3w2(1+a4) G-an@=3e9) “Substituting the above expression for E(¢#) into equation (9) then gives = 307 (+a1) Gray)? kurt) = Graj(-sai) * oF nar > 3 Proof: Given E(e#) = 1-307 “Hence, if returns follow an ARCH(1) process with a, > 0 then kurt(¥,) > 3 which implies that the unconditional distribution of returns has fatter tails than the normal distribution. azTesting, Estimating and Forecasting-ARCH(1) Model Testing for ARCH Effects: * Given an ARCH(1) model for ¥; is of the form: oY, = ete, t =1,...,T7 qd) €¢ = 042, where z ~ iid N(0,1) (2) of =wtayez4,0>0,05a,<1 (3) «+There are many tests to examine the presence of ARCH effects. “Among these, Langrange multiplier (LM) test is often used to test for the presence of ARCH effects. *+To perform the LM test, we first estimate the mean equation, which can be a regression of the variable on a constant like equation (1) or may include other variables as well (like ¥,= + 01Y;-1 + €¢,t =1,...,T) “+ Then, we save the estimated residuals é, and obtain their squares é?. 4Testing, Estimating and Forecasting-ARCH(1) Model “Testing for ARCH Effects: oY = wte,t= 1.7 (1) €¢ = o,Z¢, where Z, ~ lid N(0,1) (2) to? =w+ me21,0>0,05 a <1(3) “> Langrange multiplier (LM) test to examine the presence of ARCH effects. “To test for the first-order ARCH , we regress é? on the squared residuals lagged 62. as follows: 6? = yo + V18i-1 + % (3) «»where y, is a random error term. the null and alternative hypotheses are Ho: y, =O Hy:y, #0Testing, Estimating and Forecasting-ARCH(1) Model + Testing for ARCH Effects: “+ Regress 6? on the squared residuals lagged é?_, as follows: +6? = yo + 167-1 + Ue (3), where v; is a random error term. +The null and alternative hypotheses are: Ho: 1 = 0 Hi:71 # 0 4If there are no ARCH effects, then y; = 0 and the fit of (3) will be poor, and the equation 8? will be low. “If there exists ARCH effects, we expect the magnitude of é? to depend on its lagged values, and the R? will be relatively high. “The LM test test statistic is (7 — q)R? where T is the sample size, q is the number of @2_, terms of the right-hand side of (3), and R? is distributed as x7,), where q is the order of lag, and T — q is the number of complete observations.Testing, Estimating and Forecasting-ARCH(1) Model > Testing for ARCH Effects: Regress 6? on the squared residuals lagged 62_, as follows: 26? = Yo + Vi6i-1 + Ve (g), where v; is a random error term. +The null and alternative hypotheses are: Ho: y, = 0 My:y1 #0 “If there are no ARCH effects, then y; = 0 and the fit of (3) will be poor, and the equation R? will be low. “rIf there exists ARCH effects, we expect the magnitude of é? to depend on its lagged values, and the R? will be relatively high. “The LM test test statistic is (T — q)R? where T is the sample size, q is the number of 62., terms of the right-hand side of (3), and R? is distributed as x{,), where q is the order of lag, and 7 —q is the number of complete observations. in this case, q = 1. if (T ~ q)R* = xZ-q)r then we reject the Ho that y; = 0 and conclude that ARCH effects are present. aTesting, Estimating and Forecasting-ARCH(1) Model > Example: Testing for ARCH Effects using S&P500 Daily Return Series (RetSP500) Thus, to perform the test for ARCH(1) effects, we must. 1) Estimate a mean equation: yz = fo + €-, where y, is the S&P500 daily return series. 2) Retrieve the estimated residuals é? 3) Estimate é7 = yo + yiéi-1 + ve (3), where vy is a random error term. ##e888888808##8# ARCH heteroscedasticity test for residuals using R manually + library(aynim) © Step u: Estimate mean equation: y; = Bo + ¢r > RetSP500.mean <- dynlm(RetSP500 ~ 1, data = RetSP500) ‘© Stop 2: Retrieve the residuals from the former model and square them ehatsq <- ts(resid(RetSPs00.mean)*2) Stop 42 regress squared residuals on one-lagged squared residualsTesting, Estimating and Forecasting-ARCH(1) Model + Example: Testing for ARCH Effects using S&P500 Daily Return Series (RetSP500) Seeeeeeeaeeaaneee® ARCH heteroscedasticity test for residuals using FinTS package R package “We can check for ARCH effects directly by using the ArchTest() function. *library(FinTS) **ARCH LM-test; Null hypothesis: no ARCH effects “ data: RetSP500 “Chi-squared = 628.75, df = 1, p-value < 2.2e-16 azTesting, Estimating and Forecasting-ARCH(1) Model +sARCH Model Estimates: ARCH models are estimated by the maximum likelihood method. & We can estimate an ARCH(1) model using the garchFit() function from the fGarch package in R. “Specifically, we need to estimate the variance given in equation: 02 =w+t ayez4,w > 0,05 a, < 1(3). > ######4#4 Example: ARCH Model Estimation using R “> ####3B. ARCH(1) model with normal innovations “m4 <- garchFit(RetSP500 ~ arma(o, 0) + garch(1, 0), data = RetSP500, * trace = F) ‘summary(m4) a8 005a,<1 (3) “»The ARCH (p) model extends the autocorrelation structure of ¥? and é? in the ARCH(1) model in equation (3) to that of an AR (p) process by adding p lags of e? to the dynamic equation for o? : of = w + ayez_1 + a2 E72 ++ + pep (4)