Technical tips on time series

with Stata
Gustavo Sánchez
Senior Statistician
StataCorp

2011 Mexican Stata Users Group Meeting 2011

 Outline
 Tip 1: Specifying the time structure
 tsset
 Date formats
 Tip 2: Why some predictions with -arima- do not match my
manual computations - Kalman Filter recursions
 Tip 3: What is the initial shock for impulse response functions
after -var-
 Tip 4: How do I fit my unobserved component model with
–sspace-
 linear regression and random walk

 Tip 5: How do I specify restrictions on the long-run

cointegrating relationship in the VEC model
TIP 1:Specifying the time structure
 tsset timevar [, options]
 Date frequency (daily, weekly, monthly,…)
 Clocktime (hours, minutes, seconds,…, milliseconds)
 Generic
 delta()
Example:
tsset timevar,daily delta(7)
lags in terms of seven days
TIP 1:Specifying the time structure
 Date formats
 Example – Daily format

clear
input str12 date
"1/01/2008"
"1/02/2008"
"1/03/2008"
"1/04/2008"
"1/05/2008"
end

generate mydate1=date(date,"DMY")
format mydate1 %td
generate mydate2=date(date,"DMY“)
format mydate2 %tdmon-DD,_CCYY
TIP 1:Specifying the time structure
 Date formats
 Example – Daily format
. list date mydate1 mydate2

+--------------------------------------+
| date mydate1 mydate2 |
|--------------------------------------|
1. | 1/01/2008 01jan2008 jan-01, 2008 |
2. | 1/02/2008 01feb2008 feb-01, 2008 |
3. | 1/03/2008 01mar2008 mar-01, 2008 |
4. | 1/04/2008 01apr2008 apr-01, 2008 |
5. | 1/05/2008 01may2008 may-01, 2008 |
+--------------------------------------+
TIP 1:Specifying the time structure
 Date formats
 Example – Daily format
. tsset mydate1
time variable: mydate1, 01jan2008 to 01may2008,
but with gaps
delta: 1 day

. list mydate1 if tin(01feb2008,01apr2008)

+-----------+
| mydate1 |
|-----------|
2. | 01feb2008 |
3. | 01mar2008 |
4. | 01apr2008 |
+-----------+
TIP 1:Specifying the time structure
 Date formats
 Example – Clock format
clear
Input str20 etime y
"06feb2010 12:40:00" 2
"06feb2010 12:42:00" 5
"06feb2010 12:44:00" 7
"06feb2010 12:46:00" 6
"06feb2010 12:48:00" 9
end

generate double mytime = clock(etime, "DMY hms")

format mytime %tc DMYHH:MM:SS
TIP 1:Specifying the time structure
 Date formats
 Example – Clock format
. tsset mytime,delta(2 minute)
time variable: mytime, 06feb2010 12:40:00 to 06feb2010 12:48:00
delta: 2 minutes

. generate my_ly=l.y
(1 missing value generated)
. list mytime y ly my_ly
+---------------------------------+
| mytime y my_ly |
|---------------------------------|
1. | 06feb2010 12:40:00 2 . |
2. | 06feb2010 12:42:00 5 2 |
3. | 06feb2010 12:44:00 7 5 |
4. | 06feb2010 12:46:00 6 7 |
5. | 06feb2010 12:48:00 9 6 |
+---------------------------------+
TIP 2: Predictions with -arima-
Kalman Filter recursions
 Let’s consider the following moving average
(MA1) model:

yt     t 1   t ;  t ~ i.i.d .N (0,  )
2

Command line
to fit the model: arima y, ma(1)

And we get the

predictions with: predict double y_hat
TIP 2: Predictions with -arima-
Kalman Filter recursions

 Users try to manually reproduce the predictions with:

 However, the results do not match the predictions

obtained with:
predict double y_hat

WHY?
TIP 2: Predictions with -arima-
Kalman Filter recursions
- Code for manual predictions
use http://www.stata-press.com/data/r11/lutkepohl,clear
arima dlinvestment, ma(1)
predict double yhat
scalar b0 = _b[_cons]
scalar t1 = [ARMA]_b[L1.ma]
gen double my_yhat = b0
gen double myehat = dlinvestment - b0 in 2
forvalues i = 3/91 {
qui replace my_yhat = my_yhat ///
+ t1*L.myehat in `i'
qui replace myehat = dlinvestment - my_yhat in `i'
}
TIP 2: Predictions with -arima-
Kalman Filter recursions
List first 12 predictions
. list qtr yhat my_yhat in 1/13,sep(11)
+--------------------------------+
| qtr yhat my_yhat |
|--------------------------------|
1. | 1960q1 .01686688 .01686688 |
2. | 1960q2 .01686688 .01686688 |
3. | 1960q3 .02052151 .02062398 |
4. | 1960q4 .01478403 .0147996 |
5. | 1961q1 .01312365 .01312617 |
6. | 1961q2 .00326376 .00326418 |
7. | 1961q3 .02471242 .02471249 |
8. | 1961q4 .01691061 .01691062 |
9. | 1962q1 .01412974 .01412975 |
10. | 1962q2 .00643301 .00643301 |
11. | 1962q3 .01940009 .0194001 |
|--------------------------------|
12. | 1962q4 .01649863 .01649863 |
13. | 1963q1 .01749646 .01749646 |
+--------------------------------+
TIP 2: Predictions with -arima-
Kalman Filter recursions
 Stata uses the recursive formula for the Kalman filter
prediction based on:

 Where:

estimated variance of the white noise disturbance

TIP 2: Predictions with -arima-
Kalman Filter recursions
use http://www.stata-press.com/data/r11/lutkepohl,clear
arima dlinvestment, ma(1)
predict double yhat

** Coefficient estimates and sigma^2 from ereturn list **

scalar b0 = _b[_cons]
scalar t1 = [ARMA]_b[L1.ma]
scalar sigma2 = e(sigma)^2

** pt and shrinking factor for the first two observations**

gen double pt=sigma2 in 1/2
gen double myratio=(sigma2)/(sigma2+t1^2*pt) in 2

** Predicted series and errors for the first two observations **

gen double my_yhat = b0
generate double myehat = myratio*(dlinvestment - my_yhat) in 2

** Predictions with the Kalman filter recursions **

forvalues i = 3/91 {
qui replace my_yhat = my_yhat + t1*l.myehat in `i'
qui replace pt= (sigma2)*(t1^2)*(L.pt)/ (sigma2+t1^2*L.pt) in `i'
qui replace myratio=(sigma2)/(sigma2+t1^2*pt) in `i'
qui replace myehat=myratio*(dlinvestment - my_yhat) in `i'
}
TIP 2: Predictions with -arima-
Kalman Filter recursions
List first 10 predictions
. list qtr yhat my_yhat pt myratio in 1/10

+--------------------------------------------------------+
| qtr yhat my_yhat pt myratio |
|--------------------------------------------------------|
1. | 1960q1 .01686688 .01686688 .00192542 . |
2. | 1960q2 .01686688 .01686688 .00192542 .97272668 |
3. | 1960q3 .02052151 .02052151 .00005251 .99923589 |
4. | 1960q4 .01478403 .01478403 1.471e-06 .99997858 |
5. | 1961q1 .01312365 .01312365 4.125e-08 .9999994 |
|--------------------------------------------------------|
6. | 1961q2 .00326376 .00326376 1.157e-09 .99999998 |
7. | 1961q3 .02471242 .02471242 3.243e-11 1 |
8. | 1961q4 .01691061 .01691061 9.092e-13 1 |
9. | 1962q1 .01412974 .01412974 2.549e-14 1 |
10. | 1962q2 .00643301 .00643301 7.147e-16 1 |
+--------------------------------------------------------+
TIP 3: Initial shock for Impulse response
functions (IRF) after -var-
VAR model
Yt    1Yt 1   2 Yt  2  ...   p Yt  p   t

Where:

Yt  (Y1t , Y2t ,..., Ykt ) : I(1) Endogenous variables

i : Matrix with coefficients associated to lag i

: Vectors with coefficients

associated to the intercepts
t : Vector with innovations
TIP 3: Initial shock for Impulse response
functions (IRF) after -var-
 Orthogonalized IRF functions for a shock in Y1
order1, Y1, Y1 order1, Y1, Y2
.06

.04

.02

-.02
0 5 10 0 5 10

step
95% CI orthogonalized irf
Graphs by irfname, impulse variable, and response variable
TIP 3: Initial shock for Impulse response
functions after -var-
What is the magnitude of the shock in the IRF graph?
order1, Y1, Y1 order1, Y1, Y2

-irf graph,irf-: simple IRF

1


 correspond to one-time unit

increase
.5

 the effects do not have a causal

interpretation
0

-.5
0 5 10 0 5 10

step
95% CI impulse response function (irf)
Graphs by irfname, impulse variable, and response variable
 TIP 3: Initial shock for Impulse response
functions after -var-
What is the magnitude of the shock in the IRF graph?
order1, Y1, Y1 order1, Y1, Y2
 -irf graph,oirf-: orthogonal IRF .06

 orthogonalization is produced via the .04

Cholesky decomposition

the magnitude of the shock corresponds to

.02


one unit standard deviation

0

 -irf graph,sirf- structural IRF

-.02
 -irf graph,sirf- IRF functions are derived 0 5 10 0 5 10

step
from the constraints imposed on the SVAR
95% CI orthogonalized irf
 the magnitude of the shock corresponds to Graphs by irfname, impulse variable, and response variable
one unit standard deviation
TIP 3: Initial shock for Impulse response
functions after -var-
 Let’s fit a VAR model:

use http://www.stata-press.com/data/r11/lutkepohl
var dlinvestment dlincome, lags(1/2) dfk
TIP 3: Initial shock for Impulse response
functions after -var-
. var dlinvestment dlincome,lags(1/2) dfk
 

Vector autoregression
Equation Parms RMSE R-sq chi2 P>chi2
----------------------------------------------------------------
dlinvestment 5 .04424 0.0856 8.32989 0.0802
dlincome 5 .011403 0.1027 10.1916 0.0373
------------------------------------------------------------------------------
| Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
dlinvestment |
dlinvestment |
L1. | -.2274192 .1053092 -2.16 0.031 -.4338214 -.021017
L2. | -.1159636 .1057698 -1.10 0.273 -.3232686 .0913415
dlincome |
L1. | .7103053 .3948248 1.80 0.072 -.0635372 1.484148
L2. | .5149489 .3935121 1.31 0.191 -.2563206 1.286218
_cons | -.0012273 .0111362 -0.11 0.912 -.0230539 .0205993
-------------+----------------------------------------------------------------
dlincome |
dlinvestment |
L1. | .0597466 .0271441 2.20 0.028 .0065451 .1129481
L2. | .0563513 .0272629 2.07 0.039 .002917 .1097855
dlincome |
L1. | .0209461 .1017687 0.21 0.837 -.1785169 .220409
L2. | .0833252 .1014303 0.82 0.411 -.1154745 .2821249
_cons | .0150368 .0028704 5.24 0.000 .0094108 .0206627
------------------------------------------------------------------------------
TIP 3: Initial shock for Impulse response
functions after -var-
 Plot the IRF function for a shock in dlinvestment
use http://www.stata-press.com/data/r11/lutkepohl var
dlinvestment dlincome, lags(1/2) dfk
irf create order1, step(10) set(myirf1,replace)
irf graph oirf, impulse(dlinvestment) ///
response(dlinvestment dlincome)
TIP 3: Initial shock for Impulse response
functions after -var-
 Plot the IRF function for a shock in dlinvestment
irf graph oirf, impulse(dlinvestment) ///
response(dlinvestment dlincome)
order1, dlinvestment, dlincome order1, dlinvestment, dlinvestment
.06

.04

.02

-.02
0 5 10 0 5 10

step
95% CI orthogonalized irf
Graphs by irfname, impulse variable, and response variable
TIP 3: Initial shock for Impulse response
functions after -var-
 Table for the OIRF function for a shock in dlinvestment
. irf table oirf, irf(order1) impulse(dlinvestment) ///
> response(dlincome dlinvestment)
 

Results from order1

+--------------------------------------------------------------------------------+
| | (1) (1) (1) | (2) (2) (2) |
| step | oirf Lower Upper | oirf Lower Upper |
|--------+-----------------------------------+-----------------------------------|
|0 | .001641 -.000715 .003998 | .04424 .037741 .050739 |
|1 | .002678 .000241 .005114 | -.008895 -.018459 .000669 |
|2 | .002154 -.000283 .004592 | -.00036 -.009879 .009159 |
|3 | -.000255 -.001358 .000849 | .004022 -.001068 .009113 |
|4 | .000394 -.000347 .001136 | .000056 -.002258 .00237 |
|5 | .000217 -.000207 .000641 | -.00033 -.002245 .001585 |
|6 | .000021 -.000237 .000279 | .000426 -.000457 .001309 |
|7 | .000025 -.000101 .000152 | .000068 -.000353 .000488 |
|8 | .00003 -.000055 .000116 | -.000036 -.000356 .000284 |
|9 | 4.4e-06 -.000034 .000043 | .000035 -.000074 .000143 |
|10 | 2.7e-06 -.000022 .000027 | .000015 -.000063 .000093 |
+--------------------------------------------------------------------------------+
95% lower and upper bounds reported
(1) irfname = order1, impulse = dlinvestment, and response = dlincome
(2) irfname = order1, impulse = dlinvestment, and response = dlinvestment
TIP 3: Initial shock for Impulse response
functions after -var-
 Table for the IRF function for a shock in dlinvestment
. irf table irf, irf(order1) impulse(dlinvestment) ///
> response(dlincome dlinvestment)
Results from order1
+--------------------------------------------------------------------------------+
| | (1) (1) (1) | (2) (2) (2) |
| step | irf Lower Upper | irf Lower Upper |
|--------+-----------------------------------+-----------------------------------|
|0 | 0 0 0 | 1 1 1 |
|1 | .059747 .004985 .114509 | -.227419 -.439876 -.014963 |
|2 | .044015 -.010388 .098419 | -.021806 -.237257 .193646 |
|3 | -.008218 -.032283 .015847 | .093362 -.027102 .213826 |
|4 | .007845 -.007056 .022745 | -.001875 -.054015 .050264 |
|5 | .004629 -.004709 .013967 | -.00906 -.054602 .036483 |
|6 | .000104 -.005125 .005332 | .009605 -.010735 .029945 |
|7 | .000451 -.002119 .003022 | .001323 -.00833 .010977 |
|8 | .000638 -.001136 .002413 | -.001041 -.008544 .006462 |
|9 | .000063 -.000688 .000814 | .000769 -.001641 .003179 |
|10 | .000042 -.000454 .000538 | .00032 -.001466 .002105 |
+--------------------------------------------------------------------------------+
95% lower and upper bounds reported
(1) irfname = order1, impulse = dlinvestment, and response = dlincome
(2) irfname = order1, impulse = dlinvestment, and response = dlinvestment
TIP 4: How do I fit my unobserved
component model with –sspace-
 State Space representation
zt  Azt 1  Bxt  C t
yt  Dzt  Fwt  Gt
Where:
zt : is an m x 1 vector of unobserved state variables;
xt : is a kx x 1 vector of exogenous variables;
εt : is a q x 1 vector of state-error terms, (q ≤ m);
yt : is an n x 1 vector of observed endogenous variables;
wt : is a kw x 1 vector of exogenous variables;
Ʋt : is an r x 1 vector of observation-error terms, (r ≤ n); and
A, B, C, D, F, and G are parameter matrices.
TIP 4: How do I fit my unobserved
component model with –sspace-
 State Space representation for linear regression
zt   t
yt  zt     wt
Command specification

constraint 1 [z]L.z = 0
constraint 2 [y]z = 1
sspace (z L.z, state noconstant) ///
(y w z,noerror ), constraints(1/2)
TIP 4:
 State Space estimation for linear regression
use http://www.stata-press.com/data/r11/lutkepohl,clear
constraint 1 [z]L.z = 0
constraint 2 [dlinvestment]z = 1
sspace (z L.z, state noconstant) (dlinvestment dlincome z,noerror ), constraints(1/2) nolog

State-space model
Sample: 1960q2 - 1982q4 Number of obs = 91
Wald chi2(1) = 0.88
Log likelihood = 154.44197 Prob > chi2 = 0.3487
( 1) [z]L.z = 0
( 2) [dlinvestment]z = 1
------------------------------------------------------------------------------
| OIM
dlinvestment | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
z |
z |
L1. | (omitted)
-------------+----------------------------------------------------------------
dlinvestment |
z | 1 . . . . .
dlincome | .3668678 .3914794 0.94 0.349 -.4004178 1.134153
_cons | .0096556 .008925 1.08 0.279 -.007837 .0271483
-------------+----------------------------------------------------------------
var(z) | .0019651 .0002913 6.75 0.000 .0013941 .0025361
------------------------------------------------------------------------------
Note: Tests of variances against zero are conservative and are provided only for reference.
TIP 4: How do I fit my unobserved
component model with –sspace-
 Random Walk yt  yt 1  

zt  zt 1   t
 State Space representation
yt  z t
- Command specification
constraint 1 [z]L.z = 1
constraint 2 [y]z = 1
sspace (z L.z, state noconstant) ///
(y z,noerror noconstant), constraints(1/2)
TIP 4:
 State Space estimation for Random Walk
use http://www.stata-press.com/data/r11/lutkepohl,clear
constraint 1 [z]L.z = 1
constraint 2 [dlinvestment]z = 1
sspace (z L.z, state noconstant) (dlinvestment z,noerror noconstant), constraints(1/2) nolog

State-space model
Sample: 1960q2 - 1982q4 Number of obs = 91
Log likelihood = 112.76541
( 1) [z]L.z = 1
( 2) [dlinvestment]z = 1
------------------------------------------------------------------------------
| OIM
dlinvestment | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
z |
z |
L1. | 1 . . . . .
-------------+----------------------------------------------------------------
dlinvestment |
z | 1 . . . . .
-------------+----------------------------------------------------------------
var(z) | .0046812 .0006978 6.71 0.000 .0033135 .006049
------------------------------------------------------------------------------
Note: Model is not stationary.
Note: Tests of variances against zero are conservative and are provided only for reference.
TIP 5:VEC – Johansen identification
Reduced form for a VEC model
p 1
zt  a  bt   zt 1   i zt i   t
i 1
Where:
zt : I(1) Endogenous variables

 : Matrices containing the long-run adjustment coefficients and

coefficients for the cointegrating relationships

i : Matrix with coefficients associated to short-run dynamic effects

a,b : Vectors with coefficients associated to the intercepts and trends

t : Vector with innovations

TIP 5:
Example: VEC with three endogenous variables

 z1t   z1t 1  p 1
 z1t i 
 z   a  bt    z   z  
 2 t   2 t 1  
i 1
i  2 t  i  t


 z3t  
 z3t 1   
 z 3 t  i 


Where:
11 12  z 
   11 12 13   1t 1 
 zt 1   21  22   z
  22  23   2t 1 
 31  32   21  z3t 1 

 Identifying α and β requires r2 restrictions (r: number of cointegrating vectors).

 Johansen FIML estimation identifies α and β by imposing r2 atheoretical restrictions.

11 12   z1t 1 

 1 0  
 zt 1   21  22   13 
 z 
0 1  23   
2 t 1

 31  32    z3t 1 
TIP 5:
- Restrictions based on Johansen normalization (Default)
use http://www.stata-press.com/data/r11/lutkepohl,clear
vec linvestment lincome lconsumption, rank(2) lags(2) noetable trend(none)

Vector error-correction model

  . . .
  . . .
  . . .
Identification: beta is exactly identified

Johansen normalization restrictions imposed

------------------------------------------------------------------------------
beta | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
_ce1 |
linvestment | 1 . . . . .
lincome | (omitted)
lconsumption | -.7943718 .0125908 -63.09 0.000 -.8190493 -.7696942
-------------+----------------------------------------------------------------
_ce2 |
linvestment | (omitted)
lincome | 1 . . . . .
lconsumption | -1.013321 .0013846 -731.87 0.000 -1.016035 -1.010608
------------------------------------------------------------------------------
TIP 5:
Instead of the Johansen atheoretical restrictions we could
Use economic theory to impose restrictions to identify αβ.
For example, let’s assume the following cointegrating equations:

z1t 1  0.75 z 2t 1  13 z3t 1

z3t 1  .85 z 2t 1   21 z1t 1

Which implies

11 12   z1t 1 

   1  .75 13   
 zt 1   21  22   z
 .85 1  
2 t 1 
  21
 31  32   z3t 1 
TIP 5:
- Restrictions specified by the user
constraint define 1 [_ce1]linvestment=1
constraint define 2 [_ce1]lincome=-.75
constraint define 3 [_ce2]lconsumption=1
constraint define 4 [_ce2]lincome=-.85
vec linvestment lincome lconsumption, rank(2) lags(2) noetable trend(none) bconstraints(1/4)

Identification: beta is exactly identified

( 1) [_ce1]linvestment = 1
( 2) [_ce1]lincome = -.75
( 3) [_ce2]lconsumption = 1
( 4) [_ce2]lincome = -.85
------------------------------------------------------------------------------
beta | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
_ce1 |
linvestment | 1 . . . . .
lincome | -.75 . . . . .
lconsumption | -.0343804 .0122816 -2.80 0.005 -.0584519 -.010309
-------------+----------------------------------------------------------------
_ce2 |
linvestment | -.1745742 .00322 -54.22 0.000 -.1808852 -.1682632
lincome | -.85 . . . . .
lconsumption | 1 . . . . .
------------------------------------------------------------------------------
 Summary
 Tip 1: Specifying the time structure
 Tip 2: Predictions with –arima-. Kalman Filter recursions
 Tip 3: Initial shock for Impulse response functions after -var-
 Tip 4: Unobserved component models with –sspace-

 Tip 5: Restrictions on cointegrating relationship for VEC models

Technical tips on time series
with Stata
Gustavo Sánchez
Senior Statistician
StataCorp

2011 Mexican Stata Users Group Meeting 2011