Spurious Relationships and Cointegration

Today

Multivariate nonstationary time series
- Spurious regression
- Cointegration
“An Introduction to Nonsense”
or “How to lose money in the stock market”
Material based on Wooldridge Ch. 18

Some Important Data Relationships (1)

Marriage vs. Margarine

Some Important Data Relationships (2)

Marriage vs. Drowning

Some Important Data Relationships (3)

Space vs. Sociology

What’s going on here?

We know “correlation is not causation”
Correlations may be found between variables even with no causal relationship
Omitted variables can induce relationships between otherwise unconnected variables
What’s the omitted variable here?
Should we assume there is some hidden force relating sociologists and space exploration, and try to find it?
Alternately, should we assume that these large correlations are just highly unlikely sampling variability?
For both, probably not
Reason is related to features that make time series relationships different from cross sectional ones
With time series, Correlation may not even reflect correlation!
Sample properties can show strong relationships even for independently generated series

Simulation (Code 1)

tl<-100 #Length of simulated series
set.seed(74) #Initialize random number generator
e1<-ts(rnorm(tl)) #standard normal shocks
e2<-ts(rnorm(tl)) #additional independent standard normal
rho<-1 #AR parameter
yUR1<-numeric(tl-1)
yUR2<-numeric(tl-1)
for(s in 1:(tl-1)){
  #Random Walk initialized at 0
  yUR1[s+1]<-rho*yUR1[s]+e1[s]
  #Independently generated separate random walk
    yUR2[s+1]<-rho*yUR2[s]+e2[s]
}

Simulation (Code 2)

#Represent as same length time series
UR1<-ts(yUR1[1:tl-1])
UR2<-ts(yUR2[1:tl-1])
simulatedseries<-cbind(UR1,UR2)#Collect as Matrix
#Library for time series regression
library(dynlm)
#Regress indepent series
spuriousreg<-dynlm(formula=yUR1~yUR2,data=simulatedseries)
spursum<-summary(spuriousreg)

Simulation (Code 3)

#Plot
ts.plot(simulatedseries, gpars=
  list(main="Two Independent Simulated Random Walks", 
  ylab="Series value",lty=c(1,2),col=c("red","blue")))
  legend("bottomleft",bty="n",
  expression("Random Walk 1", "Random Walk 2"),
  lty=1:2,col=c("red","blue"))
text(60,2,"Coef= ") 
text(78,2,spursum$coef[2])
text(60,1,"t-value= ") 
text(78,1,spursum$coef[6])
text(60,0,"p-value= ") 
text(81,0,spursum$coef[8])

Simulation: 2 independent random walks

Explanation: Spurious Regression

Failing to include trend if needed, or account for integration, creates bias in regression
- Will get strong relationship (small p value, high \(R^2\)) if both series trending
- Called “spurious” or “nonsense” correlation
Two trending series regressed on each other will be correlated, even if no structural relationship
- “Time” is omitted variable
- Easy to see this with deterministic trend
- Is just as much of a problem with stochastic trend

Multiple integrated series

If \(X_t\) has unit root, and \(Y_t\) is linear function of it, \(Y_t\) must also have unit root
If tests say \(X_t\) has unit root, \(Y_t\) doesn’t, then level of \(Y_t\) can’t be linear function of level of \(X_t\)
- But may be function of growth in \(X\), or nonlinear function of level
E.g., unemployment \(\in [0,1]\) so bounded, can’t be linear unit root per se
- Farmer (2014) argues nonlinear function of it is unit root
Want to know: is it possible to find a real relationship between multiple persistent series?

Non-spurious relationships

How do we think about relationships between nonstationary series?
Consider \(Y_t=Y_{t-1}+e_t\) and \(X_t=X_{t-1}+u_t\) two series with unit roots
- E.g., price of two different stocks
If \(u_t\perp e_t\) can learn this from regressing \(\Delta Y_t=e_t\) on \(\Delta X_t=u_t\)
- Coefficient will be 0
If \(u_t\) and \(e_t\) correlated can run same regression and get nonzero coefficient
- Growth of \(X\) associated with growth of \(Y\)
Unless correlation is perfect, due to unit root behavior, series will eventually drift arbitrarily far from each other
Relationship between differences captures short run relationship only

Cointegration

Some series may instead have long run relationship
- Even though levels nonstationary, levels remain closely related over time
A model for such relationships is that the levels are related
- \(Y_t=\beta X_t +\epsilon_t\) but now \(\epsilon_t\) is stationary
If 2 (or more) series each have unit root, but linear combination of them, e.g. \(Y_t-\beta X_t\) is stationary, the series are called cointegrated
Example: price of shares of same stock traded on two different exchanges, say New York & Chicago
- Price has unit root, but two prices should not differ for longer than it takes buy/sell orders to travel from NY to Chicago
Metaphor: “A drunk walking their dog”: drunk moves in random direction from position \(X_t\) but holds dog on a leash at position \(Y_t\)
Can’t predict long run location of pair, but dog will stay within leash length of the drunk

Testing for cointegration

We would really like to use cointegration relationship:
- E.g., to know when to buy in NY and sell in Chicago
This “pairs trading” idea is one way high frequency trading makes money
Problem is, how to know if it’s real or spurious? Even if no relationship, regression of \(Y\) on \(X\) will give significant \(\beta\)
This is problem of cointegration testing
Also want to estimate \(\beta\) and use relationship for prediction
If tests suggest \(Y\) and \(X\) have unit roots, should test for cointegration before including in a regression

Lettau and Ludvigson: Code 1

#Load data on c, a, y from Lettau's website
library(readr) #Library to read .csv files
#download data from Lettau's website
cay_current<-read_csv(
  "http://faculty.haas.berkeley.edu/
  lettau/data/cay_current.csv",
  skip=1)
#Rename c to remove space
cay_current$c<-cay_current$`c (pce)`

Lettau and Ludvigson: Code 2

#Test for unit roots in each series, 
# under null with possible drift and trend
library(tseries)
#Run Phillips-Perron tests for each series
c.pp<-pp.test(cay_current$c)
a.pp<-pp.test(cay_current$a)
y.pp<-pp.test(cay_current$y)
#Set as time series for use in commands
cc<-as.ts(cay_current$c)
aa<-as.ts(cay_current$a)
yy<-as.ts(cay_current$y)

Macroeconomic Example: Lettau & Ludvigson (2004)

Long run budget constraint suggests wealth and spending cannot diverge permanently
If wealth has unit root and consumption does too, one should eventually return to the other
L&L examine data on aggregate US consumption \(c_t\), measure total wealth by financial wealth \(a_t\) and labor income \(y_t\)
Theory is that financial plus (present discounted) value of labor income make up total wealth, so these three series, if integrated, should be jointly cointegrated
Long run relationship will measure “wealth effect” of permanent wealth changes on consumer spending
Phillips Perron p-values for null of unit root with possible drift and trend are
- c: 0.8591813 , a: 0.1086362, y: 0.9128739
None of these rejects null of unit root

Cointegration tests

Difference from cointegrated \(Y_t=\beta X_t +\epsilon_t\) and spurious relationship is that \(\epsilon_t\) is stationary if cointegrated
So, if \(Y\) and \(X\) known to have unit roots (say, by tests) and \(\beta\) known, just run unit root test on \(\epsilon_t=Y_t-\beta X_t\)
Under \(H_0\) relationship is spurious (and you should difference series), under \(H_1\) they are cointegrated
Use usual unit root tests (ADF, PP, etc) with appropriate corrections for trends, serial correlation, etc
If \(\beta\) not known, can replace by OLS estimate, run unit root test on \(Y_t-\widehat{\beta}X_t\)
Since \(\widehat{\beta}\) not consistent under \(H_0\), critical values are weird, but software will calculate them for you
Version using ADF statistic called Engle-Granger test, version with robust PP statistic called Phillips-Ouliaris test
All in urca (or tseries) library in R
For more than 2 series, there can be more than one stationary cointegrating relationship
Won’t describe, but Johansen test can determine this

Application: L&L data (Code)

#Run OLS
caylm<-lm(cc~aa+yy)
library(stargazer)
stargazer(caylm,header=FALSE,type="html",
  no.space=TRUE,report="vc",omit.table.layout = "ns-=a")

#Show whether residual appears to have unit root
cayresid<-as.ts(caylm$residuals)
autoregression<-dynlm(cayresid~L(cayresid))
stargazer(autoregression,header=FALSE,type="html",
    no.space=TRUE,report="vc",
    omit.table.layout = "ns-=a")

Application: L&L data

OLS regression of \(c_t\) on \(a_t\), \(y_t\) shows series appear related, but want to be sure it’s not spurious


	Dependent variable:

	cc
aa	0.249
yy	0.782
Constant	-0.616

Regress residual on \(L \text{ residual}\) to see if persistent


	Dependent variable:

	cayresid
L(cayresid)	0.922
Constant	-0.0002

Formal test of cointegration (Code)

#Run Phillips-Ouliaris test on c,a,y
caytest<-po.test(data.frame(cc,aa,yy))
caytest

Formal test of cointegration

Apply Phillips-Ouliaris test

## 
##  Phillips-Ouliaris Cointegration Test
## 
## data:  data.frame(cc, aa, yy)
## Phillips-Ouliaris demeaned = -14.98, Truncation lag parameter = 2,
## p-value = 0.15

Can’t reject \(H_0\) of unit root in residuals
- Can’t reject no cointegration
- Might be spurious relationship
L&L paper ran several other tests, showing cointegration, with slightly different sample

Estimation under cointegration

If \(Y\), \(X\) cointegrated, \(Y_t=\beta X_t +\epsilon_t\) can estimate \(\beta\)
OLS is consistent, converges fast (at rate \(t\) instead of \(\sqrt{t}\))
But may be biased and non-normal, so other estimators preferred
Book describes leads and lags (also called dynamic OLS or Stock Watson) estimator
Regress \(Y_t\) on \(X_t\) and \(\Delta X_t\) and leads and lags of \(\Delta X_t\)
Result again converges fast, but now normal up to needing HAC standard errors

Dynamic OLS estimates (Code)

#Dynamic OLS with 4 leads and lags
dynOLS<-dynlm(cc~aa+yy+L(diff(aa),-4:4)+L(diff(yy),-4:4))
library(lmtest)
library(sandwich)
dynresults<-coeftest(dynOLS,vcovHAC) #HAC errors
stargazer(dynresults,header=FALSE,type="html",
    keep=c(1,2,21), no.space=TRUE,
    omit.table.layout = "ns-=a")

Dynamic OLS estimates of cointegrating coefficients

Assuming cointegration tests hadn’t failed, could estimate by
Regress \(c_t\) on \(y_t\), \(a_t\) and 4 leads and lags of \(\Delta y_t\), \(\Delta a_t\)
Using HAC standard errors t-stats are asymptotically normal


	Dependent variable:


aa	0.273^***
	(0.052)
yy	0.759^***
	(0.057)
Constant	-0.671^***
	(0.115)

Combining long run and short run relationships

Since \(Y_t-\beta X_t\) stationary, can be used in OLS as a predictor
Replacing with \(Y_t-\widehat{\beta} X_t\) from above method fine since \(\widehat{\beta}\) converges faster than OLS
One use is error correction model (ECM)
Predict \(\Delta Y_t\) using cointegrating residual and other stationary predictors such as lags of \(\Delta X_t\) and \(\Delta Y_t\)
\(\Delta Y_t = \pi_0+\pi_1 (Y_{t-1}-\widehat{\beta}X_{t-1}) +\pi_2 \Delta X_t + \pi_3 \Delta Y_{t-1} +\pi_4 \Delta X_{t-1}+\ldots+u_t\)
\(\pi_1\) measures how fast \(Y\) moves back to long run equilibrium
Estimate above by OLS: everything stationary, so fine up to usual concerns about HAC SEs
For stocks, might be large: return in milliseconds. For macroeconomic variables, may be small, takes several years
Can estimate also \(\beta\) and error correction coefficients jointly by Johansen procedure: a form of nonlinear MLE

Error correction model using L&L residuals (Code)

#Calculate cointegrating residual from 
#   dynamic OLS estimated coefficients
cay<-cc-dynOLS$coefficients[2]*aa
      -dynOLS$coefficients[3]*yy
      -dynOLS$coefficients[1]
#Run error correction regression
ecm<-dynlm(d(aa)~L(cay)+L(d(aa))+L(d(cc))+L(d(yy)))
#Calculate HAC standard errors
ecmresults<-coeftest(ecm,vcovHAC)
#Display Results
stargazer(ecmresults,header=FALSE,type="html", 
    no.space=TRUE,omit.table.layout = "ns-=a")

Error correction model using L&L residuals

If a variable is cointegrated, then deviations from trend revert back
Including (actually) cointegrating residual can lead to better forecasts
Try to predict \(\Delta a_t\) using lags and \(c_{t-1}-b_0-b_1a_{t-1}-b_2y_{t-1}\)


	Dependent variable:


L(cay)	0.111^*
	(0.059)
L(d(aa))	0.226^**
	(0.110)
L(d(cc))	0.265
	(0.161)
L(d(yy))	-0.153
	(0.152)
Constant	0.004^**
	(0.002)

Conclusions

Regressions using persistent time series can cause appearance of strong relationships even if absent
Spurious correlations due to (possibly stochastic) time trends in each series
If series are persistent, can test whether relationship spurious by seeing if residuals are persistent
If not, series may be cointegrated: sharing a common stochastic trend
- Signature is that linear combination of nonstationary series is stationary
Cointegration relationship can be tested for, and estimated when it exists

Spurious Relationships and Cointegration

73-374 Econometrics II

Today

Some Important Data Relationships (1)

Some Important Data Relationships (2)

Some Important Data Relationships (3)

What’s going on here?

Simulation (Code 1)

Simulation (Code 2)

Simulation (Code 3)

Simulation: 2 independent random walks

Explanation: Spurious Regression

Multiple integrated series

Non-spurious relationships

Cointegration

Testing for cointegration

Lettau and Ludvigson: Code 1

Lettau and Ludvigson: Code 2

Macroeconomic Example: Lettau & Ludvigson (2004)

Cointegration tests

Application: L&L data (Code)

Application: L&L data

Formal test of cointegration (Code)

Formal test of cointegration

Estimation under cointegration

Dynamic OLS estimates (Code)

Dynamic OLS estimates of cointegrating coefficients

Combining long run and short run relationships

Error correction model using L&L residuals (Code)

Error correction model using L&L residuals

Conclusions