The present work constitutes the second part of a two-paper project that, in particular, deals with an in-depth study of effective techniques used in econometrics in order to make accurate forecasts in the concrete framework of one of the major economies of the most productive Italian area, namely the province of Verona. It is worth mentioning that this region is indubitably recognized as the core of the commercial engine of the whole Italian country. This is why our analysis has a concrete impact; it is based on real data, and this is also the reason why particular attention has been taken in treating the relevant economical data and in choosing the right methods to manage them to obtain good forecasts. In particular, we develop an approach mainly based on vector autoregression where lagged values of two or more variables are considered, Granger causality, and the stochastic trend approach useful to work with the cointegration phenomenon.
Econometrics time seriesautoregressive modelsGranger causalitycointegrationstochastic nonstationaritytrends and breaksIntroduction
In this second part of a two-paper project, we move from theory of autoregressive, possibly multivalued, time series to the study of a concrete framework. In particular, exploiting precious economic data that the Commerce Chamber of Verona Province has put at our disposal, we successfully applied some of the relevant approaches introduced in [5] to find dependencies between economic factors characterizing the Province economy, then to make effective forecasts, very close to the real behavior of studied markets. The present part of the project is divided as follows: first, we consider an AR-approach to Verona import–export time series, then we provide a VAR model analysis of Verona relevant econometric data taken from various web databases such as Coeweb, Stockview, and Movimprese, and, within the last section, we compare such data with those coming from the whole Italian scenario. We would like to emphasize that all the theoretical background and related definitions can be retrieved from [5].
AR-approach to Verona import–export time series
In what follows, we shall apply techniques developed in previous sections to analyze our main empirical problem of forecasting export and import data for the Verona district, also using other variables such as active enterprises. These applications are based on Istat data retrieved from the database Coeweb.
EXP
We present a time series regression model in which the regressors are past values of the dependent variable, namely the Export data. We use 92 observations of variable EXP, quarterly data from 1991 to 2013 expressed in Euros. Figure 1 shows the related time series.
Export of Verona
Looking at Fig. 1, we can see that the Verona export shows relatively smooth growth, although this decreases during the years 2008–2011. Decline in exports is likely caused by economic crisis broken out in Italy in those years. Although the curve may seem apparently growing, it is also possible to notice that there are periodic trends during the years under consideration. In fact, in the fourth quarter of 1992, the curve has a significant growth, then increases fairly linearly until about the second quarter of 1994, in which one can recognize a new increasing period that slightly more obvious than the previous one. This periodicity of 18 months can also be seen in other parts of the curve, but not after the beginning of the current economic crisis, where very likely there will be a structural break. In order to test the goodness of our qualitative analysis based on historical data, we used a software called GRETL, which is particularly useful to perform statical analysis of time series. The mean and standard deviation related to the quarter of this variable EXP are respectively Mean=1579900000€andStandardDeviation=499880000€, whereas the annual mean for EXP is 1579900000×4=6319600000€. The first seven autocorrelations of EXP are ρ1=corr(EXPt,EXPt−1)=0.9718, ρ2=0.9755, ρ3=0.9450, ρ4=0.9523, ρ5=0.9165, ρ6=0.9242, ρ7=0.8931. Previous entries show that inflation is strongly positively autocorrelated; in fact, the first autocorrelation is 0.97. The autocorrelation remains large even at a lag of six quarters. This means that an increase in export in one quarter tends to be associated with an increase in the next quarter. Autocorrelation starts to decrease from the lag of seventh quarters. In what follows, we report the output obtained testing for autoregressive models according to an increasing number of delays, from 1 to 6 delays, on the variable EXP, namely:
the AR(1) case: EXP=65090000+0.971606EXPt−1
Coefficient
Standard Error
t-Statistic
p-Value
const
6.50900e+007
2.35520e+007
2.7637
0.0069
EXPt−1
0.971606
0.017392
55.8652
0.0000
SER
1.17e+08
R2
0.944426
AdjustedR2
0.943802
AIC
3641.074
BIC
3646.096
the AR(2) case: EXP=57965600+0.409313EXPt−1+0.573763EXPt−2
Coefficient
Standard Error
t-Statistic
p-Value
const
5.79656e+007
2.92851e+007
1.9794
0.0509
EXPt−1
0.409313
0.0920617
4.4461
0.0000
EXPt−2
0.573763
0.105188
5.4546
0.0000
SER
97 111 006
R2
0.60913
AdjustedR2
0.960014
AIC
3568.804
BIC
3576.303
the AR(3) case: EXP=54025100+0.618705EXPt−1+0.726958EXPt−2-0.366510EXPt−3
Coefficient
Standard Error
t-Statistic
p-Value
const
5.40251e+007
2.26874e+007
2.3813
0.0195
EXPt−1
0.618705
0.109790
5.6353
0.0000
EXPt−2
0.726958
0.063352
11.4749
0.0000
EXPt−3
−0.366510
0.115843
−3.1639
0.0022
SER
91 264 682
R2
0.964681
AdjustedR2
0.963435
AIC
3519.089
BIC
3529.044
the AR(4) case: EXP=54498000+0.748057EXPt−1+0.466614EXPt−2−0.592869EXPt−3
Coefficient
Standard Error
t-Statistic
p-Value
const
5.44980e+007
2.43509e+007
2.2380
0.0279
EXPt−1
0.748057
0.142495
5.2497
0.0000
EXPt−2
0.466614
0.075211
6.2041
0.0000
EXPt−3
−0.592869
0.156045
−3.7993
0.0003
EXPt−4
0.361048
0.065852
5.4827
0.0000
SER
86 223 417
R2
0.967898
AdjustedR2
0.966351
AIC
3470.537
BIC
3482.924
the AR(5) case: EXP=56242200+0.870848EXPt−1+0.247032EXPt−2−0.417031EXPt−3+0.648298EXPt−4−0.372917EXPt−5
Coefficient
Standard Error
t-Statistic
p-Value
const
5.62422e+007
2.12088e+007
2.6518
0.0096
EXPt−1
0.870848
0.135548
6.4246
0.0000
EXPt−2
0.247032
0.096569
2.5581
0.0124
EXPt−3
−0.417031
0.178982
−2.3300
0.0223
EXPt−4
0.648298
0.105669
6.1352
0.0000
EXPt−5
−0.372917
0.119834
−3.1119
0.0026
SER
80 872 743
R2
0.970976
AdjustedR2
0.969185
AIC
3420.938
BIC
3435.733
and the AR(6) case:
EXP=55434600+1.01304EXPt−1+0.00610464EXPt−2−0.251406EXPt−3+0.542831EXPt−4−0.737681EXPt−5+0.408104EXPt−6
Coefficient
Standard Error
t-Statistic
p-Value
const
5.54346e+007
2.23371e+007
2.4817
0.0152
EXPt−1
1.01304
0.12541
8.0777
0.0000
EXPt−2
0.006105
0.107043
0.0570
0.9547
EXPt−3
−0.251406
0.131646
−1.9097
0.0598
EXPt−4
0.542831
0.116130
4.6743
0.0000
EXPt−5
−0.737681
0.104151
−7.0828
0.0000
EXPt−6
0.408104
0.089469
4.5614
0.0000
SER
75 057 009
R2
0.974384
AdjustedR2
0.972438
AIC
3369.763
BIC
3386.943
We estimate the AR order of our autoregression related to obtained numerical results using both BIC and AIC information criteria (see Table 1).
BIC, AIC, AdjustedR2, and SER for the six AR models
p
BIC(p)
AIC(p)
AdjustedR2(p)
SER(p)
1
3646.096
3641.074
0.943802
117000000
2
3576.303
3568.804
0.960014
97111006
3
3529.044
3519,089
0.963435
91264682
4
3482.924
3470.537
0.966351
86223417
5
3435.733
3420.938
0.969185
80872743
6
3386.943
3369.763
0.972438
75057009
Both BIC and AIC are the smallest in the AR(6) model (from the seventh delay onwards the criteria begin to increase); we conclude that the best estimate of the lag length is 6, hence supporting our qualitative analysis. Previous data from Table 1 indicate that as the number of lags increases, the AdjustedR2 increases, and the SER decreases. R2, AdjustedR2, andSER measure how well the OLS estimate of the multiple regression line describes the data. The standard error of the regression (SER) estimates the standard deviation of the error term, and thus, it is a measure of spread of the distribution of a variable Y around the regression line. The regression R2 is the fraction of the sample variance of Y explained by (or predicted by) the regressors, the R2 increases whenever a regressor is added, unless the estimated coefficient on the added regressor is exactly zero. An increase in the R2 does not mean that adding a variable actually improves the fit of the model, so the R2 gives an inflated estimate of how well the regression fits the data. One way to correct this is to deflate or reduce the R2 by some factor, and this is what the AdjustedR2 does, which is a modified version of R2 that does not necessarily increase when a new regressor is added. As seen by numerical output in Table 1, the increase in AdjustedR2 is large from one to two lags, smaller from two to three, and quite small from three to four and in the next lags. Exploiting the results obtained for the AIC/BIC analysis, we can determine how large the increase in the AdjustedR2 must be to justify including the additional lag. In the AR(6) model of Eq. (1), the coefficients of EXPt−1, EXPt−4, EXPt−5, and EXPt−6 are statically significant at the 1% significance level because their p-value is less than 0.01, and the t-statistic exceeds the critical value. The constant, however, is statically significant at the 5% significance. The coefficient of EXPt−3 is statically significant at the 10% significance, and the coefficient of EXPt−2 is not statically significant. In particular, the 95% confidence intervals for these coefficient are as follows:
Variable
Coefficient
95% Confidence Interval
const
5.54346e+007
1.09738e+007
9.98955e+007
EXPt−1
1.01304
0.76341
1.26266
EXPt−2
0.006105
−0.206959
0.219168
EXPt−3
−0.251406
−0.513441
0.010627
EXPt−4
0.542831
0.311680
0.773981
EXPt−5
−0.737681
−0.944989
−0.530374
EXPt−6
0.408104
0.230022
0.586187
Large-sample critical values of the augmented Dickey–Fuller statistic
Deterministic Regressors
10%
5%
1%
Intercept only
−2.57
−2.86
−3.43
Intercept and time trend
−3.12
−3.41
−3.96
In order to check whether the EXP variable has a trend component or not, we test the null hypothesis that such a trend actually exists against the alternative EXP being stationary, by performing the ADF test for a unit autoregressive root. Large-sample critical values of the augmented Dickey–Fuller statistic yield the following ADF regression with six lags of EXPt, where the subscript t indicates a particular quarter considered:
ΔEXPˆt=55434600+δEXPt−1+γ1ΔEXPt−1+γ2EXPt−2+γ3ΔEXPt−3+γ4ΔEXPt−4+γ5ΔEXPt−5+γ6ΔEXPt−6.
The ADF t-statistic is the t-statistic testing the hypothesis that the coefficient on EXPt−1 is zero; this is t=−1.23. From Table 2, the 5% critical value is −2.86. Because the ADF statistic of −1.23 is less negative than −2.86, the test does not reject the null hypothesis at the 5% significance level. Based on the regression in Eq. (2), we therefore cannot reject the null hypothesis that export has a unit autoregressive root, that is, that export contains a stochastic trend, against the alternative that it is stationary. If instead the alternative hypothesis is that Yt is stationary around a deterministic linear trend, then the ADF t-statistic results in t=−4.07, which is less than −3.41 (from Table 2). Hence, we can reject the null hypothesis that export has a unit autoregressive root. We proceed with a test QLR, which provides a way to check whether the export curve has been stable in the period from 1993 to 2010. Specifically, we focus on whether there have been changes in the coefficients of the lagged values of export and of the intercept in the AR(6) model specification in Eq. (1) containing six lags of EXPt. The Chow F-statistics (see, e.g., [7, Sect. 5.3.3]) tests the hypothesis that the intercept and the coefficients of EXPt−1,…,EXPt−6 in Eq. (1) are constant against the alternative that they break at a given date for breaks in the central 70% of the sample. The F-statistic is computed for break dates in the central 70% of the sample because for the large-sample approximation to the distribution of the QLR statistic to be a good one, the subsample endpoints cannot be too close to the beginning or to the end of the sample, so we decide to use 15% trimming, that is, to set τ0=0.15T and τ1=0.85T (rounded to the nearest integer). Each F-statistic tests seven restrictions. Restrictions on the coefficients equaled to zero under the null hypothesis (see [5, Sect. 2.4]), and since in our case we have the coefficients of the six delays and the intercept, we get seven restrictions. The largest of these F-statistics is 13.96, which occurs in 2010:I (the first quarter of 2010); this is the QLR statistic. The critical value for seven restrictions is presented in Table 3.
Critical values of QLR statistic with 15% truncation
Number of restrictions
10 %
5 %
1 %
7
2.84
3.15
3.82
The previously reported values indicate that the hypothesis of stable coefficients is rejected at the 1% significance level. Thus, there is an evidence that at least one of these seven coefficients changed over the sample. These results also confirm the assumptions that we made earlier since the year 2010 coincides with an increasing import of the financial crisis before arriving at a partial economic recovery. A forecast of Verona export in 2014:I using data through 2013:IV can be then based on our established AR(6) model of export, which gives
EXP=55434600+1.01304EXPt−1+0.00610464EXPt−2−0.251406EXPt−3+0.542831EXPt−4−0.737681EXPt−5+0.408104EXPt−6.
Therefore, substituting the values of export into each of the four quarters of 2013, plus the two last quarters of 2012, we have
EXPˆ2014:I|2013:IV=55434600+1.013EXP2013:IV+0.006EXP2013:III−0.251EXP2013:II+0.543EXP2013:I−0.738EXP2012:IV+0.408EXP2012:III=55434600+1.013×2511098163+0.006×2326958115−0.251×2329551351+0.543×2209212521−0.738×2420606501+0.408×2265903940≅2366137617€,
so that, for 2014:II, we obtain
EXPˆ2014:II|2014:I=55434600+1.013EXP2014:I+0.006EXP2013:IV−0.251EXP2013:III+0.543EXP2013:II−0.738EXP2012:I+0.408EXP2012:IV=55434600+1.013×2366137617+0.006×2511098163−0.251×2326958115+0.543×2329551351−0.738×2209212521+0.408×2420606501≅2505454123€,
and forecasts for all 2014 quarters are as follows:
Quarter
Forecast
Error
2014:I
2 366 130 000
75 057 000
2014:II
2 505 450 000
106 841 000
2014:III
2 422 950 000
131 981 000
2014:IV
2 527 660 000
145 016 000
It is worth mentioning that the forecast error increases as the number of considered quarters increases. Figure 2 shows, through a graph, forecasts since 2002 in sample and forecasts for 2014, highlighting the confidence intervals.
Forecasts of Verona export
ΔEXP
It is also useful to analyze the time series of the growth rate in exports that we denoted by ΔEXP. Economic time series are often analyzed after computing their logarithms or the changes in their logarithms. One reason for this is that many economic series exhibit growth that is approximately exponential, that is, over the long run, the series tends to grow by a certain percentage per year on average, and hence the logarithm of the series grows approximately linearly. Another reason is that the standard deviation of many economic time series is approximately proportional to its level, that is, the standard deviation is well expressed as a percentage of the level of the series; hence, if this is the case, the standard deviation of the logarithm of the series is approximately constant. It follows that it turns to be convenient to work with the variable ΔEXPt=ln(EXPt)−ln(EXPt−1). Taking into account the data shared in Fig. 3, we retrieve the following information:
Meanonaquarterlybasis=0.014958=1.49%StandardDeviationonaquarterlybasis=0.079272=7.93%AverageGrowthRateonayearlybasis=0.014958×4=0.059832=5.98%
The first four autocorrelations of ΔEXP are ρ1=−0.6133, ρ2=0.5698, ρ3=−0.6100, ρ4=0.7029.
Rate of growth in exports
Even if it might seem contradictory that the level of export is strongly positively correlated but its change is negatively correlated, we have to consider that such values measure different things. The strong positive autocorrelation in export reflects the long-term trends in export; in contrast, the negative autocorrelation of the change of export means that, on average, an increase in export in one quarter is associated with a decrease in export in the next one. Analogously to what we have seen in Section 2.1, we perform an AIC/BIC analysis for ΔEXP obtaining that the best choice for the lag lay is 4, so that we have
ΔEXP=0.0128189−0.173627ΔEXPt−1+0.0996175ΔEXPt−2−0.189882ΔEXPt−3+0.416414ΔEXPt−4.
Coefficient
Standard Error
t-Statistic
p-Value
const
0.0128189
0.0077887
1.6458
0.1036
ΔEXPt−1
−0.173627
0.119987
−1.4470
0.1517
ΔEXPt−2
0.099618
0.100542
0.9908
0.3247
ΔEXPt−3
−0.189882
0.096363
−1.9705
0.0522
ΔEXPt−4
0.416414
0.094464
4.4081
0.0000
SER
0.052736
R2
0.576787
AdjustedR2
0.556142
AIC
−260.2402
BIC
−247.9107
In our AR(4) model, the coefficients of ΔEXPt−4 are statically significant at the 1% significance level because their p-value is less than 0.01 and the t-statistic exceeds the critical value. The coefficient of ΔEXPt−3 is statically significant at the 10% significance. The constant and the other coefficients are not statically significant. Even when the information criteria are very low, this is not a good model because R2 and AdjustedR2 are relatively small. So this AR(4) model turns out to be not very useful to predict the growth rate in exports. Figure 3 shows that the frequency in this case is annual; moreover, an increase in ΔEXP in one quarter is associated with a decrease in the next one. In this case, the results of ADF test allow us to reject the null hypothesis that rate of growth in export has a unit autoregressive root both with the alternative hypothesis of stationarity and of stationarity around a deterministic linear trend. It follows that the QLR statistic is 5.02, which occurs in 2009:I, and hence the hypothesis that the coefficients are stable is rejected at the 1% significance level. Again, the results of the software GRETL confirm that the crisis of recent years has greatly affected the exports from Verona. Consequently, by the results obtained we have that the forecast of ΔEXP for 2014, given in the table
Quarter
Forecast
Error
2014:I
−4.86%
0.052736
2014:II
5.11%
0.053525
2014:III
−1.58%
0.053961
2014:IV
6.16%
0.055304
and also sketched in Fig. 4, is not very accurate, and the predictions do not perceive the lower peaks of the variable, which is confirmed by the low value of R2.
IMP
We now turn to the empirical problem to predict Verona import by analyzing its historical series. We present an autoregressive model that uses the history of Verona import to forecast its future. We use 92 observations of variable import, quarterly data from 1991 to 2013 expressed in Euros. Figure 5 shows the time series.
Forecasts of ΔEXP
Import of Verona
Looking at Fig. 5, we can see that Verona import shows relatively smooth growth, although this decreases during the years 2008–2011; the curve is very similar to the time series of export, and hence it is reasonable to deduce that decline in import is likely caused by economic crisis broken out in Italy in those years. Although the curve may seem apparently growing, periodic trends appear during years under consideration. This curve has an annual periodicity. Looking at a minimum of the curve, exactly one year later, another minimum exists. The mean and standard deviation on a quarterly basis for IMP are Mean=2177300000€ and StandardDeviation=697420000€, whereas the annual mean export is 2177300000×4=8709200000€. The first five IMP autocorrelation values are ρ1=0.9424, ρ2=0.9280, ρ3=0.9060, ρ4=0.9260, ρ5=0.8750. These entries show that inflation is strongly positively autocorrelated; in fact, the first autocorrelation is 0.94. The autocorrelation remains large even at the lag of four quarters. This means that an increase in import in one quarter tends to be associated with an increase in the next quarter. Autocorrelation, as expected, starts to decrease starting from the lag of five quarters. As with the variable EXP, we estimated the AR order of an autoregression in IMP using both the AIC and BIC information criteria, finally obtaining that the optimal lag length is 4.
Coefficient
Standard Error
t-Statistic
p-Value
const
1.90005e+008
6.15140e+007
3.0888
0.0027
IMPt−1
0.499665
0.0997006
5.0117
0.0000
IMPt−2
0.155637
0.0746261
2.0856
0.0401
IMPt−3
−0.154911
0.0881396
−1.7576
0.0825
IMPt−4
0.434062
0.0827892
5.2430
0.0000
SER
198 000 000
R2
0.911613
AdjustedR2
0.907354
AIC
3616.812
BIC
3629.198
Therefore, we have
IMP=190005000+0.499665IMPt−1+0.155637IMPt−2−0.154911IMPt−3+0.434062IMPt−4.
We check now if the model has a trend. The null hypothesis that Verona import has a stochastic trend can be tested against the alternative that it is stationary by performing the ADF test for a unit autoregressive root. The ADF regression with four delays of IMP gives
ΔIMPˆt=190005000+δIMPt−1+γ1ΔIMPt−1+γ2IMPt−2+γ3ΔIMPt−3+γ4ΔIMPt−4.
The ADF t-statistic is the t-statistic testing the hypothesis that the coefficient on IMPt−1 is zero, and it turns to be t=−1.78. From Table 2, the 5% critical value is −2.86. Because the ADF statistic of −1.78 is less negative than −2.86, the test does not reject the null hypothesis at the 5% significance level. We therefore cannot reject the null hypothesis that import has a unit autoregressive root, that is, that import contains a stochastic trend, against the alternative that it is stationary. If the alternative hypothesis is that Yt is stationary around a deterministic linear trend, then the ADF t-statistic results in t=−2.6, which is less negative than −3.41. So, in this case, we also cannot reject the null hypothesis that export has a unit autoregressive root.
We proceed with a QLR test, which provides a way to check whether the import curve has been stable during the years sparing from 1993 to 2010. The Chow F-statistic tests the hypothesis that the intercept and the coefficients at IMPt−1,…,IMPt−4 in Eq. (3) are constant against the alternative that they break at a given date for breaks in the central 70% of the sample. Each F-statistic tests five restrictions. The largest of these F-statistics is 10.26, which occurs in 1995:III; the critical values for the five-restriction model at different levels of significance are given in Table 4. These values indicate that the hypothesis that the coefficients are stable is rejected at the 1% significance level. Thus, there is an evidence that at least one of these five coefficients changed over the sample; namely, we have a structural break, which might be caused by the devaluation that the Lira currency experienced during the period 1992–1995. According to the previous analysis, the predictions of import of Verona for the year 2014 are as follows:
Quarter
Forecast
Error
2014:I
2 775 360 000
197 957 000
2014:II
2 752 530 000
197 957 000
2014:III
2 639 510 000
235 388 000
2014:IV
2 721 670 000
236 693 000
They result in a slight increase for the next year, as shown by Fig. 6.
Critical values of QLR statistic with 15% truncation
Number of restrictions
10 %
5 %
1 %
5
3.26
3.66
4.53
Forecasts of Verona import
ΔIMP
The fourth variable of interest is represented by the logarithm of the ratio between consecutive values of IMP, that is,
ΔIMPt=ln(IMPt)−ln(IMPt−1)=ln(IMPtIMPt−1).
The first six autocorrelations values of ΔIMP are presented in Table 5.
Autocorrelations of ΔIMP
j
1
2
3
4
5
6
ρj
−0.4240
0.0631
−0.3910
0.6721
−0.3844
0.0743
In the case of the growth rate of export, the negative autocorrelation of the change of import means that, on average, an increase in import in one quarter is associated with a decrease in the next one. From the fifth lag, autocorrelation starts to be less significant. So, it can be easily seen from Fig. 7 and the autocorrelations in Table 5 that the right estimate of the lag length is 4. The consequent AR(4) model reads as follows:
ΔIMP=0.0128189−0.173627ΔIMPt−1+0.0996175ΔIMPt−2−0.189882ΔIMPt−3+0.416414ΔIMPt−4,
and the following Fig. 7 shows the time series of ΔIMP, and we can see how an increase in import in one quarter is associated with a decrease in the next one.
Rate of growth in imports
Coefficient
Standard Error
t-Statistic
p-Value
const
0.0161472
0.0110277
1.4642
0.1470
ΔIMPt−1
−0.326437
0.0950214
−3.4354
0.0009
ΔIMPt−2
−0.224760
0.0878146
−2.5595
0.0123
ΔIMPt−3
−0.280232
0.0960526
−2.9175
0.0046
ΔIMPt−4
0.431620
0.0894247
4.8266
0.0000
SER
0.083791
R2
0.531621
AdjustedR2
0.508773
AIC
−179.6755
BIC
−167.3459
The QLR statistic for AR(4) model in Eq. (5) is 22.58, which occurs in 1995:II. This value indicates that the hypothesis that the coefficients are stable is rejected at the 1% significance level. As for imports, we can associate this structural break to the last crisis of Lira occurred in that period. We observe the dynamics of the real effective exchange rate in Fig. 8.
Evolution of the Real Exchange Rate (index numbers: 1992 = 100) (Source: FMI)
As shown in Fig. 9, the devaluation of the Lira has produced some benefits for the growth of Italian exports (goods and services), especially looking at analogous economical data for Germany and France.
Growth of Exports of Goods and Services (index numbers: 1992 = 100; correct values with the GDP deflator) (Source: World Bank data)
As shown in Fig. 10, the devaluation of the Lira did not stop the value of imports, but you can still easily perceive the rupture of 1995.
Growth of Imports of Goods and Services (index numbers: 1992 = 100; correct values with the GDP deflator) (Source: World Bank data)
Active Enterprises
We would like also to briefly analyze the variable “Active Enterprises” (ACTEt), namely the time series with quarterly data from 1995 to 2013, where each observation is the number of firms operating in a given quarter in the province of Verona. With the software GRETL we obtain the AR(4) model
ACTE=9535.97+1.02210ACTEt−1−0.173385ACTEt−2+0.0152586ACTEt−3+0.0280194ACTEt−4.
The AdjustedR2 of this regression is 0.94, and the QLR statistic is 37.52, which occurs in 2011:I. This value indicates that the hypothesis that the coefficients are stable is rejected at the 1% significance level. Also, for the variable ACTEt, we can conclude that the number of active businesses were affected by the crisis of those years. However, the ADF t-statistic for this variable does not reject the null hypothesis, so we cannot reject the fact that the time series of the numbers of active enterprises has a unit autoregressive root, that is, that ACTEt contains a stochastic trend, against the alternative that it is stationary. From Fig. 11 we can see that the curve has a quite regular annual pattern and that active enterprises tend to decline in the first quarter of each year and then return generally to grow. It is worth to mention the drastic rise of the curve during the first period of the time interval under consideration. Such an increase has been caused by a particular type of bureaucratic constraints, namely by a sort of forced registration imposed to a rather large set of farms companies previously not obliged to be part of the companies register. Such a norm has been introduced in two steps, first by a simple communication (1993), and later in the form of legal disclosure (2001).
ACTEt
VAR models analysis of Verona data
In this section, we apply the theory developed in the fourth chapter to analyze the set of Verona import and export time series. Therefore, we consider a VAR model for exports (EXPt), imports (IMPt), and active companies (ACTEt) in Verona, and each of such variables is characterized by time series constituted by quarterly data from 1995 to 2013.
First model: stationary variables
As we saw in Chapter 2, the import end export of Verona are subject to a stochastic trend, so that it is appropriate to transform it by computing its logarithmic first differences in order to obtain stationary variables. Figure 12 shows a multiple graph for the time series of ΔEXPt, ΔIMPt, and ΔACTEt.
Multiple graph for ΔEXPt, ΔEXPt and ΔACTEt
The VAR for ΔEXPt, ΔIMPt, and ΔACTEt consists of three equations, each of which is characterized by a dependent variable, namely by ΔEXPt, ΔIMPt, and ΔACTEt, respectively. Because of the apparent breaks in considered time series for the years 1995 and 2010, the VAR is estimated using data from 1996:I to 2008:IV. The number of lags of this model are obtained through information criteria BIC and AIC using the software GRETL, which gives the results in Table 6, where the asterisks indicate the best (or minimized) of the respective information criteria.
VAR lag lengths
p
AIC(p)
BIC(p)
1
−13.610968
−13.119471
2
−14.685333
−13.825212*
3
−14.572491
−13.343746
4
−14.747567
−13.150199
5
−14.974180
−13.008189
6
−15.160238*
−12.825624
7
−15.048342
−12.345105
8
−15.047682
−11.975822
The smallest AIC has been obtained considering six lags; indeed, the BIC estimation of the lag length is pˆ=2. We decide to choose two delays because, for pˆ=6, we have a VAR with three variables and six lags, so we will have 19 coefficients (eight lags with three variables each, plus the intercept) in each of the three equations, with a total of 57 coefficients, and we saw in [5, Sect. 4.2] that estimation of all these coefficients increases the amount of the forecast estimation error, resulting in a deterioration of the accuracy of the forecast itself. We also prefer consider the BIC estimation for its consistency; however, the AIC overestimate p (see [5, Sect. 2.2]. Estimating the VAR model with GRETL produces the following results:
ΔEXPt=0.0014−0.44ΔEXPt−1−0.14ΔEXPt−2−0.19ΔIMPt−1+0.21ΔIMPt−2−0.15ΔACTEt−1+0.35ΔACTEt−2,ΔIMPt=0.0222−0.5ΔEXPt−1+0.57ΔEXPt−2−0.38ΔIMPt−1−0.46ΔIMPt−2+0.09ΔACTEt−1+0.2ΔACTEt−2,ΔACTEt=0.0043+0.02ΔEXPt−1+0.12ΔEXPt−2+0.07ΔIMPt−1−0.02ΔIMPt−2+0.23ΔACTEt−1+0.02ΔACTEt−2.
In the first equation (ΔEXPt) of VAR system (7), we have the coefficients of ΔEXPt−1, ΔIMPt−2, and ΔACTEt−2, which are statically significant at the 1% significance level because their p-value is less than 0.01 and the t-statistic exceeds the critical value. The constant and the coefficients of ΔIMPt−1, however, are statically significant at the 5% significance, and the other coefficients are not statically significant. The AdjustedR2 is 0.53. In the second equation (ΔIMPt) of VAR system (7), we have the coefficients of ΔEXPt−1, ΔEXPt−2, ΔIMPt−1, and ΔIMPt−2, which are statically significant at the 1% significance level. The constant, however, is statically significant at the 10% significance, and the other coefficients are not statically significant. The AdjustedR2 is 0.45. In the last equation of (7), we have only the constant statically significant, at the 5% level. The AdjustedR2 is −0.04. These VAR equations can be used to perform Granger causality tests. The results of this test for the first equation of (7) are as follows:
Variable
Test F
p-Value
ΔIMPt
12.464
0.0001
ΔACTEt
8.2240
0.0010
The F-statistic testing the null hypothesis that the coefficients of ΔIMPt−1 and ΔIMPt−2 are zero in the first equation is 12.46 with p-value 0.0001, which is less than 0.01. Thus, the null hypothesis is rejected at the level of 1%, so we can conclude that the growth rate in Verona import is a useful predictor for the growth rate in export, namely ΔIMPt Granger-causes ΔEXPt. Also, ΔACTEt Granger-causes the change in export at the 1% significance level. The results for the second equation of (7) are as follows:
Variable
Test F
p-Value
ΔEXPt
22.766
0.0000
ΔACTEt
1.5894
0.2161
For the ΔIMPt equation, we can also conclude that the growth rate in Verona export is a useful predictor for the growth rate in import, but the change in the number of active enterprises is not. The results for the last equation of (7) are as follows:
Variable
Test F
p-Value
ΔEXPt
1.0897
0.3456
ΔEXPt
1.6413
0.2059
The F-statistic testing the null hypothesis that the coefficients of ΔEXPt−1 and ΔEXPt−2 are zero in the first equation is 1.09 with p-value 0.34, which is greater than 0.10. Thus, the null hypothesis is not rejected, so we can conclude that the growth rate in Verona import is not a useful predictor for the growth rate in active enterprises, namely, ΔIMPt does not Granger-cause ΔACTEt. The F-statistic testing the hypothesis that the coefficients of the two lags of ΔEXPt are zero is 1.64 with p-value of 0.2; thus, ΔEXPt also does not Granger-cause ΔACTEt at the 10% significance level. Forecasts of the three variables in system (7) are obtained exactly as discussed in the univariate time series models, but in this case, the forecast of ΔEXPt, we also consider past values of ΔIMPt and ΔACTEt.
Forecasts of ΔEXPt
Quarter
ΔEXPt
Forecast
Error
95% Confidence Interval
2009:1
−0.02329
0.018816
0.044544
−0.071077
0.108709
2009:2
0.06674
0.017759
0.052249
−0.087683
0.123202
2009:3
−0.06221
−0.002095
0.060086
−0.123354
0.119164
2009:4
−0.003635
0.006493
0.063164
−0.120976
0.133963
2010:1
−0.1911938
0.016435
0.065429
−0.115605
0.148475
2010:2
0.0002207
0.013870
0.066425
−0.120182
0.147922
2010:3
−0.03853
0.004754
0.067609
−0.131686
0.141194
2010:4
0.08106
0.009609
0.068220
−0.128063
0.147282
2011:1
−0.002692
0.013526
0.068644
−0.125004
0.152055
2011:2
0.1127259
0.011529
0.068840
−0.127397
0.150454
2011:3
−0.02047
0.007798
0.069059
−0.131568
0.147164
2011:4
0.08649
0.010466
0.069186
−0.129156
0.150088
2012:1
−0.04747
0.011927
0.069269
−0.127863
0.151716
2012:2
0.06716
0.010711
0.069309
−0.129160
0.150583
2012:3
−0.003477
0.009260
0.069350
−0.130694
0.149213
2012:4
0.07761
0.010650
0.069377
−0.129358
0.150658
2013:1
−0.07186
0.011142
0.069393
−0.128898
0.151182
2013:2
0.05621
0.010477
0.069401
−0.129580
0.150535
2013:3
−0.04800
0.009946
0.069409
−0.130127
0.150019
2013:4
0.06604
0.010640
0.069415
−0.129445
0.150724
2014:1
−0.09138
0.010775
0.069418
−0.129315
0.150866
2014:2
0.05304
0.010436
0.069420
−0.129658
0.150530
2014:3
−0.001114
0.010259
0.069421
−0.129838
0.150356
2014:4
0.07616
0.010592
0.069422
−0.129507
0.150692
Forecast for ΔEXPt (color online)
Forecast for ΔIMPt (color online)
Forecast for ΔACTEt (color online)
By means of the forecasts from 2009 to 2013, we can establish a comparison with the real data, noting that the predictions with this VAR model are not very reliable since the error is quite high and it increases in recent years. The lack of accuracy was confirmed previously by low values of the AdjustedR2. Figures 13, 14, and 15 show the real time series of the three variables with a red line and the prediction made with the estimated models with a blue line. It can be seen from these graphs that the confidence intervals (green area in the figures) are very high.
Second model: nonstationary variables
In this section, we analyze the three variable (EXPt, IMPt, and ACTEt), considering quarterly Verona data from 1995 to 2013. We analyze these time series without avoiding structural breaks and without considering the first differences, and we check if the analysis produces different results with respect to the previous ones. Figure 16 shows a multiple graph for the time series respectively of EXPt, IMPt, and ACTEt.
Multiple Graph for EXPt, IMPt, and ACTEt
The GRETL lag length selection gives the results in Table 8; then, according to the considerations made to determine the number of delays for the model (7), we decide to choose three delays, obtaining the following model:
EXPt=−119893000+0.79EXPt−1+0.47EXPt−2−0.31EXPt−3−0.12IMPt−1+0.20IMPt−2−0.09IMPt−3+4389.61ACTEt−1+4715.09ACTEt−2−6479.85ACTEt−3,IMPt=−313115000+0.17EXPt−1+1.11EXPt−2−1.21EXPt−3+0.52IMPt−1−0.13IMPt−2+0.28IMPt−3+13719.5ACTEt−1−3215.16ACTEt−2+1103.38ACTEt−3,ACTEt=8526.18−1.62×10−6EXPt−1+1.87×10−6EXPt−2−7059×10−8EXPt−3+1.31×10−6IMPt−1−4.89×10−7IMPt−2−3.81×10−7IMPt−3+1.04ACTEt−1−0.17ACTEt−2+0.02ACTEt−3.
In the first equation (EXPt) of VAR system (8), we have the coefficients of EXPt−1, EXPt−2, IMPt−2, and ACTEt−3, which are statically significant at the 1% level because their p-value is less than 0.01 and the t-statistic exceeds the critical value. However, the coefficients of EXPt−3 and IMPt−1 are statically significant at the 5% level, and the other coefficients are not statically significant. The coefficient of IMPt−3 is statically significant at the 10% level, and the others are not statically significant. The AdjustedR2 is 0.93. In the second equation (IMPt) of VAR system (8), we have the coefficients of EXPt−2, EXPt−3, IMPt−1, IMPt−3, and ACTEt−1, which are statically significant at the 1% significance level. The constant is statically significant at the 5% level, and the other coefficients are not statically significant. The AdjustedR2 is 0.85. In the last equation of (8), we have only EXPt−1, ACTEt−1, and ACTEt−2 statically significant respectively at the 5%,1%, and 10% levels, whereas the AdjustedR2 is 0.94. If we perform Granger causality tests, then we have that all p-values of the F-statistic of the three equations are less than 0.01; only for the third equation of (8), the Granger causality test for the variable EXPt has the p-value 0.0852, and hence EXPt Granger-causes ACTEt, but in this case, the null hypothesis is rejected at the level of 10%. Notice that the model (8) has high values of the AdjustedR2, so it can be very useful to make prediction of future values of the three variables. The forecasts for EXPt concerning 2014 are given by the table
VAR lag lengths
p
AIC(p)
BIC(p)
1
98.133995
98.525673
2
97.843826
98.529262
3
97.432952
98.412147*
4
97.327951
98.600904
5
97.204547
98.771258
6
97.107478
98.967947
7
97.020628
99.174856
8
97.007994*
99.455981
Quarter
Forecast
Error
2014:I
2 415 830 000
86 744 900
2014:II
2 502 280 000
105 860 000
2014:III
2 430 160 000
143 193 000
2014:IV
2 488 770 000
158 629 000
Forecast for EXPt
Forecasts for IMPt are as follows:
Quarter
Forecast
Error
2014:I
2 764 470 000
174 343 000
2014:II
2 870 330 000
200 990 000
2014:III
2 712 960 000
237 479 000
2014:IV
2 809 610 000
249 185 000
and prediction for ACTEt reads as follows:
Quarter
Forecast
Error
2014:I
87401.59
1812.928
2014:II
87988.15
2634.586
2014:III
88266.25
3129.437
2014:IV
88495.47
3449.842
Figures 17, 18, and 19 show the time series of the three variables and their forecasts. The area of confidence interval for EXPt is rather small, which is confirmed by the value 0.93 of the AdjustedR2 of the first equation in system (8). This area is slightly wider for the second graph, and in Fig. 19 we show the confidence interval for ACTEt becoming wider at each quarter.
Forecast for EXPt
Forecast for ACTEt
No cointegration between EXPt and IMPt
We saw in Sections 2.1 and 2.3 that the time series for EXPt and IMPt are both integrated of order 1 (I(1)); hence, we perform an EG-ADF test to verify if these two variables are cointegrated. The cointegrating coefficient θ is estimated by the OLS estimate of the regression EXPt=α+θIMPt+zt; hence, we obtain EXPt=197119000+0.641536IMPt+zt, so that θ=0.641536. Then we use a Dickey–Fuller test to test for a unit root in zt=EXPt−θIMPt. The statistic test result is −2.77065, which is greater than −3.96 (see [5, Table 1] for critical values); therefore, we cannot refuse the null hypothesis of a unit root for zt, concluding that the series EXPt−θIMPt is not stationary. Moreover, we have that the variables EXPt and IMPt are not cointegrated.
VAR model with Italian data
In this section, we perform a comparison of the time series between provincial and national data. Considering the same model of system (8), but with data referring to Italy, we get a VAR(8) model of the form
EXPnt=βˆ10+βˆ11EXPnt−1+⋯+βˆ18EXPnt−8+γˆ11IMPnt−1+⋯+γˆ18IMPnt−8+δˆ11ACTEnt−1+⋯+δˆ18ACTEnt−8,IMPnt=βˆ20+βˆ21EXPnt−1+⋯+βˆ28EXPnt−8+γˆ21IMPnt−1+⋯+γˆ28IMPnt−8+δˆ21ACTEnt−1+⋯+δˆ28ACTEnt−8,ACTEnt=βˆ30+βˆ31EXPnt−1+⋯+βˆ38EXPnt−8+γˆ31IMPnt−1+⋯+γˆ38IMPnt−8+δˆ31ACTEnt−1+⋯+δˆ38ACTEnt−8,
where the letter n in the variable name indicates that we are working with national data.
Comparison between EXPnt and EXPt
Comparison of IMPnt and IMPt
Comparison of ACTEnt and ACTEt
The AdjustedR2 of the three equations in system (9) are respectively 0.95, 0.96, and 0.98. So this is a good VAR model; in fact, Granger causality tests for (9) present all p-values of the F-statistic less than 0.01. So all the three variables can be used to explain the others. In Figs. 20, 21, and 22, we note the extreme similarity of the provincial and national time series. If we perform an EG-ADF test to verify if this three couples of variables are cointegrated, then we obtain that only the variables ACTEnt and ACTEt are cointegrated with cointegrating coefficient θ=49.4948. By comparing the correlation between a variable of national data and the corresponding variables with provincial data we note a high correlation level, even taking into account the provincial variable delays. Below we present the correlation between EXPn and the delays of EXP:
p
corr(EXPnt;EXPt+p)
−4
0.7918
−3
0.8083
−2
0.8985
−1
0.9036
0
0.9823
1
0.8880
2
0.8677
3
0.7711
4
0.7557
Then we have the correlation between IMPn and the delays of IMP
p
corr(IMPnt;IMPt+p)
−4
0.7490
−3
0.7645
−2
0.8428
−1
0.8745
0
0.9641
1
0.8780
2
0.8518
3
0.7887
4
0.7948
whereas the correlation between ACTEn and the delays of ACTE are given by
p
corr(IMPnt;IMPt+p)
−4
0.6493
−3
0.7400
−2
0.8290
−1
0.9162
0
0.9947
1
0.9257
2
0.8464
3
0.7634
4
0.6771
Figures 23, 24, and 25 show the correlation diagram related to the national and provincial variables. We notice very high values, which show the strong connection between what happens at the national and the provincial levels.
Correlation between EXPn and EXP
Correlation between IMPn and IMP
Correlation between ACTEn and ACTE
Conclusion
We have presented an analysis of relevant time series related to the import and export data concerning the Province of Verona, together with a forecast analysis of the 2014 trend. Exploited techniques have been treated in our first paper, and these two articles together constitute a unitary project. In this second part, we have paid attention to the quantitative influence that certain macro economical events may have on considered time series. In particular, we extrapolated three particularly significant moments, namely the 2007–2008 world financial economic crisis, with consequent decrease of import–export, a break in 1995 probably due to the devaluation of the Lira, which did not cause a decrease of the import, but resulted in an increase in exports of Verona, and the vertical growth of the Active enterprises parameter during 1995–1998, which has been caused by a change in the related provincial regulation. It is worth to underline how our analysis shows, by obtained numerical forecasts, a concrete possibility for a partial recovery from the present economic crisis, especially when taking into account the first quarters of 2014 and particularly with regard to exports. The results obtained can be used for concrete actions aimed, for example, to the optimization of territory economic resources, even if a concrete economical program needs of a deeper treatment for which, however, our analysis constitutes a rigorous and effective basis. Concerning the latter, possible extensions may be focused on analyzing import and export time series of specific products to underline in which areas Verona is more specialized; then such results could be used to understand where to invest more. Moreover, we could perform a comparison analysis with analogous data belonging to other cities of similar economical size, both in Italy and within the European Community.
Acknowledgements
The authors would like to thanks the Camera di Commercio di Verona for the precious database that has been put at our disposal. Any effective statistical/econometric analysis cannot be realized without using real data. Any concrete forecast cannot be possible without counting on such a kind of really precious time series. Therefore, the present project would have not see the light without the concrete help of the Camera. A particular acknowledgment goes to Dr. Stefania Crozzoletti and to Dr. Riccardo Borghero.
ReferencesBaldi, P.: . The McGraw-Hill Companies, Milan (2007) Bee Dagum, E.: . Springer Verlag, Italia (2002) Bernstein, S., Bernstein, R.: . McGraw-Hill, Milan (2003) Brandt, P.T., Williams, J.T.: . Sage Publications, Thousand Oaks (2007) Di Persio, L.: Autoregressive approaches to import–export time series I: basic techniques. Mod. Stoch. Theory Appl. (2015). doi:10.15559/15-VMSTA22Harris, R., Sollis, R.: . John Wiley & Sons Ltd, West Sussex, England (2003) Kirchgässner, G., Wolters, J.: . Springer-Verlag, Berlin, Heidelberg (2007). MR2451567Stock, J.-H., Watson, M.W.: . Pearson, Italy, Milan (2012) Wei, W.W.S.: . Pearson, United States of America (2006). MR2517831