Kinh tế học - Chapter 4: Applied econometric time series 4th ed. walter enders

Chapter 4Applied Econometric Time Series 4th ed.Walter EndersThe Random Walk Modelyt = yt–1 + et (or Dyt = et).Given the first t realizations of the {et} process, the conditional mean of yt+1 isEtyt+1 = Et(yt + et+1) = ytSimilarly, the conditional mean of yt+s (for any s > 0) can be obtained fromHencevar(yt) = var(et + et–1 + ... + e1) = ts2var(yt–s) = var(et–s + et–s–1 + ... + e1) = (t – s)sRandom Walk Plus Driftyt = yt–1 + a0 + et Given the initial condition y0, the general solution for yt isEtyt+s = yt + a0s.E[(yt – y0)(yt–s – y0)] = E[(et + et–1+...+ e1)(et–s+ et–s–1 +...+e1)] = E[(et–s)2+(et–s–1)2+...+(e1)2] = (t – s)s2The autocorrelation coefficient= [(t – s)/t]0.5Hence, in using sample data, the autocorrelation function for a random walk process will show a slight tendency to decay.Table 4.1: Selected Autocorrelations From Nelson and Plosser Worksheet 4.1Consider the two random walk plus drift processes yt = 0.2 + yt1 + yt zt = 0.1 + zt1 + ztHere {yt} and {zt} series are unit-root processes with uncorrelated error terms so that the regression is spurious. Although it is the deterministic drift terms that cause the sustained increase in yt and the overall decline in zt, it appears that the two series are inversely related to each other. The residuals from the regression yt = 6.38  0.10zt are nonstationary. Scatter Plot of yt Against zt Regression ResidualsWorksheet 4.23. UNIT ROOTS AND REGRESSION RESIDUALSyt = a0 + a1zt + etAssumptions of the classical model:both the {yt} and {zt} sequences be stationary the errors have a zero mean and a finite variance.In the presence of nonstationary variables, there might be what Granger and Newbold (1974) call a spurious regressionA spurious regression has a high R2 and t-statistics that appear to be significant, but the results are without any economic meaning. The regression output “looks good” because the least-squares estimates are not consistent and the customary tests of statistical inference do not hold. Four casesCASE 1: Both {yt} and {zt} are stationary. the classical regression model is appropriate. CASE 2: The {yt} and {zt} sequences are integrated of different orders. Regression equations using such variables are meaninglessCASE 3: The nonstationary {yt} and {zt} sequences are integrated of the same order and the residual sequence contains a stochastic trend. This is the case in which the regression is spurious. In this case, it is often recommended that the regression equation be estimated in first differences. CASE 4: The nonstationary {yt} and {zt} sequences are integrated of the same order and the residual sequence is stationary. In this circumstance, {yt} and {zt} are cointegrated. The Dickey-Fuller testsThe f1, f2, and f3 statistics are constructed in exactly the same way as ordinary F-tests:Table 4.2: Summary of the Dickey-Fuller TestsTable 4.3: Nelson and Plosser's Tests For Unit Roots p is the chosen lag length. Entries in parentheses represent the t-test for the null hypothesis that a coefficient is equal to zero. Under the null of nonstationarity, it is necessary to use the Dickey-Fuller critical values. At the .05 significance level, the critical value for the t-statistic is -3.45. Quarterly Real U.S. GDPlrgdpt = 0.1248 + 0.0001t  0.0156lrgdpt–1 + 0.3663lrgdpt–1 (1.58) (1.31) (1.49) (6.26) The t-statistic on the coefficient for lrgdpt–1 is 1.49. Table A indicates that, with 244 usable observations, the 10% and 5% critical value of  are about 3.13 and 3.43, respectively. As such, we cannot reject the null hypothesis of a unit root. The sample value of 3 for the null hypothesis a2 = g = 0 is 2.97. As Table B indicates that the 10% critical value is 5.39, we cannot reject the joint hypothesis of a unit root and no deterministic time trend. The sample value of 2 is 20.20. Since the sample value of 2 (equal to 17.61) far exceeds the 5% critical value of 4.75, we do not want to exclude the drift term. We can conclude that the growth rate of the real GDP series acts as a random walk plus drift plus the irregular term 0.3663lrgdpt–1. Table 4.4: Real Exchange Rate Estimation H0:  = 0LagsMean  /DW F SD/SEE1973-1986Canada0.022(0.016)t = 1.4201.050.0591.880.1945.471.16Japan0.047(0.074)t = 0.6421.010.0072.010.22610.442.81Germany0.027(0.076)t = 0.2821.110.0142.040.85820.683.711960-1971Canada0.031(0.019)t = 1.5901.020.1072.210.434.014.004Japan0.030(0.028)t = 1.0400.980.0461.980.330.017.005Germany0.016(0.012)t = 1.2301.010.0381.930.097.026.004EXTENSIONS OF THE DICKEY–FULLER TESTyt = a0 + a1yt–1 + a2yt–2 + a3yt–3 + ... + ap–2yt–p+2 + ap–1yt–p+1 + apyt–p + etadd and subtract apyt–p+1 to obtainyt = a0 + a1yt–1 + a2yt–2 + ...+ ap–2yt–p+2 + (ap–1 + ap)yt–p+1 – apDyt–p+1 + etNext, add and subtract (ap–1 + ap)yt–p+2 to obtain:yt = a0 + a1yt–1 + a2yt–2 + a3yt–3 + ... – (ap–1 + ap)Dyt–p+2 – apDyt–p+1 + etContinuing in this fashion, we obtainRule 1:Consider a regression equation containing a mixture of I(1) and I(0) variables such that the residuals are white noise. If the model is such that the coefficient of interest can be written as a coefficient on zero-mean stationary variables, then asymptotically, the OLS estimator converges to a normal distribution. As such, a t-test is appropriate. Rule 1 indicates that you can conduct lag length tests using t-tests and/or F-tests onyt = yt–1 + 2yt–1 + 3yt–2 + + pyt–p+1 + tSelection of the Lag Lengthgeneral-to-specific methodologyStart using a lag length of p*. If the t-statistic on lag p* is insignificant at some specified critical value, re-estimate the regression using a lag length of p*–1. Repeat the process until the last lag is significantly different from zero. Once a tentative lag length has been determined, diagnostic checking should be conducted. Model Selection Criteria (AIC ,SBC)Residual-based LM testsThe Test with MA ComponentsA(L)yt = C(L)et so that A(L)/C(L)yt = et So that D(L)yt = etEven though D(L) will generally be an infinite-order polynomialwe can use the same technique as used above to form the infinite-order autoregressive modelHowever, unit root tests generally work poorly if the error process has a strongly negative MA component.Example of a Negative MA termyt = yt-1 + εt – β1εt-1; 0 t and zero otherwise.Now consider a regression of the residuals If the errors do not appear to be white noise, estimate the equation in the form of an augmented Dickey–Fuller test. The t-statistic for the null hypothesis a1 = 1 can be compared to the critical values calculated by Perron (1989). For l = 0.5, Perron reports the critical value of the t-statistic at the 5 percent significance level to be –3.96 for H2 and –4.24 for H3. Table 4.6: Retesting Nelson and Plosser's Data For Structural Change The appropriate t-statistics are in parenthesis. For a0, 1, 2, and a2, the null is that the coefficient is equal to zero. For a1, the null hypothesis is a1 = 1. Note that all estimated values of a1 are significantly different from unity at the 1% level.T k a0 1 2 a2 a1Real GNP620.3383.44(5.07)-0.189(-4.28)-0.018(-0.30)0.027(5.05)0.282(-5.03)Nominal GNP620.3385.69(5.44)-3.60(-4.77)0.100(1.09)0.036(5.44)0.471(-5.42)IndustrialProd.1110.6680.120(4.37)-0.298(-4.56)-0.095(-.095)0.032(5.42)0.322(-5.47)Power Formally, the power of a test is equal to the probability of rejecting a false null hypothesis (i.e., one minus the probability of a type II error). The power for tau-mu isNonlinear Unit Root TestsEnders-Granger Test yt = It1(yt–1 – ) + (1 – It)2(yt–1 – ) + tLSTAR and ESTAR TestsNonlinear Breaks—Endogenous BreaksSchmidt and Phillips (1992) LM TestThe overly-wide confidence intervals for  means that you are less likely to reject the null hypothesis of a unit root even when the true value of  is not zero. A number of authors have devised clever methods to improve the estimates of the intercept and trend coefficients. yt = a2 + t The idea is to estimate the trend coefficient, a2, using the regression yt = a2 + t. As such, the presence of the stochastic trend i does not interfere with the estimation of a2.LM Test ContinuedUse this estimate to form the detrended series as Then use the detrended series to estimateSchmidt and Phillips (1992) show that it is preferable to estimate the parameters of the trend using a model without the persistent variable yt-1.Elliott, Rothenberg and Stock (1996) show that it is possible to further enhance the power of the test by estimating the model using something close to first-differences. The Elliott, Rothenberg, and Stock TestInstead of creating the first difference of yt, Elliott, Rothenberg and Stock (ERS) preselect a constant close to unity, say , and subtract yt-1 from yt to obtain: = a0 + a2t  a0  a2(t  1) + et, for t = 2, , = (1  )a0 + a2[(1)t + )] + et.= a0z1t + a2z2t + etz1t = (1  ) ; z2t =  + (1)t. The important point is that the estimates a0 and a2 can be used to detrend the {yt} seriesPanel Unit Root Testsyit = ai0 + iyit–1 + ai2t + + it One way to obtain a more powerful test is to pool the estimates from a number separate series and then test the pooled value. The theory underlying the test is very simple: if you have n independent and unbiased estimates of a parameter, the mean of the estimates is also unbiased. More importantly, so long as the estimates are independent, the central limit theory suggests that the sample mean will be normally distributed around the true mean. The difficult issue is to correct for cross equation correlationBecause the lag lengths can differ across equations, you should perform separate lag length tests for each equation. Moreover, you may choose to exclude the deterministic time trend. However, if the trend is included in one equation, it should be included in all LagsEstimated it-statistic Estimated it-statistic Log of the Real RateMinus the Common Time EffectAustralia5-0.049-1.678 -0.043-1.434Canada7-0.036-1.896 -0.035-1.820France1-0.079-2.999 -0.102-3.433Germany1-0.068-2.669 -0.067-2.669Japan3-0.054-2.277 -0.048-2.137Netherlands1-0.110-3.473 -0.137-3.953U.K.1-0.081-2.759 -0.069-2.504U.S.1-0.037-1.764 -0.045-2.008Table 4.8: The Panel Unit Root Tests for Real Exchange RatesLimitationsThe null hypothesis for the IPS test is i = 2 = = n = 0. Rejection of the null hypothesis means that at least one of the i’s differs from zero. At this point, there is substantial disagreement about the asymptotic theory underlying the test. Sample size can approach infinity by increasing n for a given T, increasing T for a given n, or by simultaneously increasing n and T. For small T and large n, the critical values are dependent on the magnitudes of the various ij. The test requires that that the error terms be serially uncorrelated and contemporaneously uncorrelated. You can determine the values of pi to ensure that the autocorrelations of {it} are zero. Nevertheless, the errors may be contemporaneously correlated in that Eitjt  0The example above illustrates a common technique to correct for correlation across equations. As in the example, you can subtract a common time effect from each observation. However, there is no assurance that this correction will completely eliminate the correlation. Moreover, it is quite possible that is nonstationary. Subtracting a nonstationary component from each sequence is clearly at odds with the notion that the variables are stationary. The Beveridge-Nelson DecompositionThe trend is defined to be the conditional expectation of the limiting value of the forecast function. In lay terms, the trend is the “long-term” forecast. This forecast will differ at each period t as additional realizations of {et} become available. At any period t, the stationary component of the series is the difference between yt and the trend mt.BN 2Estimate the {yt} series using the Box–Jenkins technique. After differencing the data, an appropriately identified and estimated ARMA model will yield high-quality estimates of the coefficients. Obtain the one-step-ahead forecast errors of Etyt+s for large s. Repeating for each value of t yields the entire set of premanent componentsThe irregular component is yt minus the value of the trend. The HP FilterFor a given value of l, the goal is to select the {mt} sequence so as to minimize this sum of squares. In the minimization problem l is an arbitrary constant reflecting the “cost” or penalty of incorporating fluctuations into the trend. In applications with quarterly data, including Hodrick and Prescott (1984) l is usually set equal to 1,600.Large values of l acts to “smooth out” the trend.Let the trend of a nonstationary series be the {mt} sequence so that yt – mt the stationary component