Multivariate time series
models
Autoregressive distributive lag models
𝑝 𝑞
• 𝑦𝑡 = 𝛽0 + 𝑖=1 𝛽𝑖 𝑦𝑡−𝑖 + 𝑗=0 𝜃𝑗 𝑥𝑡−𝑗 + 𝑒𝑡
σ σ
• p and q determined using information criteria and checking the
residuals
• Long run equilibrium:
𝑦𝑡 = 𝑦𝑡−1 = 𝑦𝑡−2 … = 𝑦𝑡−𝑝 = 𝑦 and 𝑥𝑡 = 𝑥𝑡−1 = 𝑥𝑡−2 … = 𝑥𝑡−𝑞 = 𝑥
and 𝑒𝑡 = 0
𝑝 𝑞
Then: 𝑦 = 𝛽0 + σ𝑖=1 𝛽𝑖 𝑦 + σ𝑗=0 𝜃𝑗 𝑥
𝑞 𝑝
𝑦 = (𝛽0 + σ𝑗=0 𝜃𝑗 𝑥)/(1 − σ𝑖=1 𝛽𝑖 )
Error correction model
• To present both the short-run dynamics and the long-run equilibrium
• Consider p = 1 and q = 1, and assume no constant term
• Deduct 𝑦𝑡−1 from both sides and add and substract 𝜃0 𝑥𝑡−1 from the right
hand side
∆𝑦𝑡 = (𝛽1 −1)𝑦𝑡−1 + 𝜃0 ∆𝑥𝑡 + (𝜃1 − 𝜃0 )𝑥𝑡−1 + 𝑒𝑡
𝜃1 − 𝜃0
∆𝑦𝑡 = 𝜃0 ∆𝑥𝑡 + (𝛽1 −1)(𝑦𝑡−1 − 𝑥𝑡−1 ) + 𝑒𝑡
(1 − 𝛽1 )
The term in parentheses is the long-run equilibrium and is called the error
correction term.
Note that both models are identical and will produce the same results.
Non-stationarity in multivariate time series
models
• Regressions of non stationary variables, t-statistics diverge and can
not be used to perform hypothesis tests two unrelated series may
turn out to be siginificantly correlated in the test
• If a series is I(d), differencing it d times will make it stationary. Using
differenced series in the regression is one alternative
• But information is lost
• There are cases when a linear combination of two non-stationary
series is stationarycointegration
Cointegration
• If 𝑦𝑡 is I(d) and 𝑥𝑡 is I(c), but 𝑧𝑡 = 𝑦𝑡 − 𝛼 −𝛽𝑥𝑡 is I(d-c), the series are
said to be cointegrated.
• Engle – Granger representation theorem: If two series are
cointegrate,d they would be most efficiently represented by an error
correction specification. Furthermor,e if they are cointegrated and the
series is validated, then it will encompass any other dynamic
specification.
Testing for cointegration
• Let 𝑦𝑡 and 𝑥𝑡 be I(1)
• If they are cointegrated, 𝑧𝑡 = 𝑦𝑡 − 𝛼 −𝛽𝑥𝑡 should be I(0)
• That is, the residual from the regression of 𝑦𝑡 on 𝑥𝑡 should be
stationary
• Testing for cointegration is equivalent to a Dickey-Fuller test on
residuals:
∆𝑧𝑡 = 𝛿𝑧𝑡−1 + 𝑒𝑡
𝐻0: 𝛿 = 0 𝐻1: 𝛿 ≠ 0
Estimation
• Engle Granger two step procedure:
• Stage 1: Estimate the cointegrating relation by regressing 𝑦𝑡 on 𝑥𝑡 .
Test the residuals for stationarity
• Stage 2: Estimate error correction model
Theoretical properties
• If series are stationary,
• Unbiasedness requrequireso be not correlated with the error term at all time
periods.
• For consistency, 𝑥𝑡 should not be correlated with the error term of earlier
periods.
• If series is non-stationary
• Estimators do not have a proper distribution
• But if there is a cointegrating relation, estimators converge to proper
distribution even faster
Granger causality
• Questions whether past values of a variable may help forecasting
another variable even after controlling for the past values of that
variable
• Regress 𝑦𝑡 on its own lags and lags of 𝑥𝑡
• Test whether the lags of 𝑥𝑡 are jointly significant
Vector autoregression
• Addresses endogeneity between the variables by presenting the reduced-
form specification.
• A simple example with 2 variables and 1 lag in the model
• Structural form
𝑦1𝑡 = 𝛽10 + 𝛽11 𝑦1𝑡−𝑖 + 𝛽12 𝑦2𝑡−𝑖 + 𝛽13 𝑦2𝑡 + 𝑒1𝑡
𝑦2𝑡 = 𝛽20 + 𝛽21 𝑦1𝑡−𝑖 + 𝛽22 𝑦2𝑡−𝑖 + 𝛽13 𝑦1𝑡 + 𝑒2𝑡
• For reduced form, replace the endogenous variables in the equations
𝛽10 − 𝛽13 𝛽20 𝛽11 − 𝛽13 𝛽21 𝛽12 − 𝛽13 𝛽22 𝑒1𝑡 + 𝛽13 𝑒2𝑡
𝑦1𝑡 = + 𝑦1𝑡−𝑖 + 𝑦2𝑡−𝑖 +
1 − 𝛽13 𝛽23 1 − 𝛽13 𝛽23 1 − 𝛽13 𝛽23 1 − 𝛽13 𝛽23
𝑦1𝑡 = 𝛼10 + 𝛼11 𝑦1𝑡−𝑖 + 𝛼12 𝑦2𝑡−𝑖 + 𝑢𝑡
VAR – matrix form
𝐵𝑦𝑡 = 𝛽0 + 𝛽1 𝑦𝑡−1 + 𝑒𝑡
1 −𝛽13 𝛽11 𝛽12 𝑦1𝑡
•𝐵= , 𝛽1 = , 𝑦𝑡 = 𝑦
−𝛽23 1 𝛽21 𝛽22 2𝑡
• Pre multiply with the inverse of B:
𝐵−1 𝐵𝑦𝑡 = 𝐵−1 𝛽0 + 𝐵−1 𝛽1 𝑦𝑡−1 + 𝐵 −1 𝑒𝑡
𝑦𝑡 = 𝐴0 + 𝐴1 𝑦𝑡−1 + 𝑢𝑡
Stationarity of the variables
• If 𝑦1𝑡 , 𝑦2𝑡 , … , 𝑦𝑘𝑡 are stationary, estimation is as usual. One only
needs to pay attention to the correlation across the error terms of
different equations
• If they are non-stationary, one needs to inquire about cointegration.
• To better understand the methodology, lets start with a simple model
with two variables and one lag
• We make use of this to formalize the cointegration model
𝑦𝑡 = 𝐴0 + 𝐴1 𝑦𝑡−1 + 𝑢𝑡
Substract 𝑦𝑡−1 from both sides:
Δ𝑦𝑡 = 𝐴0 + (𝐴1 −𝐼)𝑦𝑡−1 + 𝑢𝑡
Δ𝑦𝑡 = 𝐴0 + Π𝑦𝑡−1 + 𝑢𝑡
The number of independent rows in Π is the number of cointegrating
relations. That, in turn, is equivalent to the number of eigenvalues of Π
that are greater than 0.
Therefore, if eigenvalues are non-zero, there is cointegration
Note that this also is equivalent to rank of Π being non-zero.
General model
𝑦𝑡 = 𝐴0 + 𝐴1 𝑦𝑡−1 + 𝐴2 𝑦𝑡−2 + ⋯ + 𝐴𝑝 𝑦𝑡−𝑝 + 𝑢𝑡
Substract 𝑦𝑡−1 from both sides, then add and substract (𝐴1 − 𝐼) 𝑦𝑡−2
Δ𝑦𝑡 = 𝐴0 + (𝐴1 − 𝐼)Δ𝑦𝑡−1 + (𝐴2 + 𝐴1 − 𝐼)𝑦𝑡−2 + ⋯ + 𝐴𝑝 𝑦𝑡−𝑝 + 𝑢𝑡
Add and substract (𝐴2 + 𝐴1 − 𝐼) 𝑦𝑡−3 and continue adding and
substracting terms in this manner to obtain:
Δ𝑦𝑡 = 𝐴0 + 𝐵1 Δ𝑦𝑡−1 + 𝐵2 Δ𝑦𝑡−2 + … + 𝐵𝑝−1 Δ𝑦𝑡−𝑝+1 − Π𝑦𝑡−𝑝 + 𝑢𝑡
Where B’s are functions of 𝐴𝑖 s, and Π = (𝐼 − 𝐴1 − 𝐴2 − ⋯ − 𝐴𝑝 )
If rank of Π is equal to 0, there is no cointegrating relation
If rank of Π is equal to k, number of variables, all variables stationary
If rank of Π is equal to r, 0<r<k, there are r cointegrating relation
• When rank is r with 0<r<k, Π can be expressed as:
Π = 𝛼𝛽
where 𝛽 a rxk matrix expressing cointegrating relations and 𝛼 is a kxr matrix
expressing adjustment to deviations from the long term equilibrium.
Estimation
• If all variables stationary, determine the lag length with information
criteria, estimate parameters and check residuals
• If all variables non-stationary and there is no cointegrating relation,
transform the variables in differences, check stationarity and estimate
the model in differences
• If there may be cointegrating relations, use Johansen technique
Johansen technique
• Estimate Π using VECM representation
• Obtain the estimates of the eigenvalues
• Test for their equality to zero
Impulse response
• Illustrates the impact of shocks, gets the reader to better understand
the results
• Usually, we set one of the error terms to one standard deviation of it,
consider all variables to have 0 as their previous period value, and
assume no further shock afterwards
• Then:𝑦0 = 0, 𝑒11 = 𝑠. 𝑑. 𝑒𝑡 , 𝑒𝑗1 = 0 𝑓𝑜𝑟 𝑗 ≠ 1, 𝑒ℎ𝑙 = 0 𝑓𝑜𝑟 𝑙 > 1
𝑦1 = 𝐴1 𝑦0 + 𝑒1 = 𝑒1
𝑦2 = 𝐴1 𝑦1 + 𝑒2 = 𝐴1 𝑒1
…
𝑦𝑠 = 𝐴1 𝑦𝑠−1 + 𝑒𝑠 = 𝐴1𝑠−1 𝑒1
Taking correlation between errors into
account
• Setting only one of the error terms to a particular value does not take
into account correlation between error terms of different equations
• We can transform the model in such a way that the original error
terms are particular functions of independent random variables
• Then we may consider a change in one of these independent
variables and see how the variables change in response
• In this method, order of the variables matters
Variance decomposition
• Remember that forecast error is largely due to the error term being
set to zero for forecasted periods.
• In a VAR setting, it involves error terms for all equations, even if
indirectly
• Variance decomposition aims to decompose the variance of forecast
error, answering the question how much of it is due to the error term
of different equations
Sources
• Diebold: 16.3 - 16.6
• Wooldridge: 18.1, 18.3-18.5
• Stock & Watson: 16.1-16.3, 17.1-17.4