al which we n raieimace in ge! ent eral ce jeuro exchange rte ¥°Y; =Bo + ByY,-1 + By Y,_9 +...4 BpYj_-p +p = (16.6) where up is a white noise error term, ‘Model (16.6) is called an autoregressive model of order p, AR (p), for it involves regressing Y at time ¢ on its values lagged p periods into the past, the value of p being determined empirically using some criterion, such as the Akaike information criterion. Recall that we discussed autoregression when we discussed the topic of autocorrelation in Chapter 6. ‘The moving average (MA) model We can also model ¥,as follows: =Co HC yty + Cott; +--+ Cqtly—g (16.7) Thatis, we express Yas a weighted, or moving, average of the current and past white noise error terms. Model (16.7) is known as the MA (q) model, the value of q being _ determined empirically. ___ The autoregressive moving average (ARMA) model _-We can combine the AR and MA models and form what is called the ARMA (p,4) model, with p autoregressive terms and q ewes average terms. Again, the values of _pand gare determined empirically. _ The autoregressive integrated moving average (ARIMA) model As noted, the BY inethodotogy is based on the assumption that the underlying time ‘series is stationary or can be made stationary by differencing it one or more times. This is known as the ARIMA (p,4,q) model, where d denotes the number of times : s has to be differenced to make it stationary. In most applications d = be a saya take only the first differences of the time series. Of course, if a time slteady stationary, then an ARIMA (p,d,q) becomes an ARMA (p,q) model. al question is to determine the appropriate model in a given situation. is question, the B] methodology follows a four-step procedure: : Determine the appropriate values of p, d and q. The main ‘are the correlogram and partial correlogram. ‘Once we identify the model, the next step is to estimate the é chosen model, In some cases we can use the method ordi- (OLS), but in some cases we have to resort to nonlinear (in methods. Since several statistical packages have built-in ahube have to worry about the actual mathematics of estimation. BJ ARIMA modeling is more an art than science ‘skill is required to choose the right ARIMA model, for we u) that the chosen model is the correct one. One simple Js from the fitted model are white noise; if they n model, but if they are not, we will have to start methodology is an iterative process. sat moony‘304 | TIME SERIES ECONOMETRICS uitimate test of a successful ARIMA model jag, within the sample period as well as outside the mgt le Step 4: Forecasting: The forecasting performance, period. 16.3 An ARMA model of IBM daily closing prices, January 3, 2009 to October 31, 2002 " wed that the logs of the IBM daily closing prices (Lciog, . BAS sacklsoes a the first differences of these prices (DLCLOSE) were Staion, ary, Since the B} methodology is based on stationary time series, we will work yi \© DLCLOSE instead of LCLOSE to model this time series, where DLCLOSE stands for oh |. the first differences of LCLOSE. ti TTo see which ARMA model fits DLCLOSE, and following the BJ methodology, xe eo show the correlogram of this series tip'to 50 lags (Table 16.4), although the picture does not change much if we consider more lags. ‘This correlogram produces two types of correlation coefficient: autocorrelation (AC) and partial autocorrelation (PAC). The ACF (autocorrelation function) shows correlation of current DLCOSE with its values at various Jags. The PACF (partial autocorrelation function) shows the correlation between observations that are periods apart after controlling for the effects of intermediate lags (i.e. lags less than 4)? The BJ methodology uses both these correlation coefficients to identity the type of ARMA model that may be appropriate in a given case. ‘Some theoretical patterns of ACF and PACF aré shown in ‘Table 16.5. Notice thet the ACFs and PACEs of AR(p) and MA(q) have opposite patterns: in the AR(p) case the ACF declines geometrically or exponentially but the PACE cuts off after a certain number of lags. The opposite happens to an MA(q) process. Keep in mind that in a concrete application we may not observe the neat patterns shown in Table 16.5. Some trial and i etror is unavoidable in practical applications. penacuinins t0 our example, we see that both ACF and PAC functions alterat? mz iat per and positive values and do not exhibit an exponential decay for any A careful examination of the correlo f ny PACE shows the neat pattern descr gram shows that neither the ACF nor nibed in Table 16.5. To. see which correlations are s™ tistis ria, < Ds ich correlations a en Bs Saiihoent rca, that the standard error of a (sample) correlation coed sib ps 1 Ma ATA * 0.087, where m is the sample size (see Ea. (132)! 041960087) - ain ee interval for the true correlation coefficients is abou" unds are statistically signi CF aid BRCr Se Saat at the 5% level, On this basis, it seems both AC » it (See the confidence bands 18, 22, 35 and 43 seem to be statistically signiica™ Since we do in the preceding fi igure), “| Med in Table 165, noe serait Pattern of the theoretical ACF and PACE %™* shots S4PPOSe We fit an AR ed by tial and error. , We can proceed i del at | esults af? own in Table 16,6, lags 4, 18, 22, 35 and 43. The results ; AS you can see, the Coefficients of AR(35) and AR(43) ‘isis akin to the \ ‘model Partial regre , — flag 2 Ante Bk OF the kth regresans aS i” 8 multiple regression. na k variable £08 Fortheinfuence ofthe otter ens PSC of that variable on the re3tes™ other repressors in the model,F of DCLOSE of IBM stock prices, ENS )oleo|fasfen}im fos fro] lo = Si 1019ORMRWAY Exponential decay neat pes through lags | eclies exponent —~ . Py : Significant »pikes warp | tia | | SB) | Ale) [0 aS aa ‘ 4) [Sanpete 7200 820720 mpe (agus 0825879 [2467745 | 1.843242 ~2.138565 [-=1.401768 | 1.332428 0.000811 0.026409, 4.445734 ~4.404059 2.089606 0.000869, din the model. fever, it should be noted that when the i Were tested for serial correlation, we did not in Table 16.6 may be a candidate for furthe® eri sc | ot significant, we can drop these fo" nly AR(4), AR(I8) and AR(22) te" from this regression also see™ Receding models, we can we the oo Ke the choice, Although there is not ¢ > inthe two tables, numerical the a ‘the model in Table 16.7 than in Table va criteria, we choose the model with ce the value that is most nee in 6 16.7 is preferred over the Om jimonious than the one Pt rameters.AR (4,18,22) model of DCLOSE, ‘Variable: D(LCLOSE) Squares 2/03/2000 8/20/2002 664 after adjustments achieved after 3 iterations 9.000937 | 0.000944] -0.990949 304 0.101286 | 0.038645 | 2.620890) 0099 0.082566 0.039024 ANSTO 067 | 0.091977 0.039053 2.355157 001% | 0.027917 Mean dependent var 0.000980 0.023499 S.D. dependent var 0.026416 Akaike info criterion ~4.447488 Schwarz criterion 4.420390 Durbin—Watson stat 2.102050 Prob(F-statistic) 0.000315 tried the counterpart of Table 16.6 using five lagged MA terms 2t lags the coefficients of lags 35 and 43 were not statistically sig- we estimated the MA equivalent of Table 16.7, and obtained che -Thé residuals from this regression were randomly distributed. 0,000878 | _-1.011247 0.038075 | —_ 2.275167 (0.038682 | _ -2.567953 0.0104 | | (0.038958 | _ 2.880715 @.0081 | | | Mean dependent var 0.000828 | SDD. dependent var 002638 | Akaike info criterion —M57% Schwarz criterion Durbin—Watson stat 2.104032 | Prob(F-statistic) ou 22)? R(4,18,22) or MA(4,18:22)? ee ptomaton esters we over the AR model, although the difference re lowest——— | simply # weighted average of the stochas Recall that the MA model is of the log closing prices of IBM are Stationay hit term, But since the bani a Hare let us see if we « the MA model, idevlog Baton ne We i tees, Aer som experimentation, We obtained the ma Table 16.9. wai able 16:9. ARMA [(4,22)(422)] model of DLCLOSE. ae ‘Vatiable: DLCLOSE). 1 Se monet ace 0.001055 | 0.934089 0.3506 | | __.|, -0.229487| 0.061210 | _-3.749152 0.0002 || | 01641421 | ~ 0.062504] -10.26202 0.0000 | | =) | 0361848 | 0.060923,| | | 5.939484 0.0000 < -[o6i8302 | ~<0.055363 | —11.16808 0.0000 | | nslsvIUpy AlVi 913 batsmiie> o> | git ehfsadependentyer | “Sosa 2b dshencen var i 1 __ Akaike info criterion 2 0 Schwargériterion Durbin—Watson stat 2.111835 _Prob(F-statistic) 0.000002 Tie. Sponge bon Bh Mreeuibe tess 823 :anci, Criteria, it seems this is the “best” model. The resid- for unit root and it was found that there was no writ ‘model are stationary. Also, on the basis autocorrelation discussed in Chapter 6, it was found Serial correlation in the residuals. “ 4,22,4,22) is probably an appropriate mo of the logs of daily closing IBM pr'* TEs BE Oy PELBIADET SO. ROE ; We can use it for forecasting, uss “two types of forecast: stat it and lagged values of te se first period forecast, We" 5. 10 This: ‘is shown in Figure 6s oe ices, @ closing IBM prices, Pnlebe : prices, the forecest* een this automatically.of forecast. The accompanyi ‘the forecast that we saw before, te error, mean absolute Percentage For our example, this coefficient ig pra quite good. ‘This can also be seen froy and forecast values track each oth: ng table gives the same measures namely, the root mean square, ertor and the ‘Theil Inequality ictically zero, Suggesting that the mn Figure 16.5, which shows how er, Forecast: LCLOSEF ‘Actual: LCLOSE Forecast Sample: 1/03/2000 10/31/2002 Adjusted Sample: 2/03/2000 4/26/2002 Included Observations: A Root Mean Squared Error 0.025754 Mean Absolute Error 0.019017 ‘Mean Abs. Percent Error 0.414809 ‘Theil Inequality Coefficient 0.002788 Bias Proportion 0.000005, Variance Proportion __0,001310 Covariance Proportion _ 0.998685 Actual: LCLOSE | Forecast Sample: 1/03/2000 10/31/2002 Adjusted Sample: 2/03/2000 10/31/2002 Included Observations: 664 Root Mean Squared Error 0.259823 | Mean Absolute Error 0.213622 Mean Abs, Percent Error 4.595556 ‘Theil Inequality Coefficient 0.028770 Bias Proportion 0.642: Variance Proportion 0.006474 Covariance Proportion 0.351020 Economic rorecasins | 309reason for this is that we use the previous forecay fovecasts and if there is an erx0r in the previous, ay - along the time axis. The » in computing sub: ent poaluet tha ror will be envied forward. ; Before proceeding further the reader is encouraged t0 acquire more recg, and see if the pattern observed in the current sample continues to hold in gy, sample. Since the ARIMA modeling is an iterative process, the reader may y,""" try other ARIMA models to see ifthey can improve on the models discuss section. 16.4 Vector autoregression (VAR) In the classical simultaneous equation models involving » endogenous (ic, depen, dent) variables, there are m equations, one for each endogenous variable. ;,., equation may contain one or more endogenous variables and some exogenous ya, Before these equations can be estimated, we have to make sure that the prob. lem of identification is solved, that is, whether the parameters or set of parameters we consistently estimated, In achieving identification, often arbitrary restrictions : sposed by excluding some variables from an equation, which may be present in the other equations in the system. "This practice was severely criticized by Sims, who argued that if there are m variables, they should all be treated on an equal footing; there should "not be any distinction between endogenous and exogenous variables.'? So each iation should haye the same number of regressors, It is for this reason that Sims veloped the VAR model. ain the ideas behind VAR, we will first consider a system of two variable ¢ we discussed the relationship between 3-month and 6-month Tis rom the point of view of cointegration. Here we discuss it f0 ng the two rates, using the VAR methodology. For this Po llowing two equations: ies (168) (169) s are wh uss the simultaneous equations models, for they ate "°° $ and 1970s. For an overview, see Gujarati/Porter °P Totes jand reality, Econometrica, 48, 1-18. ing direction, For our pu} tor. Since in VAR we are dealing na column, As we are dealing wit vectors a VAR system. «can Ht0E erg with 0° dotECONOMIC FORECASTING | 3 tem resembles a simultaneous equation system, but the funda- rence between the two is that each equation contains only its own and the lagged values of the other variables in the system. But no , of the two variables are included on the right-hand side of these number of lagged values of each variable can be different, in most e use the same number of lagged terms in each equation. te VAR system given above is known as a VAR(p) model, because we ilues of each variable on the right-hand side. If we had only one of each variable on the right-hand side, it would be a VAR( 1) model; ged terms, it would be a VAR(2) model; and so on. ealing with only two variables, the VAR system can be extend- riables, Suppose we introduce another variable, say, the Federal will have a three-variable VAR system, each equation in the gged values of each variable on the right-hand side of each plan yi relationship between them. If we have a three- m, there can be at most two cointegrating relationships variables. In general, an n-vatiable VAR system can have at Seton...We showed earlier i: det VAR for the ee a series to show how VAR forecas notat ion, we consider a VAR(I) (16.16) (16.17) fe estimated values of the ing the sample data from it to forecast the values , where 71 is specified. for time (¢ + 1) is given by (16.18) > parameters from the sample ow? 9: (16.19) al yalues of TB3 and T8612 Yote that, as usual, a hat over od (¢ + 1), namely, (16.20) same meas, but modify it® (16.21)py zapNw, Sooper aE NOFEAETONTT OAS zi z2insdoom [sutos orl introduce the reader to four important topics ting with linear regression models, th Box ins methodology, (3) multivari- ression, and (4) the nature of causality am, un ze used in forecasting sales, production, a abe: economic topics. In discussing shed between point and interval fore- and unconditional forecasts. We per capita consumption expenditure ne in the USA for the period 1960-2004 see how the fitted model performs in sed forecasting with autocorrelated of forecasting, which is popularly known n the BJ approach to forecasting, we analyze ory or purely moving average of ran- ination of AR (autoregressive) ‘the time series under study is by diffevencing it one or morec ton term, which ‘etror correction _ (Granger) causes sible that there lerlying variables are “run the causality jonstationary, but ae for causality fan 7‘STOCHASTIC REGRESSORS AND THE METHOD OF INSTRUMENTAL VARIABLES | 361 X; is random, We now distinguish three cases.> are distributed independently: n this case for all practical purposes we wseOLS.AsGreenenotes: , therefore, is that the important results we have obtained for the least squares estimator, unbiasedness, and the Gauss— rem holds whether or not we regard X as stochastic. Pilitindorg Yo Iqoon0. o¢i yniew fi 97 emporaneously (i.e. at the same time) uncorrelated: ‘This is a #1. In this case the classical OLS results hold only asymp- large samples. instte 2 independently distributed nor are contemporaneously this, the more serious case, the OLS estimators are not only inconsistent. Intuitively, the reason for this is: ")1s for Correlation between regressors and the error there ate four reasons why the Fegrentor(s) may be correlated with the surement errors in the regrebsor(s) ted variable bias ous equation bias eptession model with serial correlation in the error term, that weistudy these Sources of correlation between fegressor(s) and cd of instrumental variables, errors in the Hegressor(s) Aoted that if there are errors of mew care biaséd as well as inconsistent, To get a glimpse of this, we con- id permanent ticome hypothesis (PIM) of the late Nobel laureate which can bé explained as follows? By t+ By Xj uy; 0< By <1 (19.8) nt, or observed, Consumption experiditute, X} © permanent income ce, OF etror, term. By here represents the marginal propensity to PC), that is, the increase in consumption expenditure for ati addition. rth of increase in the permanent income, that is, the average level of tto bein the future! do not have readily available measures of Permanent income, So hanent income we use observed or current income, Xp whieh of measurement, say w). ‘Therefore, we can write ea. in surement in the tegressor(s) the (19.9) al to permanent income plus errors of measutement (19,10) anrentetcore n and measurement error is i. mean, is serially uncorrelated, and i ye composite error teem vj is fanning 2) 0) it can be shown that (19.4) Ancome (XP), but to keep pe364 | SELECTED TOPICS IN ECONOMETRICS This result shows that in the regression (19.10) thie tregressor ‘X; and the error term yjare correlated, which violates the crucial assumption of CLRM that the error terry and the regressor(s) are uncorrelated. : ; ‘As a result, it can be shown that the OLS estimate of By in Eq. (19.8) is not onty biased but is also inconsistent. It can be proved formally that (see Exercise 19.3) (19.12) where plim means the probability limit, which, as mentioned before, we use to estab. lish the consistency property of an estimator. ‘Since the term inside the bracket is expected to be less than 1, b2 will not converge by to its true MPC value whatever the sample size. If By is assumed positive, which makes sense in the present case, by will be less than the true By ~ that is, b, will underestimate By. More technically, it is biased toward zero. As this exercise shows, errors of measurement in the regressor(s) can pose serious problems in estimating the true coefficient.!? How, then, can we measure the true MPC? If somehow we can find a proxy or a tool or an instrument for permanent income so that that proxy is uncorrelated with the error term but is correlated with the regressor (presumably highly), we may be able to measure the true MPC, at least in large samples. This is the essence of the ‘method of instrumental variable(s). But how do we find a “good” proxy? We will __. answer this question shortly. apter 7 we discussed several cases of specification errors, such as the omission ‘elevant variables, incorrect functional form, and incorrect probabilistic assump- (19.13) Xp is education as measured by years of schooling, and of ability are hard to come by, suppose, instead of estimat- the following function: (19.14) " from the wage function. In this case, ee ogni! (19.15) cause nl such a problem b do not pose such a p ae n till obtained unbiased estimates o -of the estimators are larger the"STOCHASTIC REGRESSORS AND THE METHOD OF INSTRUMENTAL VARIABLES | 367. method of instrumental variables main problem with the use of OLS in regression models that contain one or regressors that are correlated with the error term is that the OLS estimators are biased as well as inconsistent, Can we find a “substitute” or “proxy” variables for the suspect stochastic regressors such that the proxy variables produce consistent tors of the true (population) regression coefficients? If we can do that success- such variables are called instrumental variables or simply instruments. How swe find such instruments? How do we know they are good instruments? Are there ways to find out if the chosen instrument is indeed a good instrument? To answer these questions, let us start with the simple linear regression given Bq, (19.2), Suppose in this regression that regressor X is stochastic and that it is Bs lated with the error term u. Suppose a variable Z is a candidate instrument for X. To bea valid instrument, Z must satisfy the following criter' Instrument relevance: ‘That is, Z must be correlated, positively or negatively, __ with the stochastic variable for which it acts as an instrument, variable X in the it case. The greater the extent of correlation between the two variables, the er is the instrument. Symbolically, (19.24) ity: Z must not be correlated with the error term u. That is, (19.25) ‘we proceed further, it may be noted that if we have a multiple regression ral regressors and some of them are correlated with the error term, we must ent for each of the stochastic regressors. In other words, there must be instruments as the number of stochastic regressors in the model. But nore to say about this later. n see, all thése conditions may be hard to satisfy at the same time. So 3d good instruments in every application. That is why sometimes ntal variables sounds chimerical, although there are examples of 15 of cn interesting but somewhat questionable example of IV appli- wanted to find out the relationship between student perfor- mpétition. She estimated the following regression: By + By (Number of school districts) + error term or is stochastic, she used the number of streams in a for the number of school districts, for she observed “districts also had a lot of streams; presumably the {for the school districts,!® a atau to oncate the eft of @ siversity of Chicago, 48, 2005, pp. 267-79. ‘schools benefit students and taxpayers? etn368 | setecteD tones ecoNomETRICS How does IV estimation work? The answer follows. nh Tee IV estimation sonore ( s To show how IV works, we will continue with the two-variable regression. As ye know the OLS estimator of By in Eq, (19.2) is: phe! by Bx? oe 2%, -. id 9 = (Y; we 20 Pad ED ng a and ba eit gad EL ni tt ro i where 2 =2Z;-Z. yf : slovtregsy ‘Caution: Do not just put 2; for ; in the formula for bp given above and note care- ant ni “fully that the denominator has both zand x terms, 4 aie |. Now noting that ¥j= B; + BoX; +m; and, therefore, xillsgilodmny?, ages Yt = Bary +4 -i2) we obtain 2a gg no Bad tl, 9 agkton dausn Sedo a (19.27) Dikinbabedi aay biti ad BHP adorn ieore Poe heen Of course, if Z = X, A an 7 (19.28) fsa the only dlference; from the usual OLS estimator of B, is that we nt estimated from the IV estimator. ing that in the soon ion coy (Z, u)= 0, taking the probability (19.29) nator of By, in small samples it is biased. Furthes the IV estimator is distributed as follows: Meeretirenii4 ) (19.30) or involves the squared (population) co" i in large samples the IV estimator mn Eq, (19.27) involves sample quantiSIOCHASHIE REGHESSORS AND THE MEIHOD OF INSTRUMENTAL VARIABLES | 36% lly diatributed With mean equal te 1 population value and varianee By contrant, the usual OLS, eatinator haw the varlanee COS) <1 the variance of the 1V ostinato will be larger than the variance We OLS ton enpectally #6 I p%y In mall, In other wards, the IV estimator eeffictent than the OLS estimator, It p%., tn umall It suggents that Z isa weake nt for X, On the other hand, (fit in large, IC nuggente that itis a strong, Mt for X, Romielidea how far the variuneas of 1V and OLS estimators can diverge, In this case, variance of the IV estimator in 25 times an large Whew © OL, ls 100 tlinen larger: In the extreme case, Boi {he varlance of the LV estimator ts infintte, Of courne, If Pry = 1, the two a er way of saying that the variable X is its own Note that in practice we estimate py. by its sample counterpart, r,.. the variance of the IV estimator given in Eq, (19,30) to establish con- and test hypotheses, asstiming that our sample sine te reasonably the variance of the IV estimator is heteroscedastic.!8 Therefore eattns Point to note about the preceding discussion is that in obtaining lates via the IV method we Pay a price in terms of wider confidence of the larger variance of IV estimators, especially if the selecrot veal Proxy for the original regressor, Again, there is no such thing os ic regressor(s) correlat- Cameron and ‘Trivedi conducted a Monte Carlo simulation assumed the following: (19.32) (19.33) (19.34) tin the reression of ¥) on X; is assumed known sor X} is equal to the instrumental variable Z, imed that Z; was distributed normally with ‘were jointly normally distributed, each with teste ce ore ail Ty