Vector Autoregression

1
 Overview
 Vector Autoregression (VAR) model is an extension of univariate
autoregression model to multivariate time series data

 VAR model is a multi-equation system where all the variables are

treated as endogenous (dependent)

 There is one equation for each variable as dependent variable. In

its reduced form, the right-hand side of each equation includes
lagged values of all dependent variables in the system, no
contemporaneous variables

2
 VAR Model
VAR(p) model:
Yt = a + A1Yt-1 + A2Yt-2 + … + ApYt-p + εt

where:
Yt = (y1t, y2t, …, ynt)’ : an (nx1) vector of time series variables
a: an (nx1) vector of intercepts
Ai (i=1, 2, …, p): (nxn) coefficient matrices
εt : an (nx1) vector of unobservable i.i.d. zero mean error term
(white noise)

3
 VAR Model
Example: Bivariate VAR(2) Model

 y1t   a1  b111 b12

1
  y1t 1  b112 b122   y1t  2   1t 
 y   a    1 1    2 2    
 2t   2  b21 b22   y2t 1  b21 b22   y2t  2   2t 
Or:
y1t = a1 + b111y1t-1 + b112y2t-1 + b211y1t-2 + b212y2t-2 + ε1t
y2t = a2 + b121y1t-1 + b122y2t-1 + b221y1t-2 + b222y2t-2 + ε2t

4
 The Simplest VAR Model
Example: Bivariate VAR(1) Model with no intercepts

5
The Marginal Models

6
 The Simplest VAR Model
Example: Bivariate VAR(1) Model with no intercepts

7
The Final Equation

8
The Simplest VAR Model

9
 VAR Stationarity
Based on the Final Equation
 The LEFT-HAND SIDE of the univariate series {Yt} is a
stationary AR(2) if both roots of ϕ(L)=0 are outside the unit
circle.
 This is equivalent to: both roots of P(λ)=0 are inside the unit
circle.
 (What process does the error term on the right-hand side
follow? White noise? MA(1)? Other model?)
 Question: What about the univariate series {Xt}?

10
Check for Stationarity

11
 The Simplest VAR Model
Example: Bivariate VAR(1) Model with no intercepts

12
The Companion Form

13
The Companion Form II

14
 The Companion Form
 F is an example of a companion matrix.
 Since (9) is the same equation as the characteristic equation
for the final equation for yt , we have that [yt , xt ] is a
stationary process if and only if the eigenvalues of F have
moduli less than 1 (inside the unit circle).

15
The Companion Form

16
The Companion Form

17
 The Companion Form
 The VAR(p) is a stationary process if and only if all the
eigenvalues of F have moduli less than 1 (inside the unit
circle).
 Recall, the eigenvalues are the solutions of

18
 Why do we need VAR?
 Time-series data with autoregressive in nature (serially
correlated)
 VAR model is one of the most successful and flexible models
for the analysis of multivariate time series
 Especially useful for describing the dynamic behavior of
economic and financial time series
 Useful for forecasting

19
 Applications of VAR

 Analysis of system response to different shocks/impacts

 Model-based forecast. In general VAR encompasses
correlation information of the observed data and use this
correlation information to forecast future movements or
changes of the variable of interest

20
 Applications of VAR

 In economics, VAR is used to forecast macroeconomic

variables, such as GDP, money supply, and unemployment

 In finance, predict spot prices and future prices of securities;

foreign exchange rates across markets

21
 Applications of VAR

 In accounting, predict different accounting variables such as

sales, earnings, and accruals

 In marketing, VAR can be used to evaluate the impact of

different factors on consumer behavior and forecast its future
change.

22
Forecasting

23
 Applications of VAR:
Forecasting
 1-step forecast based on information available at time T:
YT+1|T = a + A1YT + A2YT-1 + … + ApYT-p+1

 h-step forecast:
YT+h|T = a + A1YT+h-1|T + A2YT+h-2|T + … + ApYT+h-p|T

24
 Implementation
 All data have to have same frequency
 Data with mixed frequency need to be converted to the same
frequency

Convert higher-frequency data to the frequency of the lowest-

frequency data). For example: if we have daily, weekly and
monthly data then we will need to convert everything to monthly
frequency

Interpolate lower-frequency data into high frequency

25
SAS & R for VAR

26
 Example of VAR usage

 Testable hypothesis: there has to be a dependence of DJIA

index on its own lag and on lag of total market capitalization
and vice versa
 Use return on DJIA index and return on market capitalization
 Monthly observation

27
 Dataset
Obs year month ret_dji ret_totval

1 1961 1 . .

2 1961 2 0.02619 0.032994

3 1961 3 0.02575 0.033013

4 1961 4 0.01442 0.006127

5 1961 5 0.02067 0.022185

6 1961 6 -0.02172 -0.030350

7 1961 7 0.01913 0.033709

8 1961 8 0.03243 0.022871

9 1961 9 -0.03279 -0.020618

28
SAS

29
 SAS Implementation
 PROC VARMAX
proc varmax data=comb;

model ret_dji ret_totval / p=1;

run;
 ret_djit = a1 + b11 ret_djit-1 + b12 ret_totvalt-1 + ε1t
 ret_totvalt = a2 + b21 ret_djit-1 + b22 ret_totvalt-1 +ε2t

30
 SAS Output
Mode l Param e te r Es tim ate s
S tandard
Equation Param e te r Es tim ate Error t Value Pr > |t| Variable
re t_dji CONS T1 0.00542 0.00195 2.78 0.0057 1
AR 1_1_1 -0.29801 0.06973 -4.27 0.0001 ret_dji(t-1)
AR 1_1_2 0.37788 0.07499 5.04 0.0001 ret_totval(t-1)
re t_totval CONS T2 0.00802 0.00185 4.32 0.0001 1
AR 1_2_1 0.06674 0.06627 1.01 0.3144 ret_dji(t-1)
AR 1_2_2 -0.04813 0.07127 -0.68 0.4998 ret_totval(t-1)

Ret_djit = 0.005 – 0.298 ret_djit-1 + 0.378 ret_totvalt-1 + e1t

Ret_totvalt = 0.008 + 0.067 ret_djit-1 – 0.048 ret_totvalt-1 + e2t

31
SAS Output

32
R

33
 R for VAR
 VARs are implemented in the vars package in R. It contains
a function VARselect to choose the number of lags p using
four different information criteria: AIC, HQ, SC and FPE.
We have met the AIC before, and SC is simply another name
for the BIC (SC stands for Schwarz Criterion after Gideon
Schwarz who proposed it). HQ is the Hannan-Quinn
criterion and FPE is the “Final Prediction Error” criterion.
Care should be taken using the AIC as it tends to choose
large numbers of lags. Instead, for VAR models, we prefer to
use the BIC.

34
 Concerns
 Assuming all variables are endogenous
 If time-series data are nonstationary (containing stochastic
trends), while it is possible to estimate VAR in levels, it is
preferable to estimate VAR in first differences
 Uncertainty about number of lags (using LR test,
Information criteria: AIC, BIC etc.)

35
 Concerns
 Data requirements (long time series)
 Imprecise estimated coefficients (overfitting the model).
Solution – restrict or weight coefficients

 Computationally intensive

36
 References
 courses.ttu.edu/isqs6348-westfall/VARPresentation.ppt
 Chapter 1: Vector Autoregressions.
https://www2.bc.edu/~iacoviel/teach/0809/EC751_files/var.p
df
 Chapter 6: Multivariate time series models.
www.nek.lu.se/.../Ch6%20Multivar
iate%20time%20series%20models

37
 References
 http://www.uio.no/studier/emner/sv/oekonomi/ECON5101/v
11/undervisningsmateriale/Lect7v11.pdf
 Dwyer, Gerald P., Jr. Why Are Vector Autoregressions Useful
in Finance? http://jerrydwyer.com/pdf/lectvar.pdf
 Vector Autoregressions: Forecasting and Reality.http://
www.frbatlanta.org/filelegacydocs/robtallman.pdf

38
 References
 Chapter 11: Vector Autoregressive Models for Multivariate
Time Series.
http://faculty.washington.edu/ezivot/econ584/notes/varMode
ls.pdf
 https://www.otexts.org/fpp/9/2

39