Introduction to Time Series Regression and Forecasting

Introduction to Time Series Regression and Forecasting (SW Chapter 14) Time series data are data collected on the same observational unit at multiple time periods  Aggregate consumption and GDP for a country (for example, 20 years of quarterly observations = 80 observations)  Yen/$, pound/$ and Euro/$ exchange rates (daily data for 1 year = 365 observations)  Cigarette consumption per capita in a state, by year 14-1 Example #1 of time series data: US rate of price inflation, as measured by the quarterly percentage change in the Consumer Price Index (CPI), at an annual rate 14-2 Example #2: US rate of unemployment 14-3 Why use time series data?  To develop forecasting models o What will the rate of inflation be next year?  To estimate dynamic causal effects o If the Fed increases the Federal Funds rate now, what will be the effect on the rates of inflation and unemployment in 3 months? in 12 months? o What is the effect over time on cigarette consumption of a hike in the cigarette tax?  Or, because that is your only option … o Rates of inflation and unemployment in the US can be observed only over time! 14-4 Time series data raises new technical issues  Time lags  Correlation over time (serial correlation, a.k.a. autocorrelation)  Forecasting models built on regression methods: o autoregressive (AR) models o autoregressive distributed lag (ADL) models o need not (typically do not) have a causal interpretation  Conditions under which dynamic effects can be estimated, and how to estimate them  Calculation of standard errors when the errors are serially correlated 14-5 Using Regression Models for Forecasting (SW Section 14.1)  Forecasting and estimation of causal effects are quite different objectives.  For forecasting, o R 2 matters (a lot!) o Omitted variable bias isn’t a problem! o We will not worry about interpreting coefficients in forecasting models o External validity is paramount: the model estimated using historical data must hold into the (near) future 14-6 Introduction to Time Series Data and Serial Correlation (SW Section 14.2) First, some notation and terminology. Notation for time series data  Yt = value of Y in period t.  Data set: Y1,…,YT = T observations on the time series random variable Y  We consider only consecutive, evenly-spaced observations (for example, monthly, 1960 to 1999, no missing months) (missing and non-evenly spaced data introduce technical complications) 14-7 We will transform time series variables using lags, first differences, logarithms, & growth rates 14-8 Example: Quarterly rate of inflation at an annual rate (U.S.) CPI = Consumer Price Index (Bureau of Labor Statistics)  CPI in the first quarter of 2004 (2004:I) = 186.57  CPI in the second quarter of 2004 (2004:II) = 188.60  Percentage change in CPI, 2004:I to 2004:II  2.03   188.60  186.57  = 100    = 1.088%  = 100   186.57  186.57     Percentage change in CPI, 2004:I to 2004:II, at an annual rate = 4´1.088 = 4.359%  4.4% (percent per year)  Like interest rates, inflation rates are (as a matter of convention) reported at an annual rate.  Using the logarithmic approximation to percent changes yields 4´100´[log(188.60) – log(186.57)] = 4.329% 14-9 Example: US CPI inflation – its first lag and its change 14-10 Autocorrelation The correlation of a series with its own lagged values is called autocorrelation or serial correlation.  The first autocorrelation of Yt is corr(Yt,Yt–1)  The first autocovariance of Yt is cov(Yt,Yt–1)  Thus corr(Yt,Yt–1) = cov(Yt ,Yt 1 ) =1 var(Yt ) var(Yt 1 )  These are population correlations – they describe the population joint distribution of (Yt, Yt–1) 14-11 14-12 Sample autocorrelations The jth sample autocorrelation is an estimate of the jth population autocorrelation: ˆ j = cov(Yt , Yt  j ) var(Yt ) where 1 T cov(Yt , Yt  j ) =  (Yt  Y j 1,T )(Yt  j  Y1,T  j ) T t  j 1 where Y j 1,T is the sample average of Yt computed over observations t = j+1,…,T. NOTE: o the summation is over t=j+1 to T (why?) o The divisor is T, not T – j (this is the conventional definition used for time series data) 14-13 Example: Autocorrelations of: (1) the quarterly rate of U.S. inflation (2) the quarter-to-quarter change in the quarterly rate of inflation 14-14  The inflation rate is highly serially correlated (1 = .84)  Last quarter’s inflation rate contains much information about this quarter’s inflation rate  The plot is dominated by multiyear swings  But there are still surprise movements! 14-15 Other economic time series: 14-16 Other economic time series, ctd: 14-17 Stationarity: a key requirement for external validity of time series regression Stationarity says that history is relevant: For now, assume that Yt is stationary (we return to this later). 14-18 Autoregressions (SW Section 14.3) A natural starting point for a forecasting model is to use past values of Y (that is, Yt–1, Yt–2,…) to forecast Yt.  An autoregression is a regression model in which Yt is regressed against its own lagged values.  The number of lags used as regressors is called the order of the autoregression. o In a first order autoregression, Yt is regressed against Yt–1 o In a pth order autoregression, Yt is regressed against Yt–1,Yt–2,…,Yt–p. 14-19 The First Order Autoregressive (AR(1)) Model The population AR(1) model is Yt = 0 + 1Yt–1 + ut  0 and 1 do not have causal interpretations  if 1 = 0, Yt–1 is not useful for forecasting Yt  The AR(1) model can be estimated by OLS regression of Yt against Yt–1  Testing 1 = 0 v. 1 ¹ 0 provides a test of the hypothesis that Yt–1 is not useful for forecasting Yt 14-20 Example: AR(1) model of the change in inflation Estimated using data from 1962:I – 2004:IV: Inf t = 0.017 – 0.238Inft–1 R 2 = 0.05 (0.126) (0.096) Is the lagged change in inflation a useful predictor of the current change in inflation?  t = –.238/.096 = –2.47 > 1.96 (in absolute value)  ÞReject H0: 1 = 0 at the 5% significance level  Yes, the lagged change in inflation is a useful predictor of current change in inflation–but the R 2 is pretty low! 14-21 Example: AR(1) model of inflation – STATA First, let STATA know you are using time series data generate time=q(1959q1)+_n-1; _n is the observation no. So this command creates a new variable time that has a special quarterly date format format time %tq; Specify the quarterly date format sort time; Sort by time tsset time; Let STATA know that the variable time is the variable you want to indicate the time scale 14-22 Example: AR(1) model of inflation – STATA, ctd. . gen lcpi = log(cpi); variable cpi is already in memory . gen inf = 400*(lcpi[_n]-lcpi[_n-1]); quarterly rate of inflation at an annual rate This creates a new variable, inf, the “nth” observation of which is 400 times the difference between the nth observation on lcpi and the “n-1”th observation on lcpi, that is, the first difference of lcpi compute first 8 sample autocorrelations . corrgram inf if tin(1960q1,2004q4), noplot lags(8); LAG AC PAC Q Prob>Q ----------------------------------------1 0.8359 0.8362 127.89 0.0000 2 0.7575 0.1937 233.5 0.0000 3 0.7598 0.3206 340.34 0.0000 4 0.6699 -0.1881 423.87 0.0000 5 0.5964 -0.0013 490.45 0.0000 6 0.5592 -0.0234 549.32 0.0000 7 0.4889 -0.0480 594.59 0.0000 8 0.3898 -0.1686 623.53 0.0000 if tin(1962q1,2004q4) is STATA time series syntax for using only observations between 1962q1 and 1999q4 (inclusive). The “tin(.,.)” option requires defining the time scale first, as we did above 14-23 Example: AR(1) model of inflation – STATA, ctd . gen dinf = inf[_n]-inf[_n-1]; . reg dinf L.dinf if tin(1962q1,2004q4), r; Linear regression L.dinf is the first lag of dinf Number of obs F( 1, 170) Prob > F R-squared Root MSE = = = = = 172 6.08 0.0146 0.0564 1.6639 -----------------------------------------------------------------------------| Robust dinf | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------dinf | L1. | -.2380348 .0965034 -2.47 0.015 -.4285342 -.0475354 _cons | .0171013 .1268831 0.13 0.893 -.2333681 .2675707 -----------------------------------------------------------------------------. dis "Adjusted Rsquared = " _result(8); Adjusted Rsquared = .05082278 14-24 Forecasts: terminology and notation  Predicted values are “in-sample” (the usual definition)  Forecasts are “out-of-sample” – in the future  Notation: o YT+1|T = forecast of YT+1 based on YT,YT–1,…, using the population (true unknown) coefficients o Yˆ = forecast of YT+1 based on YT,YT–1,…, using the T 1|T estimated coefficients, which are estimated using data through period T. o For an AR(1):  YT+1|T = 0 + 1YT  YˆT 1|T = ˆ0 + ˆ1 YT, where ˆ0 and ˆ1 are estimated using data through period T. 14-25 Forecast errors The one-period ahead forecast error is, forecast error = YT+1 – YˆT 1|T The distinction between a forecast error and a residual is the same as between a forecast and a predicted value:  a residual is “in-sample”  a forecast error is “out-of-sample” – the value of YT+1 isn’t used in the estimation of the regression coefficients 14-26 Example: forecasting inflation using an AR(1) AR(1) estimated using data from 1962:I – 2004:IV: Inf t = 0.017 – 0.238Inft–1 Inf2004:III = 1.6 (units are percent, at an annual rate) Inf2004:IV = 3.5 Inf2004:IV = 3.5 – 1.6 = 1.9 The forecast of Inf2005:I is: Inf 2005:I |2000:IV = 0.017 – 0.238´1.9 = -0.44 » -0.4 so Inf 2005:I |2000:IV = Inf2004:IV + Inf 2005:I |2000:IV = 3.5 – 0.4 = 3.1% 14-27 The AR(p) model: using multiple lags for forecasting The pth order autoregressive model (AR(p)) is Yt = 0 + 1Yt–1 + 2Yt–2 + … + pYt–p + ut  The AR(p) model uses p lags of Y as regressors  The AR(1) model is a special case  The coefficients do not have a causal interpretation  To test the hypothesis that Yt–2,…,Yt–p do not further help forecast Yt, beyond Yt–1, use an F-test  Use t- or F-tests to determine the lag order p  Or, better, determine p using an “information criterion” (more on this later…) 14-28 Example: AR(4) model of inflation Inf t = .02 – .26Inft–1 – .32Inft–2 + .16Inft–3 – .03Inft–4, (.12) (.09) (.08) (.08) (.09) R 2 = 0.18  F-statistic testing lags 2, 3, 4 is 6.91 (p-value < .001)  R 2 increased from .05 to .18 by adding lags 2, 3, 4  So, lags 2, 3, 4 (jointly) help to predict the change in inflation, above and beyond the first lag – both in a statistical sense (are statistically significant) and in a substantive sense (substantial increase in the R 2 ) 14-29 Example: AR(4) model of inflation – STATA . reg dinf L(1/4).dinf if tin(1962q1,2004q4), r; Linear regression Number of obs F( 4, 167) Prob > F R-squared Root MSE = = = = = 172 7.93 0.0000 0.2038 1.5421 -----------------------------------------------------------------------------| Robust dinf | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------dinf | L1. | -.2579205 .0925955 -2.79 0.006 -.4407291 -.0751119 L2. | -.3220302 .0805456 -4.00 0.000 -.481049 -.1630113 L3. | .1576116 .0841023 1.87 0.063 -.0084292 .3236523 L4. | -.0302685 .0930452 -0.33 0.745 -.2139649 .1534278 _cons | .0224294 .1176329 0.19 0.849 -.2098098 .2546685 -----------------------------------------------------------------------------NOTES  L(1/4).dinf is A convenient way to say “use lags 1–4 of dinf as regressors”  L1,…,L4 refer to the first, second,… 4th lags of dinf 14-30 Example: AR(4) model of inflation – STATA, ctd. . dis "Adjusted Rsquared = " _result(8); Adjusted Rsquared = .18474733 . test L2.dinf L3.dinf L4.dinf; ( 1) ( 2) ( 3) result(8) is the rbar-squared of the most recently run regression L2.dinf is the second lag of dinf, etc. L2.dinf = 0.0 L3.dinf = 0.0 L4.dinf = 0.0 F( 3, 147) = Prob > F = 6.71 0.0003 14-31 Digression: we used Inf, not Inf, in the AR’s. Why? The AR(1) model of Inft–1 is an AR(2) model of Inft: Inft = 0 + 1Inft–1 + ut or Inft – Inft–1 = 0 + 1(Inft–1 – Inft–2) + ut or Inft = Inft–1 + 0 + 1Inft–1 – 1Inft–2 + ut = 0 + (1+1)Inft–1 – 1Inft–2 + ut 14-32 So why use Inft, not Inft? AR(1) model of Inf: AR(2) model of Inf: Inft = 0 + 1Inft–1 + ut Inft = 0 + 1Inft + 2Inft–1 + vt  When Yt is strongly serially correlated, the OLS estimator of the AR coefficient is biased towards zero.  In the extreme case that the AR coefficient = 1, Yt isn’t stationary: the ut’s accumulate and Yt blows up.  If Yt isn’t stationary, our regression theory are working with here breaks down  Here, Inft is strongly serially correlated – so to keep ourselves in a framework we understand, the regressions are specified using Inf  More on this later… 14-33 Time Series Regression with Additional Predictors and the Autoregressive Distributed Lag (ADL) Model (SW Section 14.4)  So far we have considered forecasting models that use only past values of Y  It makes sense to add other variables (X) that might be useful predictors of Y, above and beyond the predictive value of lagged values of Y: Yt = 0 + 1Yt–1 + … + pYt–p + 1Xt–1 + … + rXt–r + ut  This is an autoregressive distributed lag model with p lags of Y and r lags of X … ADL(p,r). 14-34 Example: inflation and unemployment According to the “Phillips curve,” if unemployment is above its equilibrium, or “natural,” rate, then the rate of inflation will increase. That is, Inft is related to lagged values of the unemployment rate, with a negative coefficient  The rate of unemployment at which inflation neither increases nor decreases is often called the “non-accelerating rate of inflation” unemployment rate (the NAIRU).  Is the Phillips curve found in US economic data?  Can it be exploited for forecasting inflation?  Has the U.S. Phillips curve been stable over time? 14-35 The empirical U.S. “Phillips Curve,” 1962 – 2004 (annual) One definition of the NAIRU is that it is the value of u for which Inf = 0 – the x intercept of the regression line. 14-36 The empirical (backwards-looking) Phillips Curve, ctd. ADL(4,4) model of inflation (1962 – 2004): Inf t = 1.30 – .42Inft–1 – .37Inft–2 + .06Inft–3 – .04Inft–4 (.44) (.08) (.09) (.08) (.08) – 2.64Unemt–1 + 3.04Unemt–2 – 0.38Unemt–3 + .25Unempt–4 (.46) (.86) (.89) (.45)  R 2 = 0.34 – a big improvement over the AR(4), for which R 2 = .18 14-37 Example: dinf and unem – STATA . reg dinf L(1/4).dinf L(1/4).unem if tin(1962q1,2004q4), r; Linear regression Number of obs F( 8, 163) Prob > F R-squared Root MSE = = = = = 172 8.95 0.0000 0.3663 1.3926 -----------------------------------------------------------------------------| Robust dinf | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------dinf | L1. | -.4198002 .0886973 -4.73 0.000 -.5949441 -.2446564 L2. | -.3666267 .0940369 -3.90 0.000 -.5523143 -.1809391 L3. | .0565723 .0847966 0.67 0.506 -.1108691 .2240138 L4. | -.0364739 .0835277 -0.44 0.663 -.2014098 .128462 unem | L1. | -2.635548 .4748106 -5.55 0.000 -3.573121 -1.697975 L2. | 3.043123 .8797389 3.46 0.001 1.305969 4.780277 L3. | -.3774696 .9116437 -0.41 0.679 -2.177624 1.422685 L4. | -.2483774 .4605021 -0.54 0.590 -1.157696 .6609413 _cons | 1.304271 .4515941 2.89 0.004 .4125424 2.196 -----------------------------------------------------------------------------14-38 Example: ADL(4,4) model of inflation – STATA, ctd. . dis "Adjusted Rsquared = " _result(8); Adjusted Rsquared = .33516905 . ( ( ( ( test L1.unem L2.unem L3.unem L4.unem; 1) 2) 3) 4) L.unem = 0 L2.unem = 0 L3.unem = 0 L4.unem = 0 F( 4, 163) = Prob > F = 8.44 0.0000 The lags of unem are significant The null hypothesis that the coefficients on the lags of the unemployment rate are all zero is rejected at the 1% significance level using the F-statistic 14-39 The test of the joint hypothesis that none of the X’s is a useful predictor, above and beyond lagged values of Y, is called a Granger causality test “causality” is an unfortunate term here: Granger Causality simply refers to (marginal) predictive content. 14-40 Forecast uncertainty and forecast intervals Why do you need a measure of forecast uncertainty?  To construct forecast intervals  To let users of your forecast (including yourself) know what degree of accuracy to expect Consider the forecast YˆT 1|T = ˆ0 + ˆ1 YT + ˆ1 XT The forecast error is: YT+1 – YˆT 1|T = uT+1 – [( ˆ0 – 0) + ( ˆ1 – 1)YT + ( ˆ1 – 2)XT] 14-41 The mean squared forecast error (MSFE) is, E(YT+1 – YˆT 1|T )2 = E(uT+1)2 + + E[( ˆ0 – 0) + ( ˆ1 – 1)YT + ( ˆ1 – 2)XT]2  MSFE = var(uT+1) + uncertainty arising because of estimation error  If the sample size is large, the part from the estimation error is (much) smaller than var(uT+1), in which case MSFE » var(uT+1)  The root mean squared forecast error (RMSFE) is the square root of the MS forecast error: RMSFE = E [(YT 1  YˆT 1|T ) 2 ] 14-42 The root mean squared forecast error (RMSFE) RMSFE = E [(YT 1  YˆT 1|T ) 2 ]  The RMSFE is a measure of the spread of the forecast error distribution.  The RMSFE is like the standard deviation of ut, except that it explicitly focuses on the forecast error using estimated coefficients, not using the population regression line.  The RMSFE is a measure of the magnitude of a typical forecasting “mistake” 14-43 Three ways to estimate the RMSFE 1. Use the approximation RMSFE » u, so estimate the RMSFE by the SER. 2. Use an actual forecast history for t = t1,…, T, then estimate by T 1 1 2 ˆ MSFE = ( ) Y  Y  t 1 t 1|t T  t1  1 t t1 1 Usually, this isn’t practical – it requires having an historical record of actual forecasts from your model 3. Use a simulated forecast history, that is, simulate the forecasts you would have made using your model in real time….then use method 2, with these pseudo out-ofsample forecasts… 14-44 The method of pseudo out-of-sample forecasting  Re-estimate your model every period, t = t1–1,…,T–1  Compute your “forecast” for date t+1 using the model estimated through t  Compute your pseudo out-of-sample forecast at date t, using the model estimated through t–1. This is Yˆ . t 1|t  Compute the poos forecast error, Yt+1 – Yˆt 1|t  Plug this forecast error into the MSFE formula, T 1 1 2 ˆ MSFE = Y  Y ( )  t 1 t 1|t T  t1  1 t t1 1 Why the term “pseudo out-of-sample forecasts”? 14-45 Using the RMSFE to construct forecast intervals If uT+1 is normally distributed, then a 95% forecast interval can be constructed as YˆT |T 1 ± 1.96´ RMSFE Note: 1. A 95% forecast interval is not a confidence interval (YT+1 isn’t a nonrandom coefficient, it is random!) 2. This interval is only valid if uT+1 is normal – but still might be a reasonable approximation and is a commonly used measure of forecast uncertainty 3. Often “67%” forecast intervals are used:  RMSFE 14-46 Example #1: the Bank of England “Fan Chart”, 11/05 14-47 Example #2: Monthly Bulletin of the European Central Bank, Dec. 2005, Staff macroeconomic projections Precisely how, did they compute these intervals? 14-48 Example #3: Fed, Semiannual Report to Congress, 7/04 Economic projections for 2004 and 2005 Federal Reserve Governors and Reserve Bank presidents Indicator Range Central tendency 2005 Change, fourth quarter to fourth quarter Nominal GDP Real GDP PCE price index excl food and energy Average level, fourth quarter Civilian unemployment rate 4-3/4 to 6-1/2 3-1/2 to 4 1-1/2 to 2-1/2 5-1/4 to 6 3-1/2 to 4 1-1/2 to 2 5 to 5-1/2 5 to 5-1/4 How did they compute these intervals? 14-49 Lag Length Selection Using Information Criteria (SW Section 14.5) How to choose the number of lags p in an AR(p)?  Omitted variable bias is irrelevant for forecasting  You can use sequential “downward” t- or F-tests; but the models chosen tend to be “too large” (why?)  Another – better – way to determine lag lengths is to use an information criterion  Information criteria trade off bias (too few lags) vs. variance (too many lags)  Two IC are the Bayes (BIC) and Akaike (AIC)… 14-50 The Bayes Information Criterion (BIC) ln T  SSR ( p )  BIC(p) = ln    ( p  1) T  T   First term: always decreasing in p (larger p, better fit)  Second term: always increasing in p. o The variance of the forecast due to estimation error increases with p – so you don’t want a forecasting model with too many coefficients – but what is “too many”? o This term is a “penalty” for using more parameters – and thus increasing the forecast variance.  Minimizing BIC(p) trades off bias and variance to determine a “best” value of p for your forecast. p o The result is that pˆ BIC  p! (SW, App. 14.5) 14-51 Another information criterion: Akaike Information Criterion (AIC) 2  SSR ( p )  AIC(p) = ln    ( p  1) T  T  ln T  SSR ( p )  BIC(p) = ln    ( p  1) T  T  The penalty term is smaller for AIC than BIC (2 < lnT) o AIC estimates more lags (larger p) than the BIC o This might be desirable if you think longer lags might be important. o However, the AIC estimator of p isn’t consistent – it can overestimate p – the penalty isn’t big enough 14-52 Example: AR model of inflation, lags 0 – 6: # Lags 0 1 2 3 4 5 6 BIC 1.095 1.067 0.955 0.957 0.986 1.016 1.046 AIC 1.076 1.030 0.900 0.884 0.895 0.906 0.918 R2 0.000 0.056 0.181 0.203 0.204 0.204 0.204  BIC chooses 2 lags, AIC chooses 3 lags.  If you used the R2 to enough digits, you would (always) select the largest possible number of lags. 14-53 Generalization of BIC to multivariate (ADL) models Let K = the total number of coefficients in the model (intercept, lags of Y, lags of X). The BIC is, ln T  SSR ( K )  BIC(K) = ln  K T  T   Can compute this over all possible combinations of lags of Y and lags of X (but this is a lot)!  In practice you might choose lags of Y by BIC, and decide whether or not to include X using a Granger causality test with a fixed number of lags (number depends on the data and application) 14-54 Nonstationarity I: Trends (SW Section 14.6)  So far, we have assumed that the data are well-behaved – technically, that the data are stationary.  Now we will discuss two of the most important ways that, in practice, data can be nonstationary (that is, deviate from stationarity). You need to be able to recognize/detect nonstationarity, and to deal with it when it occurs.  Two important types of nonstationarity are: o Trends (SW Section 14.6) o Structural breaks (model instability) (SW Section 14.7)  Up now: trends 14-55 Outline of discussion of trends in time series data: 1. What is a trend? 2. What problems are caused by trends? 3. How do you detect trends (statistical tests)? 4. How to address/mitigate problems raised by trends 14-56 1. What is a trend?  A trend is a long-term movement or tendency in the data.  Trends need not be just a straight line!  Which of these series has a trend? 14-57 14-58 14-59 What is a trend, ctd. The three series:  log Japan GDP clearly has a long-run trend – not a straight line, but a slowly decreasing trend – fast growth during the 1960s and 1970s, slower during the 1980s, stagnating during the 1990s/2000s.  Inflation has long-term swings, periods in which it is persistently high for many years (70s’/early 80s) and periods in which it is persistently low. Maybe it has a trend – hard to tell.  NYSE daily changes has no apparent trend. There are periods of persistently high volatility – but this isn’t a trend. 14-60 Deterministic and stochastic trends  A trend is a long-term movement or tendency in the data.  A deterministic trend is a nonrandom function of time (e.g. yt = t, or yt = t2).  A stochastic trend is random and varies over time  An important example of a stochastic trend is a random walk: Yt = Yt–1 + ut, where ut is serially uncorrelated If Yt follows a random walk, then the value of Y tomorrow is the value of Y today, plus an unpredictable disturbance. 14-61 Deterministic and stochastic trends, ctd. Two key features of a random walk: (i) YT+h|T = YT  Your best prediction of the value of Y in the future is the value of Y today  To a first approximation, log stock prices follow a random walk (more precisely, stock returns are unpredictable (ii) var(YT+h|T – YT) = h u2  The variance of your forecast error increases linearly in the horizon. The more distant your forecast, the greater the forecast uncertainty. (Technically this is the sense in which the series is “nonstationary”) 14-62 Deterministic and stochastic trends, ctd. A random walk with drift is Yt = 0 +Yt–1 + ut, where ut is serially uncorrelated The “drift” is 0: If 0  0, then Yt follows a random walk around a linear trend. You can see this by considering the hstep ahead forecast: YT+h|T = 0h + YT The random walk model (with or without drift) is a good description of stochastic trends in many economic time series. 14-63 Deterministic and stochastic trends, ctd. Where we are headed is the following practical advice: If Yt has a random walk trend, then Yt is stationary and regression analysis should be undertaken using Yt instead of Yt. Upcoming specifics that lead to this advice:  Relation between the random walk model and AR(1), AR(2), AR(p) (“unit autoregressive root”)  A regression test for detecting a random walk trend arises naturally from this development 14-64 Stochastic trends and unit autoregressive roots Random walk (with drift): Yt = 0 + Yt–1 + ut AR(1): Yt = 0 + 1Yt–1 + ut  The random walk is an AR(1) with 1 = 1.  The special case of 1 = 1 is called a unit root*.  When 1 = 1, the AR(1) model becomes Yt = 0 + ut *This terminology comes from considering the equation 1 – 1z = 0 – the “root” of this equation is z = 1/1, which equals one (unity) if 1 = 1. 14-65 Unit roots in an AR(2) AR(2): Yt = 0 + 1Yt–1 + 2Yt–2 + ut Use the “rearrange the regression” trick from Ch 7.3: Yt = 0 + 1Yt–1 + 2Yt–2 + ut = 0 + (1+2)Yt–1 – 2Yt–1 + 2Yt–2 + ut = 0 + (1+2)Yt–1 – 2(Yt–1 – Yt–2) + ut Subtract Yt–1 from both sides: Yt – Yt–1 = 0 + (1+2–1)Yt–1 – 2(Yt–1 – Yt–2) + ut or Yt = 0 + Yt–1 + 1Yt–1 + ut, where  = 1 + 2 – 1 and 1 = –2.. 14-66 Unit roots in an AR(2), ctd. Thus the AR(2) model can be rearranged as, Yt = 0 + Yt–1 + 1Yt–1 + ut where  = 1 + 2 – 1 and 1 = –2. Claim: if 1 – 1z – 2z2 = 0 has a unit root, then 1 + 2 = 1 (you can show this yourself!) Thus, if there is a unit root, then  = 0 and the AR(2) model becomes, Yt = 0 + 1Yt–1 + ut If an AR(2) model has a unit root, then it can be written as an AR(1) in first differences. 14-67 Unit roots in the AR(p) model AR(p): Yt = 0 + 1Yt–1 + 2Yt–2 + … + pYt–p + ut This regression can be rearranged as, Yt = 0 + Yt–1 + 1Yt–1 + 2Yt–2 + … + p–1Yt–p+1 + ut where  =  1 + 2 + … + p – 1 1 = –(2 +… + p) 2 = –(3 +… + p) … p–1 = –p 14-68 Unit roots in the AR(p) model, ctd. The AR(p) model can be written as, Yt = 0 + Yt–1 + 1Yt–1 + 2Yt–2 + … + p–1Yt–p+1 + ut where  = 1 + 2 + … + p – 1. Claim: If there is a unit root in the AR(p) model, then  = 0 and the AR(p) model becomes an AR(p–1) model in first differences: Yt = 0 + 1Yt–1 + 2Yt–2 + … + p–1Yt–p+1 + ut 14-69 2. What problems are caused by trends? There are three main problems with stochastic trends: 1. AR coefficients can be badly biased towards zero. This means that if you estimate an AR and make forecasts, if there is a unit root then your forecasts can be poor (AR coefficients biased towards zero) 2. Some t-statistics don’t have a standard normal distribution, even in large samples (more on this later) 3. If Y and X both have random walk trends then they can look related even if they are not – you can get “spurious regressions.” Here is an example… 14-70 Log Japan gdp (smooth line) and US inflation (both rescaled), 1965-1981 lgdpjs infs 4 2 0 -2 1965q1 1970q1 1975q1 time 1980q1 1985q1 14-71 Log Japan gdp (smooth line) and US inflation (both rescaled), 1982-1999 lgdpjs infs 4 2 0 -2 -4 1980q1 1985q1 1990q1 time 1995q1 2000q1 14-72 3. How do you detect trends? 1. Plot the data (think of the three examples we started with). 2. There is a regression-based test for a random walk – the Dickey-Fuller test for a unit root. The Dickey-Fuller test in an AR(1) Yt = 0 + 1Yt–1 + ut or Yt = 0 + Yt–1 + ut H0:  = 0 (that is, 1 = 1) v. H1:  < 0 (note: this is 1-sided:  < 0 means that Yt is stationary) 14-73 DF test in AR(1), ctd. Yt = 0 + Yt–1 + ut H0:  = 0 (that is, 1 = 1) v. H1:  < 0 Test: compute the t-statistic testing  = 0  Under H0, this t statistic does not have a normal distribution!!  You need to compare the t-statistic to the table of DickeyFuller critical values. There are two cases: (a) Yt = 0 + Yt–1 + ut (intercept only) (b) Yt = 0 + t + Yt–1 + ut (intercept & time trend)  The two cases have different critical values! 14-74 Table of DF critical values (a) Yt = 0 + Yt–1 + ut (intercept only) (b) Yt = 0 + t + Yt–1 + ut (intercept and time trend) Reject if the DF t-statistic (the t-statistic testing  = 0) is less than the specified critical value. This is a 1-sided test of the null hypothesis of a unit root (random walk trend) vs. the alternative that the autoregression is stationary. 14-75 The Dickey-Fuller test in an AR(p) In an AR(p), the DF test is based on the rewritten model, Yt = 0 + Yt–1 + 1Yt–1 + 2Yt–2 + … + p–1Yt–p+1 + ut (*) where  = 1 + 2 + … + p – 1. If there is a unit root (random walk trend),  = 0; if the AR is stationary,  < 1. The DF test in an AR(p) (intercept only): 1. Estimate (*), obtain the t-statistic testing  = 0 2. Reject the null hypothesis of a unit root if the t-statistic is less than the DF critical value in Table 14.5 Modification for time trend: include t as a regressor in (*) 14-76 When should you include a time trend in the DF test? The decision to use the intercept-only DF test or the intercept & trend DF test depends on what the alternative is – and what the data look like.  In the intercept-only specification, the alternative is that Y is stationary around a constant  In the intercept & trend specification, the alternative is that Y is stationary around a linear time trend. 14-77 Example: Does U.S. inflation have a unit root? The alternative is that inflation is stationary around a constant 14-78 Example: Does U.S. inflation have a unit root? ctd DF test for a unit root in U.S. inflation – using p = 4 lags . reg dinf L.inf L(1/4).dinf if tin(1962q1,2004q4); Source | SS df MS -------------+-----------------------------Model | 118.197526 5 23.6395052 Residual | 380.599255 166 2.2927666 -------------+-----------------------------Total | 498.796781 171 2.91694024 Number of obs F( 5, 166) Prob > F R-squared Adj R-squared Root MSE = = = = = = 172 10.31 0.0000 0.2370 0.2140 1.5142 -----------------------------------------------------------------------------dinf | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------inf | L1. | -.1134149 .0422339 -2.69 0.008 -.1967998 -.03003 dinf | L1. | -.1864226 .0805141 -2.32 0.022 -.3453864 -.0274589 L2. | -.256388 .0814624 -3.15 0.002 -.417224 -.0955519 L3. | .199051 .0793508 2.51 0.013 .0423842 .3557178 L4. | .0099822 .0779921 0.13 0.898 -.144002 .1639665 _cons | .5068071 .214178 2.37 0.019 .0839431 .929671 ------------------------------------------------------------------------------ DF t-statstic = –2.69 Don’t compare this to –1.645 – use the Dickey-Fuller table! 14-79 DF t-statstic = –2.69 (intercept-only): t = –2.69 rejects a unit root at 10% level but not the 5% level  Some evidence of a unit root – not clear cut.  This is a topic of debate – what does it mean for inflation to have a unit root?  We model inflation as having a unit root. Note: you can choose the lag length in the DF regression by BIC or AIC (for inflation, both reject at 10%, not 5% level) 14-80 4. How to address and mitigate problems raised by trends If Yt has a unit root (has a random walk stochastic trend), the easiest way to avoid the problems this poses is to model Yt in first differences.  In the AR case, this means specifying the AR using first differences of Yt (Yt)  This is what we did in our initial treatment of inflation – the reason was that inspection of the plot of inflation, plus the DF test results, suggest that inflation plausibly has a unit root – so we estimated the ARs using Inft 14-81 Summary: detecting and addressing stochastic trends 1. The random walk model is the workhorse model for trends in economic time series data 2. To determine whether Yt has a stochastic trend, first plot Yt, then if a trend looks plausible, compute the DF test (decide which version, intercept or intercept+trend) 3. If the DF test fails to reject, conclude that Yt has a unit root (random walk stochastic trend) 4. If Yt has a unit root, use Yt for regression analysis and forecasting. 14-82 Nonstationarity II: Breaks and Model Stability (SW Section 14.7) The second type of nonstationarity we consider is that the coefficients of the model might not be constant over the full sample. Clearly, it is a problem for forecasting if the model describing the historical data differs from the current model – you want the current model for your forecasts! (This is an issue of external validity.) So we will:  Go over two ways to detect changes in coefficients: tests for a break, and pseudo out-of-sample forecast analysis  Work through an example: the U.S. Phillips curve 14-83 A. Tests for a break (change) in regression coefficients Case I: The break date is known Suppose the break is known to have occurred at date . Stability of the coefficients can be tested by estimating a fully interacted regression model. In the ADL(1,1) case: Yt = 0 + 1Yt–1 + 1Xt–1 + 0Dt() + 1[Dt()´Yt–1] + 2[Dt()´Xt–1] + ut where Dt() = 1 if t  , and = 0 otherwise.  If 0 = 1 = 2 = 0, then the coefficients are constant over the full sample.  If at least one of 0, 1, or 2 are nonzero, the regression function changes at date . 14-84 Yt = 0 + 1Yt–1 + 1Xt–1 + 0Dt() + 1[Dt()´Yt–1] + 2[Dt()´Xt–1] + ut where Dt() = 1 if t  , and = 0 otherwise The Chow test statistic for a break at date  is the (heteroskedasticity-robust) F-statistic that tests: H 0 : 0 = 1 = 2 = 0 vs. H1: at least one of 0, 1, or 2 are nonzero  Note that you can apply this to a subset of the coefficients, e.g. only the coefficient on Xt–1.  Often, however, you don’t have a candidate break date… 14-85 Case II: The break date is unknown Why consider this case?  You might suspect there is a break, but not know when  You might want to test for stationarity (coefficient stability) against a general alternative that there has been a break sometime.  Often, even if you think you know the break date, that “knowledge” is based on prior inspection of the series – so that in effect you “estimated” the break date. This invalidates the Chow test critical values (why?) 14-86 The Quandt Likelihod Ratio (QLR) Statistic (also called the “sup-Wald” statistic) The QLR statistic = the maximal Chow statistics  Let F() = the Chow test statistic testing the hypothesis of no break at date .  The QLR test statistic is the maximum of all the Chow Fstatistics, over a range of , 0    1: QLR = max[F(0), F(0+1) ,…, F(1–1), F(1)]  A conventional choice for 0 and 1 are the inner 70% of the sample (exclude the first and last 15%.  Should you use the usual Fq,¥ critical values? 14-87 The QLR test, ctd. QLR = max[F(0), F(0+1) ,…, F(1–1), F(1)]  The large-sample null distribution of F() for a given (fixed, not estimated)  is Fq,¥  But if you get to compute two Chow tests and choose the biggest one, the critical value must be larger than the critical value for a single Chow test.  If you compute very many Chow test statistics – for example, all dates in the central 70% of the sample – the critical value must be larger still! 14-88  Get this: in large samples, QLR has the distribution,  1 q Bi ( s ) 2  , max a  s1a     q i 1 s(1  s )  where {Bi}, i =1,…,n, are independent continuous-time “Brownian Bridges” on 0 ≤ s ≤ 1 (a Brownian Bridge is a Brownian motion deviated from its mean), and where a = .15 (exclude first and last 15% of the sample)  Critical values are tabulated in SW Table 14.6… 14-89 Note that these critical values are larger than the Fq, critical values – for example, F1, 5% critical value is 3.84. 14-90 Has the postwar U.S. Phillips Curve been stable? Recall the ADL(4,4) model of Inft and Unempt – the empirical backwards-looking Phillips curve, estimated over (1962 – 2004): Inf t = 1.30 – .42Inft–1 – .37Inft–2 + .06Inft–3 – .04Inft–4 (.44) (.08) (.09) (.08) (.08) – 2.64Unemt–1 + 3.04Unemt–2 – 0.38Unemt–3 + .25Unempt–4 (.46) (.86) (.89) (.45) Has this model been stable over the full period 1962-2004? 14-91 QLR tests of the stability of the U.S. Phillips curve. dependent variable: Inft regressors: intercept, Inft–1,…, Inft–4, Unempt–1,…, Unempt–4  test for constancy of intercept only (other coefficients are assumed constant): QLR = 2.865 (q = 1). o 10% critical value = 7.12  don’t reject at 10% level  test for constancy of intercept and coefficients on Unempt,…, Unempt–3 (coefficients on Inft–1,…, Inft–4 are constant): QLR = 5.158 (q = 5) o 1% critical value = 4.53  reject at 1% level o Break date estimate: maximal F occurs in 1981:IV  Conclude that there is a break in the inflation – unemployment relation, with estimated date of 1981:IV 14-92 14-93 B. Assessing Model Stability using Pseudo Out-of-Sample Forecasts  The QLR test does not work well towards the very end of the sample – but this is usually the most interesting part – it is the most recent history and you want to know if the forecasting model still works in the very recent past.  One way to check whether the model is working is to see whether the pseudo out-of-sample (poos) forecasts are “on track” in the most recent observations. Because the focus is on only the most recent observations, this is an informal diagnostic (not a formal test) – this complements formal testing using the QLR. 14-94 Application to the U.S. Phillips Curve: Was the U.S. Phillips Curve stable towards the end of the sample?  We found a break in 1981:IV – so for this analysis, we only consider regressions that start in 1982:I – ignore the earlier data from the “old” model (“regime”).  Regression model: dependent variable: Inft regressors: 1, Inft–1,…, Inft–4, Unempt–1,…, Unempt–4  Pseudo out-of-sample forecasts: o Compute regression over t = 1982:I,…, P o Compute poos forecast, Inf P 1|P , and forecast error o Repeat for P = 1994:I,…, 2005:I 14-95 POOS forecasts of Inf using ADL(4,4) model with Unemp There are some big forecast errors (in 2001) but they do not appear to be getting bigger – the model isn’t deteriorating 14-96 poos forecasts using the Phillips curve, ctd. Some summary statistics:  Mean forecast error, 1999:I – 2004:IV = 0.11 (SE = 0.27) o No evidence that the forecasts are systematically to high or too low  poos RMSFE, 1999:I – 2004:IV: 1.32  SER, model fit 1982:I – 1998:IV: 1.30 o The poos RMSFE  the in-sample SER – another indication that forecasts are not doing any worse (or better) out of sample than in-sample This analysis suggests that there was not a substantial change in the forecasts produced by the ADL(4,4) model towards the end of the sample 14-97 Summary: Time Series Forecasting Models (SW Section 14.8)  For forecasting purposes, it isn’t important to have coefficients with a causal interpretation!  The tools of regression can be used to construct reliable forecasting models – even though there is no causal interpretation of the coefficients: o AR(p) – common “benchmark” models o ADL(p,q) – add q lags of X (another predictor) o Granger causality tests – test whether a variable X and its lags are useful for predicting Y given lags of Y. 14-98 Summary, ctd.  New ideas and tools: o stationarity o forecast intervals using the RMSFE o pseudo out-of-sample forecasting o BIC for model selection o Ways to check/test for for nonstationarity:  Dickey-Fuller test for a unit root (stochastic trend)  Test for a break in regression coefficients:  Chow test at a known date  QLR test at an unknown date  poos analysis for end-of-sample forecast breakdown 14-99