jonal Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation Robert F. Bugle Econometrica, Volume 50, Issue 4 (Tul., 1982), 987-1008, Stable URL: Lhtp:flinks,jstor-org/sici?sici~0012-9682%28 198207%29S0%IA4%ICI87%3 AACHWEO%3E2,0,CO®IB2.3 ‘Your use of the ISTOR archive indicates your acceptance of ISTOR’s Terms and Conditions of Use, available at up: srw jstor org/aboutterms.html. ISTOR's Terms and Conditions of Use provides, in part, at unless you have obtained prior permission, you may aot download an entite issue of a journal or multiple copies of articles, and ‘you may use content in the ISTOR archive only for your personal, non-commercial use Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the sereen or pfinted page of such transmission. Econometrica is published by The Bconomene Society. Please contact the publisher for further permissions regarding the use of chis work, Publisher contact information may be obtained at ftps jor. on@ounalsfeconoso- el Economerrica ©1982 The Econometric Society ISTOR and the JSTOR logo are trademarks of JSTOR, and are Registered in the U.S, Patent and Trademark Otic. For more information on ISTOR contact jstor-nfo@umnich edu, ©2002 JSTOR hup:thrww stor orgy Sat Sep 28 18:27:24 2002Beanomatriee, Vl. 0, Ne. 4 (uly, 1982) AUTOREGRESSIVE CONDITIONAL HETEROSCEDASTICITY. WITH ESTIMATES OF THE VARIANCE OF UNITED KINGDOM INFLATION! By Roser F. ENOLE Traditions) econometric models assume a convent one-petiad forecast varianes. Ta seneraize tis (plausible assumption, 9 ee elas of siochasleproceses called autore- {retve condlional hefroseecestic (ARCH) processes are introduced in this paper. Thee fre mean zero, eislly uecorrdaied proceses with aonconstant varanees canditoral oa the par, but constant unconditional variances For such pracescs, the Feat past Bice information about te one-pried forecast variance ‘A reazesion model is then introduced with disuurbancesfllowing an ARCH proces. Maximus lielvood estimators are described and a siple scoring Herston formulated Gccinaty lest scares maitaina its opuraty. properties vt hs setup, Dut MANS likelocd is more efficient. Tae elauveeffiieney i calculated and ean be infinite, Te st Whether he disturbances follow an ARCH proces, the Lagrange muller procedure nppayed. The testis based simply an the autoenceation of the squaced OLS resis "This rode ie ase to esine the means and vanances of inflaton sn the OK, The ARCH effect (ound 1 be significant nd the esrpated variances merease substartally during the chaoui seventies, 1, INTRODUCTION IF & RANDOM VARIABLE py, is drawn from the conditional density function ‘fOr, the forecast of today’s value based upon the past information, under standard assumptions. is simply (|), which depends upon the value of the conditioning variable y,_. The variance of this one-period forecast is given by Viyrly,)- Such an expression recognizes that the conditional forecast variance depends upon past information and may therefore be random variable. For conventional econometrie models, however, the conditional variance does nat depend upon y,—.. This paper will propose a class of models where the variance does depend upon the past and will argue for their usefulness in economics. Estimation methods, tests for the presence of such models, and an empirical ‘example will be presented, Consider initially the first-order autoregression Bea Date where ¢ is white noise with ¥(¢)= 02. The conditional mean of y, is yy, while the unconditional mean is zero. Clearly, the vast improvement in forecasts duc to time-series models stems from the use of the conditional mean. The conditional ‘This paper was wien while dhe author was visting the Eordon School of Economies. He nefits greatly fom many auvalting oniverations with Devid Hendry and help suezstions ‘by Denia Sargan nd Andfew Harvey. Special thanks ace due Freak Seba who carted ut the ampitations Further ansghtfal comments are die to Cive Granger, Tom Rothenberg, Eamond Malievaud. Iesn-Fearcas Rickard, Wayne Foller, and two cnouymols refecees, The cescrch Was supported by NSF SOC 760947 acd ‘The latemational Cente for Ezoncmies and Related Disciplines, Ail errors remain dhe aos resporabliy 97988 ROBERT F. ENGLE variance of y, is a? while the unconditional variance is a?/1— y2, For real processes one might expect better forecast intervals if additional information from the past were allowed to affect the forecast variance; a more general class of models seems desirable ‘The standard approach of heteroscedastcity is to introduce an exogenous variable x, which predicts the variance. With a known zeto mean, the model might be Jat where again V(«) = 0%. The variance of y, is simply o%x}_, and, therefore, the forecast interval depends upon the evoiulion of an exogenous variable. This standard solution to the problem seems unsatisfactory, as it requires a specifica- tion of the causes of the changing variance, rather than recognizing that both conditional means and variances may jointly evolve over time. Pethaps because of this difficulty, heteroscedasticity corrections are rarely considered in time- seies dala. ‘A model which allows the conditional variance to depend on the past realiza~ tion of the series is the bilinear model described by Granger and Andersen [13]. ‘A simple ease is Heo ‘The conditional variance is now a’y?_,. However, the unconditional variance is either zero of infinity, which makes this aa unattractive formulation, although slight generalizations avoid this problem. ‘A preferable model is wren, y= ay a yoy. with ¥(&)= L This is an example of what will be called an autoregressive conditional heteroscedasticity (ARCH) model. [tis not exactly a bilinear model, ‘but is very close to one. Adding the assumption of normality, it ean be more directly expressed in terms of q,, the information set available at time ¢. Using conditional densities, wy Felt N OA), Q — haagtayy The variance function ean be expressed more generally as O) BBO nea Pears deep) where p is the order of the ARCH process and a is a vector of unknown parameters,HETEROSCEDASTICITY 989 ‘The ARCH regression model is obtained by assuming that the mean of y, is siven as x,f, a linear combination of lagged endogenous and exogenous variabies included in the information set g,_, with fa vector of unknown parameters, Formally, Jd MOP oh, BAG eo trer Gapet nn xB. ‘The variance function can be further generalized to include current and Jagged x's as these also enter the information sel. The A function then becomes eC poses Mapa) or simply B= ROA. This generalization will not be treated in this paper, but represents a simple extension of the results, In particular, if the 4 function factors into hows OheCRy os Beph the two types of heteroscedasticity can be dealt with sequentially by first correcting for the x component and then ‘iting the ARCH model on the transformed data. ‘The ARCH regression model in (4) has a variety of characteristics which make it attractive for econometric applications. Fconometric forecasters have found that their ability to predict the {uture varies from one period to another. McNees (17, p. 52] suggests that, “the inherent uncertainty or randomness associated with different forecast periods seems to vary widely over time.” He aiso documents thas, “large and small errors tend to cluster together (in contiguous time peri ods).” This analysis immediately suggests the usefulness of the ARCH model where the underlying forecast variance may charge over time and is predicted by ppast forecast errors, The results presented by McNecs also show some serial ‘correlation during the episodes of large variance. ‘A second example is found in monetary theory and the theory of finance. By the simplest assumptions, portfolios of finaneial assets are held as functions of the expected means and variances of the rates of return, Any shifts in asset demand must be associated with changes in expected means and variances of the rates of retum. If the mean is assumed to follow a standard regression of time-series model, the variance is immediately constrained to be constant over time. The use of an exogenous variabie to explain changes in variance is usually ot appropriate.990 ROBERT F. ENGLE A third interpretation is (hat the ARCH regression model is an approximation to a more complex repression which has non-ARCH disturbances. The ARCH specification might then be picking up the effect of variables omitted from the estimated model. The existence of an ARCH effect would be interpreted as evidence of misspecification, either by omitted variables or through structural change. If this is the case, ARCH may be a better approximation to reality than making standard assumptions about the disturbances, but trying to find the ‘omitted variable or determine the nature of the structural change would be even better. Empirical work using time-series data frequently adopts ad hoc methods 10 measure (and allow) shifts in the variance over time, For example, Kiein [15] obtains estimates of variance by constructing the five-period moving variance about the ten-periad moving mean of annual inflation rates, Others, such as Khan (14], resort to the notion of “variability” rather than variance, and use the absolute value of the first difference of the inflation rate, Engle [10] compares these with the ARCH estimates for U.S. data. 2. THE LIKELIHOOD FUNCTION Suppose y, is generated by an ARCH process described in equations (1) and. (@). The properties of this process can easily be determined by repeated applica- tion of the relation £x = E(Ex|¥)). The mean of, is 7ero and ll auton covariances are zero, The unconditional variance is given by 6? = Ey? = Eh, For many functions A and values of «, the variance is independent of ¢, Under such conditions », is covariance stationary; a set of sufficient conditions for this is ferived below. Although the process defined by (1) and (3) has all observations conditionally ormally distributed, the vector of y snot jointly normally distributed. The joint density is the product of all the conditional densities and, therefore, the loz likelihood is the sum of the conditional normal log likelihoods corresponding to (1) and (3). Let / be the average log likelihood and /, be the log likelihood of the 1th observation and T the sample size. Then iz Fle oy Gm plogh, 5 ¥2/hy apart from some constants in the likelihood, To estimate the unknown parameters a, this likelihood function can be maximized. The ficst-ocder conditions are m1 aa 3h, Iw oyHETEROSCEDASTICITY 998 and the Hessian is ee cuaectal Bada ~~ 943 a Ba"\ A} "| "YBa? | 2K Ga ‘The conditional expectation of the second term, given is zero, and of the last factor in the frst, is just one. Hence, the information matrix, which is simply the negative expectation of the Hessian averaged over all observations, becomes shn| 4 th Oh = iF { at Fat which is consistently estimated by 1. Bh, ah, ay 4, ~ $3 ab | OS, If the A function is pth order linear (in the squares), so that it can be written as GN ea taht tap then the information matrix and gradient have particularly simple form. Let 2 =U yhges ses ¥ag) ated a (p50), +4) 80 that (L1) can be rewntten as (yh, ‘The gradient then becomes simply ay at and the estimate of the information matrix 04) 8g = ap Dem) 3. DISTRIBUTION OF THE FIRST-ORDER LINEAR ARCH PROCESS ‘The simplest and often very useful ARCH mode! isthe first-order linear mode! sziven by (1) and (2). A large observation for y will lead to a large variance for the next period's distribution, but the memory is coafined to one period. If a, = 0, of course y will be Gaussian white noise and if itis a positive number, successive cebsecvations will be dependent through higher-order moments. As shown below, if ais oo large, the variance of the process will be infinite To determine the conditions for the process (0 be stationary and to find the marginal distribution of the y's, & recursive argument is required. The odd992 ROBERT F ENGLE moments are immediately seen to be zero by symmetry and the even moments are computed using the following theorem, In all cases it is assumed that the process begins indefinitely far in the past with 2r finite initial moments. TreoreM |: For integer 1, the 2rth moment of a firsteorder linear ARCH process with ap > 0, a 20, exises if, and only if, a fha-v 0, ay + % 20, is covariance stationary if, and ony if, the astociated characteristic equation hhas all raots outside the unit eirele. The stationary variance is given by E(y2) = ag/ O34 Paoor: See Appendin. ‘Although the pth-order linear model is a convenient specication, itis kel that other formulations of the variance model may be more appropriate for particular applications. Two simple alternatives are the exponential and absolute value forms: (16) = exp(ag + yin), CD b= a+ yeh ‘These provide an interesting contrast, The exponential form has the advantage that the vaciance is positive forall values of alpha, bu itis not difficult to show that data generated from such a model have infinite variance for any value of @, #0, The implications of this deserve further study. The absolute value form requires both parameters (o be positive, but ean be shown to have finite variance for any parameter values In order (o find estimation results which are more general than the linear model, general conditions on the variance model will be formulated and shown to be implied for the linear process. Let & be ap XI random vector drawn from the sample space Z, which has elements £; =(&_1,.... 4) For any &,, let g be identical, except that the mith, element has been multiplied’ by ~1, where m fies between. 1 and p. Dermimon: The ARCH process defined by (1) and 3) is symmetric if @) AE) ~ AEE) for any m and €.<2, (b) h(E )/4a,= BREE )/9a, for any mi and eZ, Oo DACE )/ 28,4 = ~ SAG 5/86, for amy om and g All the fudctions described have been. symmetric, This condition is the main distinction between mean and variance models, Another characterization of general ARCH models is in terms of regutarity conditions, Dertwiriox: The ARCH model defined by (1) and (3) is regular if (@) ——minh(Zy>4 for some > O and GeF, ) E([BA(E, )/ Bq FBACE )/0E, | [rag 1)—_ ORistS For all i, om,994 ROBERT F, ENGLE The first portion of the definition is very important and casy to check, as it requires the variance always ta he pasitive. This eliminates, for example, the logelog autoregression. The second portion is difficult to check in some cases, yet should generally be (rue if the process is stationary with bounded derivatives, since conditional expectations are finite if unconditional ones are. Condition (b) ig 4 sufficient condition far the existence of some expectations of the Hessian used in ‘Theorem 4, Presumably weaker conditions could be found. Tueonen 3: The poh-order linear ARCH model satisfies the regularity condi fons, if ag > Gand a, 4,20. Proor: See Appendix. In the estimation portion of the paper, a very substantia! simplification results if the ARCH process is symmetric and regular. 5. ARCH REGRESSION MODELS If the ARCH random variables discussed thus far have a non-zero mean, which can be expressed as 2 linear combination of exogenous and lagged dependent variables, then 2 regression framework is appropriate, and the model ‘ean be written as in (4) or (5). An alternative interpretation for the model is that the disturbances in a linear regression follow an ARCH process. Th the pth-order linear ease, the specification and likelihood are given by JilB~ NL Bsh) Kya agt ays + ae, (8) gy xB iz Ta = = Hlogh, —36/h, where x, may include lagged dependent and exogenous variables and an irtele- vant constant has been omitted fram the tikelihood. This likelihood function can be maximized with respect to the unknown parameters a and f, Atractive methods for computing such an estimate and its properties are discussed below. Under the assumptions in (18), the ordinary least squares estimator of f! is still consistent as x and ¢ are uncarcelated through the definition of the regression as a conditional expectation, If the x's can be treated as fixed constants then the Teast squares standard errors will be correct; however, if there are lagged dependent variables in x,, the standard errors as conventionally computed will not be consistent, since the squares of the disturbances will be correlated withHETEROSCEDASTICITY 995, squares of the 2's. This is an extension of White's [18] argument on heterosce- dasticity and it suggests that using his alternative form for the covariance matrix ‘would give a consistent estimate of the least-squares standard errors. If the regressors include no lagged dependent variables and the process is stationary, then letting y and x be the Tx 1 and TX K vector and matrix of dependent and independent variables, respectively, E(y|x) = x8, as) Vat(y| x)= 0% and the Gauss-Markov assumptions are statisfied, Ordinary least squares is the best linear unbiased estimator fos the model in (18) and the variance estimates are unbiased and consistent. However, maximum likelihood is different and consequently asymptotically superior; ordinary least squares does not achieve the Cramer-Rao bound, The maximum-likelisoad estimator is nonlinear and is more efficient than OLS by an amount calculated in Section 6. ‘The maximum likelihood estimator is found by solving the first order condi- sions. The derivative with respect to fis at, ak aL the _ | eB 2h, alk} ‘The first term is the familiar first-order condition for 2n exogenous heterosee- dastic correction; the second term results because fis also 4 function of the 2's, as in Amemiya {1}. Substituting the linear variance function gives oy en ‘geen which can be rewritten approximately by collecting terms in x and ¢ as om Bat Beef" - Som Hees ne)| eh Dane The Hessian is996 ROBERT F. ENGLE, Taking conditional expectations of the Hessian, the last two terms vanish Dhecause his entirely a function of the past. Similarly, /h, becomes one, since it is the only current value in the second term. Notice that these results hold regardless of whether x, includes Sagged-dependent varigbles, The information matrix is the average over all ¢ of the expected value of the conditional expectation and is therfore, given by By gathering tetms in x/x,, (24) can be rewritten, except for end effects, as salts re Soe aly BEN Ta a similar fashion, the off-diagonal blocks of the information matrix can be expressed as a1 1 34, a, 26) fag pees Bap The important result to be shown in Theorem 4 below is that this off-diagonal block is zero, The implications are far-reaching in that estimation of a and f can be undertaken separately without asymptotic loss of efficiency and their vari- ances can be calculated separately. Tuwonem 4: If an ARCH regression model is symmetric and regular, then Sag = Proor: See Appendix. 6, ESTIMATION OF THE ARCH REGRESSION MODEL, Because of the black diagonality of the information matrix, the estimation of a and fi can be considered separately without loss of asymptotic efficiency.HETEROSCEDASTICITY 997 Furthermore, either can be estimated with full efficiency based only on a consistent estimate of the other. See, for example, Cox ané Hinkley (6, p. 308). The procedure recommended here is to initially estimate A by ordinary least squares, and obtain the residuals. From these residuats, an efficient estimate of a can be constructed, and based upon these & estimates, efficent estimates of fare found. The iterations ace calculated using the scocing algorithm. Each step for a parameter vector @ produces estimates 4°"! based on 9" according (0 gat rae where J and 4///d@ are evaluated at $/. The advantage of this algorithm is partly that it requires only Fist derivatives of the ikelihood funetion ia this case and partly that it uses the statistical properties of the problem to tailor the algorithm to this application. For the plh-order linear model, the scoring step for a can be rewritten by substituting (J2), (12), and (14) into (27) and interpreting y} as the residuals. ‘The iteration ie simply 08) ait mal (FEES an 6! =o! 4 [84] where mw (Lelie eh )dals FH (F000. 5) Fa (G hihi fhe) In these expressions, ¢, is the residual from iteration i) is the estimated conditional variance, and ai the estimate of the vector of unkown parameters from iteration i, Each step is, therefore, easily constructed from a least-squares regression on transformed variables. The variance-covariance matrix of the parameters is consistently estimated by the inverse of the estimate of the information matrix divided by 7, which is simply 2(2'2)"', This differs slightly from 422)" computed by the auxiligry regression. Asymptotically, 6° = 2, if the distributional assumptions are corsect, but it is not clear which formulation is better in practice, ‘The parameters in a must satisfy some nonnegativity conditions and some stationarity conditions, These could be imposed via penalty functions or the parameters could be estimated and ebecked for conformity. The latter approach is used here, although a perhaps useful reformulation of tae madel might employ squares to impose the nonnegativity constraints directly: O09 hazed bald +--+ ae998, ROBERT F. ENGLE ‘Convergence for such an iteration can be formulated in many ways. Following Besley [3], 2 simple criterion is the gradient eround the inverse Hessian, For & parameter vector, @, this is x ot, y aL ag Bpae" | Bg” Using @ as the convergence criterion is attractive, as it provides a natural normalization and as iti interpretable as the remainder term in a Taylor-series expansion about the estimated maximum. In any case, substcuting the gradient and estimated information matrix in (30), @ = R? of the auxiliary regression, For 2 given estimate of a, & scoring step can be computed to improve the estimate of beta. The scoring algorithm for f is 30) Gl B= B+ [San] 5 Defining £,= x,r, and 2 = @s,/r, with @ and 2 as the corresponding matrix and veetor, (31) ean be rewritten using (22) and (24) and for the estimate of «, on the ith iteration, as BD Aa Ae (eRe ‘Thus, an ordinary least-squares program can again perform the scoring iteration, and (¥.£)"' from this calculation will be the final variance-covariance matrix of the maximum likelihood estimates of A. Under the conditions of Ccowder's (7] theorem for martingales, it can be established that the maximum likelihood estimators and fl are asymptotically normally distributed with limiting distrisution T(&- a) 3N(0,52.), es ¥T(&~ a) > N(O,821), {T(B-B)3.N(0.558). 17. GAINS IN EFFICIENCY FROM MAXIMUM LIKELIHOOD ESTIMATION ‘The gain in efficiency from using the maximunt-likelihood estimation rather than OLS has been asserted above. In this section, the gains are celeulated for a special case. Consider the linear stationary ARCH model with p = | and all x, ‘exogenous. This is the case where the Gauss-Markov theorem applies and OLS has a variance matrix o%(x'x)"' = BE, x/x)"'. The stationary variance is 8 = ay /(1- @). ‘The information matrix for this case becorses, from (25), ef Smen(te' + Baal)HETEROSCEDASTICITY 999 With x exogenous, the expectation is only necessary over the scale factor. Because the disturbance process is stationary, the variance-covariance matrix 18 proportional to that for OLS and the relative efficiency depends only upon the scale factors. The relative efficieney of MLE to OLS is, therefore, Rm Bm! + 2Cai/Ah Now substitute y= ay + aie, 07 ag/I— my, and y= a/} ~ ay. Recogniz ing that «7, and ¢? have the same density, define for each uae a)/ag ‘The expression for the relative efficiency becomes ity a we on) aar( th) saree, Te Te ey where thas variance one and mean zero, From Jensen’s inequality, tie expected value of a reciprocal exceeds the reciprocal of the expected value and, therefore, the first term is greater than unity. The second is positive, s0 there is a gain in efficiency whenever +70. Eu"? is infinite because u° is conditionally chi squared with one degree of freedom. Thus, the limit of the relative efficiency goes to infinity with y: Jim, R20 For a, close to unity, the gaia in efficiency from using a maximum likelinood estimator may be very large, 8. TESTING FOR ARCH DISTURBANCES, In the linear regression model, with or without lagged-dependent variables, OLS is the appropriate procedure if the disturbances are not conditionally heteroscedastic, Because the ARCH model requires iterative procedures, i¢ may be desirable to test whether itis appropriate before going to the effort to estimate it The Lagrange multiplier test procedure is ideal for this as in many similar cases. Sec, for example, Brousch and Pagan [4, 5], Godfrey (12), and Engle (9). Under the null hypothesis, «, = a, +++ = a = 0. The test ig based upon the score under the ull and the information matrix under the null. Consider the ARCH model with f= A(2,a), where f is some differentiable function which, therefore, includes both the linear and exponential cases as well as lots of others and z,=(l.eh,....,¢2.,) where ¢, are the ordinary least squares residuals Under the null, h, is 2 constant denoted °. Writing 3h, /da = W’s,, where A is1900 ROBERT F ENGLE the scalar derivative of A, the score and information can be written as aL aa and, therefore, the LM test statistic can be consistently estimated by ya fee's) where 2 (hos 2ph fis the column vector of “This is the form used by Breusch and Pagan [4] and Godfrey [12] for testing for heterascedasticity. As they point out, ail reference to the & function has dis- appeared and, thus, the test is the same for any k which is a function only of 2, In this problem, the expectation required in the information miateix could be evaluated quite simply under the null; this could have superior finite sample performance. A second simplification, which is appropriate for this model as well Js the heteroscedasticity model, is to note that plim f”/?/T = 2 because normal- ity has already been assumed. Thus, an asymptotically equivalent statistic would be 06) Ea TY 2(2'2y P/O? = TR? where R? is the squared multiple correlation between /? and 2. Since adding a constant and multiplying by a scalar will not change the R? of a regression, this fs ako the R? of the regression of e? on an intercept and p lagged values of <>. ‘The statistic will be asymptotically distributed as chi square with p degrees of freedom when the null hypothesis is true, “The test procedure is to run the OLS regression and save the residuals. Regress the squared residuals on a constant and p lags and test TR? as a x3. This will be an asympioticaly locally most powerful test, a characterization it shares with Tikelihood ratio and Wald tests, The same test has been proposed by Granger and Anderson [13] to test for higher moments in bilinear time series. 9, ESTIMATION OF THE VARIANCE OF INFLATION Economic theory frequently suggests that economic agents respond not only £9 the mean, but also to higher moments of economic random variables. In financial theory, the variance as well as the mean of the rate of return are determinants of portfolio decisions. In macroeconomics, Lucas (16) for example,HETEROSCEDASTICITY 1001 argues that the variance of inflation is 2 determinant of the response to various shocks. Furthermore, the variance of inflation may be of independent interest as itis the unanticipated component which is responsible for the bulk of the welfare loss due to inflation. Friedman [11] also argues that, as high inflation will generally be associated with high varlability of inflation, the statistical relation- ship between inflation and unemployment should have a pasitive slope, not & negative one as in the traditional Phillips curve. Measuring the variance of inflation over time has presented probiems to various researchers, Khan [4] has used the absolute value of the first difference of inflation wale Klein (15] has used a moving variance around a moving mean. Each of these approaches makes very simple assumptions about the mean of the distribution, which are inconsistent with conventional econometric approaches. ‘The ARCH method allows a conventional regression specification for the mean funetion, with a variance which is permitted to change stochastically over the sample period. For a comparison of several measures for U.S. data, see Engle [10 A conventional price equation was estimated using British data from 1958-[1 through 197-11. It was assumed that price inflation followed wage increases; thus the mosel is a restricted transfer function Letting p be the first difference of the log of the quarterly consumer price index and w be the log of the quarterly index of manual wage rates, the mode! chosen after some experimentation was BT P= Bip. + Bibs Bsp_s+ Bp m)_\+ Bs. ‘The model has typical seasonal behavior with the firs, Fourth, and fifth lags of the first difference. The lagged value of the real wage is the error correction mechanism of Davidson, et al. (8}, which restricts the lag weights to give a constant real wage in the long run. As this is a reduced form, the current wage tate cannot enter. ‘The Teast squares estimates of this model are given in Table I. The fit is quite ‘good, with less than I per cent standard error of forecast, and all ¢ statistics ‘greater than 3. Notice that _, and j.. have equal and opposite signs, suggesting that ic isthe acceleration of inflation one year ago which explains much of the short-run behavior in prices. TABLE OnpIvany Least Souanes (6) ie Smt HO Goel. 0534 G40e —0una assy ommsT Seb. 010 110 ltd onia6 —oausT2 rsa 395° 37388 ga ORNS ORL IRS RETAsoo2 ROBERT F, ENGLE, To establish the reliability of the model by conventional criteria, it was tested for setial correlation and for coefficient restrictions. Godfrey's [12] Lagrange multiplier test, for serial correlation up to sixth order, yields a chi-squared statistic with 6 degrees of freedom of 4.53, which is not significant, and the square of Durbin's fis 0.57. Only the 9th autocorrelation of the least squares residuals exceeds two asymptotic standard errors and, thus, the hypothesis of white noise disturbances can be accepted. The model was compared with an unrestricted regression, including all lagged p and w from one quarter through six. The asymptotic F statistic was 2.04, which is not significant et the 5 per cent level, When (37) was tested for the exclusion of through wg. the statistic ‘was 2.4, which is barely significant at the 5 per cent but not the 2.5 per cent level. The oaly variabie which enters significantly in either of these regressions is w_g and if seems unatiractive (o include this alone. ‘The Lagrange multiplic: test fora first-order linear ARCH effect for the model in GT) was not significant. However, testing for a fourth-order linear ARCH process, the chi-squared statistic with 4 degrees of freedom was 15.2, which is highly significant. Assuming that agents discount past residuals, a linearly declining set of weights was formulated to give the model C8) bya t ade, +034 +022 54 01e2.) which is used in the balance of the paper. A two-parameter variance function ‘was chosen because it was suspected that the nonnegativity and stationarity constraints on the a's would be hard to satisfy in an unrestricted model. The chicsquared test for a, = 0 in (38) was 6.1, which has one degree of freedom. One step of the scoring algorithm was employed to estimate model (37) and (38), The scoring step on a was performed first and then, using the new efficient 4, the algorithm obtains in one step, efticient estimates of fl. These are given in ‘Table TL The procedure was also iterated to convergence by doing three steps om 4, followed by three steps on f, followed by three more steps on a, and so forth. ‘Convergence, within 0.1 per cent of the final value, occurred after two sets of a and ff steps. These results are given in Table IT ‘The maximum likelihood estimates differ from the least squares effects primar- ily im decreasing the sizes of the short-run dynamic coefficients and increasing TABLET Maxmaen Lieeuttoon EsrnwaTes oF ARCH Monet. (36) (37), ‘One-Stem SCORING ESTs" Goel 0310210 0ai 00097 oom ose See. 010 ode 108 OmliTonawse 00 fsa 190° 288 30g S98 Ga La wy Aen eae pea) nah) wee Pig) OX erm ee aeHETEROSCEDASTICITY 1003 TABLE I Masavtnt Ligetatoon Eetivares oF ARCH Manes (26) (47) Trexare Estaaros* weir a é ° Geo Geil. 0162204 oazs O00?” Oss a 988 Seer 0108 00892 O87 IS 0K] $0298 Sat SO 286 329 si) 687 Te 30 Depend veep — ou) keh abe? oy Keeton sehr UR east natal vane as Snipe pees BSE TP the coefficient on the long run, as incorporated in the error correction mecha nism, The acceleration term is not $0 clearly implied as in the least squares estimates. These seem reasonable results, since much of the inflationary dynam- ics are estimated by @ period of very severe inflation in the middie seventies This, however, is also the period of the largest forecast errors and, henee, the ‘maximum likelihood estimator will discount these observations. By the end of the sample period, inflationary levels were rather modest and one might expect that the maximum likelihood estimates would provide a better forecasting equation The standard errors for ordinary least squares are generally greater than for maximum likelihood. The least squares standard errors are 45 per cent to 25 per cent greater, with one exception where the standard error actually falls by 5 per cent 10 7 per eent, As mentioned earlier, however, the least squares estimates are biased when there are lagged dependent variables, The Wald test for a = 0 is also significant “The final estimates of A, are the one-step-ahead forecast variances. For the ‘one-siep scoring estimator, these vary from 23 x 10°® to 481 x 10-6, That is, the forecast standard deviation ranges from 0.5 per cent to 22 per cent, which is more than a factor of 4. The average of the 4, since 1974, is 230% 10-6, as compared with 42 x 10-* during the last four years of the sixties. Thus, the standard deviation of inflation increased from 06 per cent to 1.5 per cent over a few years, as the economy moved from the rather predictable sixties into the chaotic. seventies In order to determine whether the confidence intervals arising from the ARCH. model were superior to the least squares model, the outliers were exainined. The ‘expected number of residuals exceeding two (conditional) standard deviations is 3.5. For ordinary least squares, there were $ while ARCH produced 3. For least squares these occurred in "74-1, "75-1, "75.11, "75-1V, and "76-Ll; they all occur within three years of each other and, in fact, tree of them are in the same year. For the ARCH reodel, they are much more spread out and osly one of tte least squares points remains an outlier, although the others are still large. Examining. the observations exceeding one standard deviation shows similar effects. In the seventies, there were 13 OLS and 12 ARCH residuals outside one sigma, which are both above the expected value of 9, tn the sixties, there were 6 for OLS, 10 for ARCH and an expected number of (2. Thus, the number of outliers fors004 RORERT F. ENGLE, ordinary least squares is reasonable; however, the timing of their occurrence is far from random. The ARCH model comes closer to truly random residuals after standardizing for their conditional distributions. ‘This example illustrates the usefulness of the ARCH model for improving the performance of a least squares model and far obtaining more realistic forecast variances, University of California, San Diego Manaseript received July, 1979; ital reuton reeived Hh, 1961 APPENDIX, roar oF Taeont 1: Let 2) = ORB Wt) Fist ls shown that there i an uppettangular yr race A andr f vector b such that (82) lw) yaad b+ Aig For any zero-mean nornal random variable, with variance Fe) =o Th i- ty Ay cause the conditional istriboton ofy is normal a EUIIH.d =A fh aio mt ay” TL Gi 0. Exparding chs gression establishes tha the moment sa finear combination of; . Purhermors, only powers of fess than oe equ tp 2 ate regured, therelace, A ia (42) upper ancalae Now EG |g) BF ACY Awd oc in general Boba EAL APE AMD EAB ‘Beenie the series stats indefinitely far in the past with 2r frie moments, th limi as O65 (0 Infinity exss if and ony al the eigenvalues of 4 Te within dhe unit cir "The limit cam be ween a8 odo ay sia, BO) which does nat deperd upon the conditioning variables and does nat depend sport ene, ‘hs is an exptesion for the stallonary morunts of the ucconsiwoaal distububan of A Bema ate.HETEROSCEDASTICITY loos Teemains ony esti thatthe condition ip the theogen is necessary ard suficient to have all eigenvalues le within the unl cee. As the mate fas aleady een ove to be upper Wangs, ‘he disgnnal elements ace ‘Ne eigenvales. From (A), 18 See thatthe diagonal elemels ace singly or ther n= faa =a tore ee, ecu a na any, the gle oo be ea cle Tea thea bee Sten oe tian 7 Neha ya pee ch Ne oust suru Whe mtn eater oes dyn nec Satna he ec os am le er are 2 belo aoe se Rete st tio ue satya aan tae tsa ta te The rahi tha ancy tn tien asin fo gon that's anh ne atoret inte sre roar or Tox 3: Let w(t uP a Then i te af the companion matrix 4, A9) nL a bE Ay whete B= fa0,0,-., 0) and cm a ¢ aa|l 9 oo ot ag oo 1a ‘Taleing successive expectations BOG MAME A EAD RANE AM Because the series starts indefinitely far in the post with fine variance, if, and only fal eigenveues “ie within the unit eel, the Jest exists ais given by 88 Lim FOL Geead OA As this does not depend upon intial conditions oF or 1 his veto is the comman varanee (otal ‘Ass wall known in time series analysis, ths condone equivalent to the ccadisn that al the cots ofthe characteristic equation, formed Tom fe a's le oubice the wat cice. See Anderson [2 p. TM. isa he ut of the Gee elersent can he rowzten a8 a stenf(i- $s) ono. roar oF Trane 3: Cleads, under the conditions, AG} 2 ag > 0, etabshing pat (a). Let Ae = EPR RR 38 DoE Pon Now there ace thiee cases; im fem, ard Gem Mh s> mt thea f-,E dh 4 and the ‘cotalitons! expectation of [2 y|sfinste, because the condisoral density i natal el m, hen the expeciavon becomes EE") |fp-y Again, because the conitional densiy ¥ normal, all1006 ROBERT F. ENGLE moments ens ncuding the expectation of the third power of the absolute value. If (< my, the eapectaion i taker in (Wo pars fist wi eeszect fo ¢~ ~ L Fons RE (Il EE Led Merae atone} wanel de $a wa0gF bel domtl * Sates Ia the final expression, the inal index on 6 Is lager and, cherefore, may fall into either of the preceding eases, whieh, therefore, ceablisties the existence af the tera (thre rema terms wit Pape, the tecursion cane repeated. Ae al as are finite, an expression for, can be Weiler ‘cd constant times she tied abuaite moment of £» conan oN Jy» led apethercorscant Times the fist thea momen. As case are Hoch condivarally mrmal abd 3 the constants must be fine ay they have 3 fle puraber of terns, the second part of the regulanty condition has been cntablshes OLD. “To exes Theorem 4 «care symmetry argument es required, beginning with the following lemma Linon: Let wand » be any tao random variables. B(g(u, |») wil be an endcymmate faction af if gis oniismmmeiie ine, the conditinal demay of tl ets symaetvic trv, and te speciation Proor: BEglu 0) oy = EL g(4,0)|—0) — becaoseg is ancsymmettic in 0 = —Flgtaoi[o) because the condiional density is symuetric a Proor of Tuonine 4: They element of fy é given by by the aia rule. tthe expectation af the term in square brakes, condivonal on fy 18200 fe all J. then {he theorems proven 1%, ah, ah (aan 8 Re gets eens tithe exogenous or it is Lapued dependent variable, in which ease ii teluded in ef 4 a Ah . WF ta, Beg MYHETEROSCEDASTICITY 1007 boy parc (2) of the cegulaity condions and this inter is frite by pact (b of the condition. Hence cach ter 15 fice. Naw lake the expectation in fo eps, fir Bath respect [0 THIS must Thatefore alga be finite ad (5 ae ey ton) ae By the symmetry assumption," ie ayes i 6 ony Oh /A5 it snteayeaetrie. Therefore the whole expression 6 aalt-syeametric m4-n, ich part of the Conditioning se fy Recanse iesymmetic, te conditions! density mus be syrmerrie (9G. and the lemma cae be invoked to show that gC} is ant-symmette. Fiaally, aking expectations of g conditional on gives 2, because the density of 6 ‘conditional onthe pasts ¢spenietic (anwal} densi and We theorem sestabished. QED. REFERENCES (8) avers, ‘Ts “Regression Analisis when the Variance of the Deperdeat Vavable is Propor Tonal to ine Square of Us Expectation,” Jounal of the American Static Arsciatian (1973), 924-534 (2) Anpessan, T. We The Stausueat Analy of Tine Series. New York: Soha Wiley and Sos, 1s. {3) BEESLEY, Daven: “On the Ellicient Computation of Now Lines Full Taformation Maxine ikelitond Esiimaio,” paper sreseted to the European Meetiegs of he Econometric Soceiy, Athens, 1918 (4) Barusct, TS. ano A. R, Paoaw: “A Simple Test for Heterowedasticiy and Random Coethclent Variation,” Econometrica, 46578, 1287-1254. (S| "The Lagrange Muller Tes and es Applications to Model Specitsason." Rewtew of Beonomic Sti, 411980), 239-254, (6) Cox, DR. ave DV. iver: Theoretical Stonatic, Landon: Chapman ang Hal, 1974 17] Caowben, M. 1 "Maximum Likelihood Eststation for Depencent Onsevations” Javad of the Ropal Staisueal Soe. Series By 24.976), 45-53. [8] Davtoson, IE. HC, D. F. Hlewoey, F. Seta, axD S. Yeo" "Econometric Modeling of the “Aggregate Tine-Series Relationship Between Coasuimers' Expenditure and Tacome the alted Kingdom,” The canon Fourat, $8 1978, 661-68 [9] Exot, RFA General Appeaach to the Corstuction of Model Diagnostics Based upon the Lagrange Mulipiee Paveipie” University of Calfoca, San Diego Discussion Paper 79-43, , (19) “Eniaes ofthe Varagee of U.S fllation Based om the ARCH Movie,” University of California, San Diego Diseusion Paper +14, 1980. (11) Pareoman, Atcron: “Nobel Leoue: fflation aed Unemployment" Sauna! of Petiea! Econ comp, ASISTD), ASL-AT2 (12) Goursey, [Gu “Testing Aasinst General Autoresresive and Moving Avezage Ercor Models Wien she Reqressore Inchade Legged Depeacent Variables” Econometrica, 41978), (29% 13a, 113] Gravors, C. W. Ta AND A, ANDERSEN: te Juroduction 1a Bilmear Tine. Serer Models (Gottingen: Vandertoeck ant Rupees, (97% [D4] Kees, M.S. “The Varabiity of Expectations in Hyperintiations," Jounal of Political Eanamy. 85197), 817-827, 15) Kuss, Bs “The Damand for Qualiy-Adjusted Cash Balances: Peice Uncerainty in the US, Demand for Money Function," Journal of Pottiea! Economy, 85(1977), 692-115. [16] Eaters R. E, In "Some International Evicence on Output-llaion Tadeo” Amencan Econone Review, 41973), 326-334 [10] MeNees, 5. 5. "Phe Forecasting Recotd foc the 1910" New England Economic Review, September /Oetober 1979, 33-88 [ia] Wate, FA Heteoncedasieuy Consistent Covariance Matrix Estiatae and a Direct Tes for Heteroscedascy,” Beanzmatien, 481980), 817-838,