International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 337 Dynamic Time Series Regression: A Panacea for Spurious Correlations Emmanuel Alphonsus Akpan*, Imoh Udo Moffat** * Department of Mathematics and Statistics, University of Uyo, Nigeria Department of Mathematics and Statistics, University of Uyo, Nigeria ** Abstract- The study examined that the linear relationship between Gross Domestic Product ( ) and Money Supply ( ) from 1981 to 2014 is spurious and could be avoided by dynamic regression modeling. The fact that spurious regression always results in misleading correlations between two time series was a big motivation for undertaking this study. Therefore, exploring data from the Central Bank of Nigeria Statistical Bulletin, we found that the linear relationship between the dependent variable ( ) and the independent variable ( ) seemed spurious as the errors of the regression model were found to be autocorrelated. In a bid to correct this problem of spurious regression, we identified lags 0, -1 and -2 of as predictors of using cross correlation function. Hence, the dynamic regression of current lag and past lags 1, 2 of as predictors of revealed that the errors are uncorrelated and the coefficient of determination is as low as 0.2086, indicating that and are totally unrelated. Index Terms- autocorrelated errors; cross correlation function; dynamic regression; prewhitening; spurious regression. I. INTRODUCTION I t is always common to see in literature that most regression models are reported with very high coefficient of determination, which indicates that the model is a good fit without considering the problem posed by autocorrelated errors (Granger and Newbold, 1974). Though in classical regression model, the error terms are assumed to be a white noise, that is, a sequence of independent and identically distributed random variables. In practice, however, the error terms always appear to be autocorrelated (Wei, 2006; Box, Jenkins and Reinsel, 2008). But if the autocorrelated error terms are ignored, the regression becomes spurious resulting in misleading correlations, a situation where a significant regression can be achieved for totally unrelated series (Pankratz, 1991; Wei, 2006; Cryer and Chan, 2008). The problem associated with spurious regressions can be avoided by including lagged values of time independent variables in the regression model (Pankratz, 1991; Brockwell and Davis, 2002; Fuller, 1996; Wei, 2006; and, Box, Jenkins and Reinsel, 2008). Different authors use different names to refer to regression models in which the current value of time dependent variable is a function of current and lagged values of time independent variables. For instance, Fuller (1996) referred to such regression models as transfer function models while Pankratz, (1991) called it dynamic regression models. However, prior studies revealed that differencing both the time dependent variable and time independent variables is one approach that spurious regressions could be avoided but failed to take into consideration lagged values of time independent variables, thereby, creating a gap in knowledge that the dependent variable may be related to independent variables with time lags which often results in loss of useful information about the roles of lagged values in explaining movements in the dependent variable. Thus, this paper contributes towards filling the gap by analyzing the relationship between Gross Domestic Product ( ) – dependent variable and Money Supply ( ) – independent variable. II. METHODOLOGY Regression Model Rawlings, Pantula and Dickey (1998) defined a standard regression model as = + , , + , +⋯ + , + � (2.1) where = dependent variable � = regression parameters, i = 1,…, k � = independent variables, i = 1,…, k � = error term assumed to be i.i.d. N(0, � ) Thus, the dependent variable for a time series regression model with independent variables is a linear combination of independent variables measured in the same time frame as the dependent variable. Estimates of the parameters of the model in (2.1) can be obtained by Least Squares Estimation Method (see for example Drasper and Smith, 1998; Rawlings, Pantula and Dickey, 1998). Dynamic Regression Model Dynamic regression model as specified by Pankratz (1991) is as follows: = + + − +⋯ + − + � (2.2) The intuition is that equation (2.2) is built to take into account useful information about the roles of time (past) lag ( − , … , − ) in explaining the movements in , which is not possible with equation (2.1). The parameters of dynamic regression models are estimated using maximum likelihood method, see Pankratz (1991) for details. Autoregressive Moving Average (ARMA) Processes A natural extension of pure autoregressive and pure moving average processes is the mixed autoregressive moving average www.ijsrp.org International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 � processes, which includes the autoregressive and moving average as special cases (Wei, 2006). � , process if { } A stochastic process { } is an is stationary and if for every �, =� � (2.3) = − − −⋯ − is the autoregressive coefficient polynomial. � = − � −� −⋯ −� is the moving average coefficient polynomial. Box, Jenkins and Reinsel (2008) considered the extension of ARMA model in (2.3) to deal with homogenous non-stationary time series in which , is non-stationary but its � ℎ difference is a stationary ARMA model. Denoting the � ℎ difference of by =� ∇� = � � (2.4) where is the nonstationary autoregressive operator such that d of the roots of = are unity and the remainder lie outside the unit circle while � is a stationary autoregressive operator. Thus, (2.4) is called an autoregressive integrated moving average model and can be referred to as an � , �, model. Prewhitening and Cross Correlation Function (CCF) According to Wei (2002), assuming that independent variable, follows an ARMA(p,q) process, = �� ( �( (2.5) where is white noise. The series − = �� �( (2.6) is called the prewhitened series. Applying the same prewhitening transformation to the dependent variable, we obtain a filtered independent series, − = �� �( (2.7) Let Y = { } be time dependent variable, X = { } be time independent variable, and the cross covariance = � , ) for each pair of integers t and function , , s. The cross correlation between X and Y at lag k can be defined � , = , − ) = by . In general, the √� � cross correlation function is not an even function since , − ) need not equal , + ). Moreover, the sample cross correlation function (CCF) is useful for identifying lags of independent variable that might be useful predictors of dependent variable (Cryer and Chan, 2008). However, the CCF can be obtained by prewhitening by considering a more general regression model relating X to Y, = ∑∞ −∞ − + � (2.8) where X is independent of �. Applying the filter to both sides of (2.8), we have ̃ = ∑∞ ̃ − + �̃ −∞ (2.9) where �̃ = � − � − − � − − ⋯ 338 The prewhitening procedure thus orthogonalizes the various lags of X in the original regression model (Cryer and Chan, 2008). Model Selection Criteria For a given data set, when there are multiple adequate models, the selection criterion is normally based on summary statistics from residuals of a fitted model (Wei, 2006). There are several model selection criteria based on residuals (see Wei, 2006). For the purpose of this study, we consider the well-known Akaike’s information criterion (AIC), (Akaike, 1973) defined as AIC = − likelihood + number of parameters where the likelihood function is evaluated at the maximum likelihood estimates. The optimal order of the model is chosen by the value of the number of parameters, so that AIC is minimum (Wei, 2006). Model Diagnostic Checking Box and Pierce (1970) proposed the Portmanteau statistics: ∗ (m) T ∑ ̂ = = (2.10) where T is the number of observations. Ljung and Box (1978) modify the ∗ (m) statistic to increase the power of the test in finite samples as follows: Q(m) = T(T + (2.11) where T is the number of observations. 2) ∑ = ̂�2 � �− The decision rule is to reject if Q(m) > � , where � denotes the 100 (1 – )th percentile of a Chi-squared distribution with m – (p + q) degree of freedom (see for example Akpan, Moffat and Ekpo, 2016). III. DATA ANALYSIS AND DISCUSSION This study considers the Gross Domestic Product (N’ Billion) as the dependent variable and the Money Supply (N’ Billion) as the independent variable. The data were obtained from the Central Bank of Nigeria Statistical Bulletin for a period spanning from 1981 to 2014. Each series consists of 34 observations. First, we regress on , and obtain the estimated regression model presented in equation (3.1) below: = . + . s.e (294.2920) (0.5383) t-value ( . (7.564) (3.1) p-value (0.85) (1.98e-09) = 0.5709 [Excerpts from Table 1]. From the fitted model in (3.1), it is observed that the inclusion of the in the model is significant since the p - value = (1.98e-09) < 0.05 level of significance, implying that there is a very strong evidence to conclude that has a significant linear contribution to . The coefficient of determination ( ) indicates www.ijsrp.org International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 that the is able to explain about 57.09% of the total variation in . If the error term is found to be autocorrelated, then the regression model could be termed spurious. Table 1: Output of Regression Model Call: lm(formula = Y ~ X) Residuals: Min 1Q Median 3Q Max -3985.9 -72.3 59.2 577.8 3080.8 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 56.0350 294.2920 0.190 0.85 X 4.0717 0.5383 7.564 1.98e-09 *** --Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 1595 on 43 degrees of freedom Multiple R-squared: 0.5709, Adjusted R-squared: 0.5609 F-statistic: 57.21 on 1 and 43 DF, p-value: 1.984e-09 In order to check if autocorrelations exist in the residuals obtained from the regression model in equation (3.1), we consider the ACF [Figure 1] of the residuals from the regression model; 339 independent variable, we examine the cross correlation function (CCF), that is, a smooth tapering pattern in the CCF shows which lags of independent variable we should used. Considering the CCF for and in [Figure 2], we noticed that it is unclear and misleading. To identify which lags of may predict , we therefore, apply the prewhitening technique to help us identify the lags of CCF. Figure 2: Cross Correlation Function of Money Supply and Gross Domestic Product By prewhitening, we mean fitting an ARIMA model to and reducing the residuals to white noise. Thereafter, we filter with the fitted ARIMA model to obtain the white noise residual series. Lastly, is filtered with the same model and then the cross correlation function is computed using the prewhitened and prewhitened . Now, fitting an ARIMA model to , we allow the data to talk. The ACF and PACF of [Figures 3and 4] respectively, indicate a tentative ARIMA(2,0,2) model alongside ARIMA(2,0,0) ARIMA(0,0,2) models. Both the ACF and PACF of residuals from ARIMA(2,0,2) model, [Figures 5 and 6], ARIMA(2,0,0) model, [Figures 7 and 8], and ARIMA(0,0,2) model, [Figures 9 and 10] respectively, are near white noise. Figure 1: ACF of Residuals from Regression Model it is observed that there are significant spikes at lags 2, 4 and 6 which are more than 5% of the total lags while all other lags fall within the confidence bounds, as such the residuals from the regression model appear to be autocorrelated and the regression model in (3.1) is termed spurious. The reason for autocorrelated error term of the regression model in (3.1) is not farfetched. The nonstationarity in and is more likely the cause of spurious correlations found in the error term of the regression model. To cop-out the menace of spurious correlations in the error term, we therefore, employ the dynamic regression method which takes into account both the lagged and current values of . In order to build the dynamic regression model successfully, we make use of Box and Jenkins three iterative methods; model identification, model estimation and model diagnostic checking. Model Identification To identify the nature of the time-lagged relationship between a dependent variable, and current and past values of an Figure 3: ACF of Money Supply Figure 4: PACF of Money Supply www.ijsrp.org International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 340 Figure 5: ACF of the Residuals from ARIMA (2,0,2) Model Figure 9: ACF of the Residuals from ARIMA (0,0,2) Model Figure 6: PACF of the Residuals from ARIMA(2,0,2) Model Figure 10: PACF of the Residuals from ARIMA (0,0,2) Model Since all the three tentative models seem appropriate, their outputs as seen in Table 2, Table 3 and Table 4 for ARIMA(2,0,2), ARIMA(2,0,0) and ARIMA(0,0,2) respectively, Figure 7: ACF of the Residuals from ARIMA(2,0,0) Model Figure 8: PACF of the Residuals from ARIMA(2,0,0) Model Table 2: Output of ARIMA(2,0,2) Model Call: arima(x = X, order = c(2, 0, 2)) Coefficients: ar1 ar2 ma1 ma2 intercept -0.1704 0.5977 0.5434 0.0290 297.0807 s.e. 0.1884 0.1892 0.2383 0.2389 134.1617 sigma^2 estimated as 116088: log likelihood = -326.69, aic = 665.38 Table 3: Output of ARIMA(2,0,0) Model Call: arima(x = X, order = c(2, 0, 0)) Coefficients: ar1 ar2 intercept 0.2188 0.4599 281.2463 s.e. 0.1281 0.1289 154.9215 sigma^2 estimated as 126455: log likelihood = -328.5, aic = 665 Table 4: Output of ARIMA(0,0,2) Model Call: arima(x = X, order = c(0, 0, 2)) Coefficients: ma1 ma2 intercept 0.3487 0.4537 312.8221 s.e. 0.1588 0.1082 96.1965 www.ijsrp.org International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 sigma^2 estimated as 132106: log likelihood = -329.42, aic = 666.84 we observed that both ARIMA(2,0,2) model and ARIMA(2,0,0) model have a smaller information criteria (aic) than ARIMA(0,0,2) model. Parsimoniously, we chose ARIMA(2,0,0) model over ARIMA(2,0,2) model. Thus, the estimated ARIMA(2,0,0) model is presented in (3.2) below; (1− . − . = � ) ∗ (3.2) where ∗ = ( − 281.2463) [Excerpts from Table 3] Then, we filter the series using the model for as shown in (3.3) (1− . − . ) (3.3) Now, the cross correlation function (CCF) for the prewhitened (which is the product of the ARIMA(2,0,0) model and its residuals) and the prewhitened is presented in [Figure 11]. We observed clear spikes at lags 0, -1,-2 and 1. Since we are only interested in the past values of , the spike at lag1 is ignored. Thus, , − and − should be included as predictors of . Figure 11: CCF for Prewhitened Money Supply and Prewhitened GDP Model Estimation The estimated dynamic regression model is presented in (3.4) below: = . − . Xt− − . . s.e (37.62991) (0.08255) (0.07183) t-value ( . (-1.673) (-1.978) (3.4) p-value (1.31e-05) (0.1017) (0.0545) = 0.2086 [Excerpts from Table 5]. Xt− − (0.12893) (-0.241) (0.8109) From the dynamic regression model in (3.4), the effect of using lagged values in overcoming spurious correlations is clearly seen as both lagged (Xt− , Xt− ) and current ( ) values in the model appear to show no sign of relationship with since their p - values are all less than 5% level of significance. Moreover, the coefficient of determination ( = . ) indicates that there is no linear relationship between and . Model Diagnostic Checking 341 The results from Box-Ljung test indicate that the residuals from the model in (3.4) are uncorrelated since = 17.46, df =22 with corresponding p-value = 0.7375 > 0.05 level of significance. [Excerpts from Table 6] Table 5: Output of Dynamic Regression Model Call: dynlm(formula = Y ~ XLag2 + XLag1 + X) Residuals: Min 1Q Median 3Q Max -158.84 -80.45 -28.63 20.12 715.01 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 185.75550 37.62991 4.936 1.31e-05 *** XLag2 -0.13813 0.08255 -1.673 0.1017 XLag1 -0.03104 0.12893 -0.241 0.8109 X -0.14208 0.07183 -1.978 0.0545 . --Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual standard error: 156.5 on 42 degrees of freedom (10 observations deleted due to missingness) Multiple R-squared: 0.2086, Adjusted R-squared: 0.1521 F-statistic: 3.691 on 3 and 42 DF, p-value: 0.01908 Table 6: Box – Ljung Test for Residuals from Dynamic Regression Model Box-Ljung test data: residuals(regmod01) X-squared = 17.46, df = 22, p-value = 0.7375 IV. CONCLUSION Although it is well documented in the literature that differencing both the dependent and independent variables offers a good solution to avoiding spurious correlations, our study takes a look at dynamic regression as another approach to avoiding spurious correlations. This was achieved by allowing the dependent variable to be expressed as a function of both lagged and current values of the independent variable. Regressing on , we found that the errors were correlated being a clear evidence of spurious correlation. However, the cross correlation function of the prewhitened variables indicated that lags 0, -1 and -2 of should be included as predictors of . Subsequently, we modeled a dynamic regression of with past lags 1, 2 and current lag of t as explanatory variables and the resulting residual series was diagnostically checked using Ljung and Box Q – statistic. The residuals were confirmed to be uncorrelated, revealing that and are totally unrelated. Therefore, we concluded that the problem of spurious correlations could be avoided by modeling a series using dynamic regression. This study could be extended to include the lagged values of the dependent variable, and both lagged and current values of multiple independent variables as predictors of the dependent variable. REFERENCES [1] H. Akaike, A New Look at the Statistical Model Identification, www.ijsrp.org International Journal of Scientific and Research Publications, Volume 6, Issue 10, October 2016 ISSN 2250-3153 [2] [3] [4] [5] [6] [7] [8] [9] IEEE Transactions on Automatic Control, 1973, Vol. 19, no 6, pp. 716 – 723. E. A. Akpan, I. U Moffat and N. B. Ekpo, Arma – Arch Modeling of the Returns of First Bank of Nigeria, European Scientific Journal, 2016, Vol.12, no.8, pp. 257 – 266. G.E.P. Box, G. M Jenkins and G. C. Reinsel, Time Series Analysis: Forecasting and Control. 3rd Ed., New Jersey: Wiley and Sons, 2008, pp. 522. G. E. P. Box and D. Pierce, Distribution of Residual Autocorrelations in Autoregressive Integrated Moving Average Time Series Models, Journal of the American Statistical Association, 1970, 65, pp.1509-1526. P. J. Brockwell and R. A. Davis, Introduction to Time Series and Forecasting, 2nd Ed,. Springer, 2002. J. D. Cryer, and K. Chan, Time Series Analysis with Application in R, 2nd Ed., Springer, 2008. N. R. Drasper and H. Smith, Applied Regression Analysis, 3rd Ed., New York, John Wiley and Sons, 1998. W. A. Fuller, Introduction to Statistical Time Series, 2nd Ed., New York, John Wiley and Sons, 1996. C. W. J. Granger and P. Newbold, Spurious Regressions in Econometrics, Journal of Econometrics, 1974, 2, pp.111 – 120. 342 [10] G. Ljung and G. C. Box, On a Measure of Lack of Fit in Time series models, Biometrica, 1978, Vol. 2 no.66, pp. 265-270. [11] A. Pankratz, Forecasting with Dynamic Regressions Models, 3nd Ed., New York, John Wiley and Sons, 1991. [12] J. O. Rawlings, S. G. Pantula and D. A. Dickey, Applied Regression Analysis: AResearch Tool, 2nd Ed., Springer, 1998. [13] W. W. S. Wei, Time Series Analysis Univariate and Multivariate Methods, 2nd Ed., Adison Westley, 2006. AUTHORS First Author – Akpan, E. A., Department of Mathematics and Statistics, University of Uyo, eubong44@gmail.com, +2348036200343 Second Author – Moffat, I. U. Ph. D, Department of Mathematics and Statistics, University of Uyo, moffitto@yahoo.com, +2348064497511 www.ijsrp.org