INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Lesson 21: Univariate and Multi-

Variate Models
Module: Advanced Time-Series Models
Introduction

• Univariate models of prediction and forecasting

• White noise iid process
• Stationarity of AR processes
• Unconditional mean of the AR process
• ACF and PACF plots: AR process
• Autocorrelation coefficient
• Autocorrelation coefficient: numerical example
Introduction

• Moving average (MA) processes

• A simple example of MA(2) process
• ACF and PACF plots: MA process
• ARMA process
• Building ARMA model
Introduction

• Determining the accuracy of the forecast

• Introduction to time-series stationarity
• Time series and stationarity
• Stochastic non-stationarity
• Deterministic non-stationarity
• Time-series and stationarity: visual examination
Introduction

• Testing for unit-roots

• Introduction to mean reversion
• Cointegration
• Error correction models (ECM)
• Engle-Granger approach to ECM and cointegration
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Univariate Models of Prediction and

Forecasting
Univariate Models of Prediction and
Forecasting
• Univariate time-series models are a class of specifications where one
attempts to model and predict financial variables using only information
contained in their own past values and possibly current and past values of an
error term
• Time series models are usually a-theoretical. (HFT trading)
• An important class of models is autoregressive integrated moving average
(ARIMA) models.
• Often in HFT, fundamental variables affecting 𝑃𝑡 are not available in the same
frequency as 𝑃𝑡 .
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Introduction to Time-Series
Stationarity
Time-Series and Stationarity

• A strictly stationary process

• A time series process {Xt} is said to be strictly stationary if the joint probability
distribution of a sequence {yt1 to ytn} is the same as that for {yt1+m to ytn+m}  m: i.e.,
𝐹 𝑦𝑡1 , … , 𝑦𝑡𝑛 = 𝐹 𝑦𝑡1+𝑚 , … , 𝑦𝑡𝑛+𝑚
Time-Series and Stationarity

• A weakly stationary process

• A weakly stationary (or covariance stationary) process satisfies the following
three equations:
• 𝐸 𝑦𝑡 = 𝑢, 𝑡 = 1,2,3 … ∞
• 𝐸 𝑦𝑡 − 𝑢 𝑦𝑡 − 𝑢 = 𝜎 2 < ∞
• 𝐸 𝑦𝑡1 − 𝑢 𝑦𝑡2 − 𝑢 = γ𝑡 2−𝑡1 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡1 𝑎𝑛𝑑 𝑡2
• To summarize, a weak stationary process should have a constant mean,
constant variance, and a constant auto-covariance structure.
• What is auto-covariance: 𝐸 𝑦𝑡 − 𝑢 𝑦𝑡+𝑠 − 𝑢 = γ𝑠 ; s = 0,1,2, ...
γ𝑠
• Autocorrelation: 𝜏𝑠 = , s = 0,1,2, ... ; 𝜏𝑠 lies in the interval ±1
γ0
Time-Series and Stationarity

• Problems with non-stationary series

• Shocks to non-stationary systems do not die away and are highly persistent
• Estimation with non-stationary data can lead to spurious regressions, i.e., artificially high t-stat.
and adjusted R-sq, even if the variables are not related to each other. t-ratio will not follow t-
distribution
Adj. R sq T-stat

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

White Noise IID Process

White Noise IID Process

• It can be defined as follows

1. E(𝜇𝑡 ) = 0 , t = 1,2,...,
2. Var(𝜇𝑡 )= 𝜎 2
3. 𝐸[ 𝜇𝑡𝑘 − 𝑢 𝜇𝑡𝑚 − 𝑢 ] = γ𝑡 𝑘 −𝑡𝑚 = 0 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡𝑘
𝑎𝑛𝑑 𝑡𝑚 for all 𝑡 𝑘 ≠ 𝑡𝑚
• Thus, white noise has a constant (zero) mean, finite constant variance, and
is uncorrelated across time
• Often, we assume a fix distribution (e.g., normal iid), then it is a special case
of white noise process
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Introduction to AR Processes
Autoregressive (AR) Processes

• Current value of the variable 𝑦𝑡 depends upon its own previous

values
• An AR model of order p, AR(p) can be expressed as below
• 𝑦𝑡 = 𝜇 + 𝜑1 𝑦𝑡−1 + 𝜑2 𝑦𝑡−2 + ⋯ + 𝜑𝑝 𝑦𝑡−𝑝 + 𝑢𝑡
𝑝 𝑝
• 𝑦𝑡 = 𝜇 + σ𝑖=1 𝜑𝑖 𝑦𝑡−𝑖 + 𝑢𝑡 or 𝑦𝑡 = 𝜇 + σ𝑖=1 𝜑𝑖 𝐿𝑖 𝑦𝑡 + 𝑢𝑡
• 𝜑(𝐿)𝑦𝑡 = 𝜇 + 𝑢𝑡 ; where 𝜑 𝐿 = 1 − 𝜑1 𝐿 − 𝜑2 𝐿2 − ⋯ − 𝜑𝑝 𝐿𝑝
• Here 𝑢𝑡 is the white noise iid error term [distributed as N (0, 𝜎 2 )]
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Stationarity of AR Processes
Stationarity of AR Processes

• 𝜑(𝐿)𝑦𝑡 = 𝜇 + 𝑢𝑡 ; where 𝜑 𝐿 = 1 − 𝜑1 𝐿 − 𝜑2 𝐿2 − ⋯ − 𝜑𝑝 𝐿𝑝

• 𝑦𝑡 = 𝜑 𝐿 −1 𝑢
𝑡 [Ignore the drift]
• As the lag length is increased, the function 𝜑 𝐿 −1 should converge to zero
• If we compute the roots of this characteristic equation

• 𝜑 𝐿 = 1 − 𝜑1 𝐿 − 𝜑2 𝐿2 − ⋯ − 𝜑𝑝 𝐿𝑝 = 0

• What if a root is more or less than 1, what if it is equal to 1

• This is non-stationarity
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Stationarity of AR Processes: A simple

Example
Stationarity of AR Processes

• For example, consider the process below

• 𝑦𝑡 = 𝑦𝑡−1 + 𝑢𝑡
• 1 − 𝐿 𝑌𝑡 = 𝑢𝑡

• Characteristic equation =1-z=0

• What do we infer?
Unconditional Mean of the AR process

• If the AR (p) process is stationary the unconditional mean of the

process is given below
𝜇
• 𝐸 𝑦𝑡 =
1−𝜑1 −𝜑2 −⋯−𝜑𝑝

• For example, for AR(1) process: 𝑦𝑡 = 𝜇 + 𝜑1 𝑦𝑡−1 + 𝑢𝑡

𝜇
• 𝐸 𝑦𝑡 =
1−𝜑1
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Autocorrelation Coefficient
Autocorrelation Coefficient

• 𝐸[ 𝑦𝑡 − 𝐸 𝑦𝑡 𝑦𝑡+𝑠 − 𝐸 𝑦𝑡+𝑠 ] = 𝛾𝑠 ; s = 0,1,2, ….

• When s = 0, the autocovariance at lag zero is obtained, which is the
variance of 𝑦𝑡 ; 𝛾0
• Autocovariances also depend on the units of measurement and are
not easy to interpret directly
𝛾𝑠
• Autocorrelation measure 𝜏𝑠 =
𝛾0

• These autocorrelations (𝜏𝑠 ) lie in the interval ±1

Autocorrelation Coefficient

• 𝜏𝑠 are ~ (approximately or asymptotically) normally distributed.

𝜏ෝ𝑠 ~ 𝑁(0,1/𝑇 )
• For example, a 95% non-rejection region would be given by the
following interval: −1.96 ∗ 1/ 𝑇 to 1.96 ∗ 1/ 𝑇
• If the sample autocorrelation coefficient, 𝜏ෝ𝑠 , falls outside this
region, for a given lag s, then the null hypothesis that the true
value is zero (at lag s) is rejected
Autocorrelation Coefficient

• Box–Pierce (BP) test stat.: 𝑄 = 𝑇 σ𝑚 𝜏

𝑘=1 𝑘
2

2
𝜏𝑘
• Ljung–Box (LB) statistic 𝑄 ∗ = 𝑇(𝑇 + 𝑚
2) σ𝑘=1
𝑇−𝑘

• These statistics are asymptotically distributed as 𝜒𝑚

2

• What happens if T increases towards infinity

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Autocorrelation Coefficient:
Numerical Example
Autocorrelation Coefficient

• Consider a series of 100 observations. At lags, 1, 2, 3, 4, 5, you

observe the following auto-correlation coefficients: 𝜏1 =
0.207; 𝜏2 = −0.013; 𝜏3 = 0.086; 𝜏4 = 0.005, 𝑎𝑛𝑑 𝜏5 = −0.022 .
We will test these coefficients individually, using Box-Pierce (BP)
test and jointly Ljung-Box (LB) test.
1
• (1) 95% confidence interval : =±1.96 ∗
𝑇

• (2) Using BP and LB test conduct the joint hypothesis

Autocorrelation Coefficient

• (1) 95% confidence interval

• For each coefficient, we can construct the confidence interval
1
=±1.96 ∗ ; here T=100. Thus, the confidence interval lies in
𝑇
between -0.196 to +0.196.
• Let us come to the joint test. The null hypothesis is that all the first
five coefficients are jointly zero. That is, 𝐻0 : 𝜏1 = 0; 𝜏2 = 0; 𝜏3 =
0; 𝜏4 = 0, 𝑎𝑛𝑑 𝜏5 = 0.
Autocorrelation Coefficient

• Let us come to the joint test. The null hypothesis is that all the first
five coefficients are jointly zero. That is, 𝐻0 : 𝜏1 = 0; 𝜏2 = 0; 𝜏3 =
0; 𝜏4 = 0, 𝑎𝑛𝑑 𝜏5 = 0. The relevant critical values are obtained
from 𝜒 2 distribution with five degrees of freedom. These are 11.1
at 5% level, 15.1 at 1% level.
• The test statistic for BP test is given below. 𝑄 = 𝑇 σ𝑚 𝜏
𝑘=1 𝑘
2

2
𝜏𝑘
• The LB statistics is computed as follows. 𝑇(𝑇 + 2) σ𝑚
𝑘=1 𝑇−𝑘
Autocorrelation Coefficient

• Let us come to the joint test. The null hypothesis is that all the first five
coefficients are jointly zero. That is, 𝐻0 : 𝜏1 = 0; 𝜏2 = 0; 𝜏3 = 0; 𝜏4 =
0, 𝑎𝑛𝑑 𝜏5 = 0. The relevant critical values are obtained from 𝜒 2 distribution
with five degrees of freedom. These are 11.1 at 5% level, 15.1 at 1% level.
• The test statistic for BP test is given below.
𝐵𝑃 = 100 ∗ (0.2072 + −0.013 2 + 0.0862 + 0.0052 + −0.022 2 ) = 5.09.
0.2072
• The LB statistics is computed as follows. 𝐿𝐵 = 100 ∗ 100 + 2 ∗ ቀ +
100−1
−0.013 2 0.0862 0.0052 −0.022 2
+ + + ቁ = 5.26
100−2 100−3 100−4 100−5
Autocorrelation Coefficient

• So, in both cases, the joint null hypothesis that all of the first five
autocorrelation coefficients are zero cannot be rejected. Note
that, in this instance, the individual test caused a rejection while
the joint test did not.
• Thus, the effect of the significant autocorrelation coefficient is
diluted in the joint test by the insignificant coefficients.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

ACF and PACF Plots: AR Process

ACF and PACF Plots: AR Process

• Auto-correlation function (ACF) plot and partial ACF plot

• An AR process has
• A geometrically decaying acf
• A number of non-zero points of pacf = AR order

Chris Brooks; Introductory Econometrics for Finance. 4th Edition. Chapter 6

ACF and PACF Plots: AR Process

• Auto-correlation function (ACF) plot and partial ACF plot

• An AR process has
• A geometrically decaying acf
• A number of non-zero points of pacf = AR order

Chris Brooks; Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Moving Average (MA) Processes

Moving Average (MA) Processes

• Current value of the variable 𝑦𝑡 depends upon the error terms

• An MA model of order q, MA(q) can be expressed as below
• 𝑦𝑡 = 𝜇 + 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 + ⋯ + 𝜃𝑞 𝑢𝑡−𝑞
𝑞
• 𝑦𝑡 = 𝜇 + σ𝑖=1 𝜃𝑖 𝑢𝑡−𝑖 + 𝑢𝑡 or in terms of lag operator
• 𝑦𝑡 = 𝜇 + 𝜃 𝐿 𝑢𝑡 ; 𝑤ℎ𝑒𝑟𝑒 𝜃 𝐿 = 1 + 𝜃1 𝐿 + 𝜃2 𝐿2 + ⋯ + 𝜃𝑞 𝐿𝑞
• Here 𝑢𝑡 is the white noise IID error term [distributed as N (0, 𝜎 2 )]
• A moving average model is simply a linear combination of white noise
processes
Moving Average (MA) Processes

• The following properties define this moving average [MA(q)] process

• 𝑦𝑡 = 𝜇 + 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 + ⋯ + 𝜃𝑞 𝑢𝑡−𝑞

• 𝐸(𝑦𝑡 ) = 𝜇
• Var(𝑦𝑡 ) = 𝛾0 = (1 + 𝜃12 + 𝜃22 + 𝜃32 + ⋯ + 𝜃𝑞2 ) 𝜎 2

𝜃𝑠 + 𝜃𝑠+1 𝜃1 + 𝜃𝑠+2 𝜃2 + ⋯ + 𝜃𝑞 𝜃𝑠 𝜎 2 ; 𝑓𝑜𝑟 𝑠 = 1,2,3. . 𝑞

• Auto-Cov.=(𝛾𝑠 )=൝
0 𝑓𝑜𝑟 𝑠 > 𝑞

• An MA (q) process has constant mean, constant variance, and auto-

covariances that are non-zero up to lag ‘q’ and zero thereafter.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

A Simple Example of MA(2) Process:

Part I
Example MA (2) Processes: Variance

• The following properties define this moving average [MA(q)] process

• 𝑦𝑡 = 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2
• 𝐸(𝑦𝑡 ) = 𝐸[𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 ]=𝐸(𝑢𝑡 ) + 𝜃1 𝐸(𝑢𝑡−1 ) + 𝜃2 𝐸(𝑢𝑡−2 )=0
• Var(𝑦𝑡 ) = 𝐸 𝑦𝑡 − 𝐸 𝑦𝑡 𝑦𝑡 − 𝐸 𝑦𝑡 = 𝐸 𝑦𝑡2
• 𝐸 𝑦𝑡2 = 𝐸[(𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 )(𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 )]
• 𝐸 𝑦𝑡2 = 𝐸[(𝑢𝑡2 + 𝜃12 𝑢𝑡−1
2
+ 𝜃22 𝑢𝑡−2
2
+ 𝐶𝑟𝑜𝑠𝑠 𝑝𝑟𝑜𝑑𝑢𝑐𝑡𝑠]
• 𝐸 𝐶𝑟𝑜𝑠𝑠 − 𝑝𝑟𝑜𝑑𝑢𝑐𝑡 =? [𝐻𝑖𝑛𝑡: 𝐸 𝑢𝑡 ∗ 𝑢𝑡−𝑠 = 𝐶𝑜𝑣 𝑢𝑡 , 𝑢𝑡−𝑠 =? For s≠0
• For s=0; 𝐸 𝑢𝑡 ∗ 𝑢𝑡 =𝐶𝑜𝑣 𝑢𝑡 , 𝑢𝑡 =𝑉𝑎𝑟 𝑢𝑡 , 𝑢𝑡 = 𝜎 2
Example MA (2) Processes: Variance

• The following properties define this moving average [MA(q)] process

• 𝑦𝑡 = 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2

• 𝐸(𝑦𝑡 ) = 𝐸[𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 ]=𝐸(𝑢𝑡 ) + 𝜃1 𝐸(𝑢𝑡−1 ) + 𝜃2 𝐸(𝑢𝑡−2 )=0

• Var(𝑦𝑡 ) = 𝐸 𝑦𝑡 − 𝐸 𝑦𝑡 𝑦𝑡 − 𝐸 𝑦𝑡

• 𝐸 𝑦𝑡2 = 𝐸[ 𝑢𝑡2 + 𝜃12 𝑢𝑡−1

2
+ 𝜃22 𝑢𝑡−2
2
= (1 + 𝜃12 + 𝜃22 ) 𝜎 2
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

A Simple Example of MA(2) Process:

Part 2
Example MA (2) Processes: Auto-Covariance

• The following properties define this moving average [MA(q)] process

• 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2
• 𝛾1 = 𝐸 𝑦𝑡 − 𝐸 𝑦𝑡 𝑦𝑡−1 − 𝐸 𝑦𝑡−1 = 𝐸[𝑦𝑡 ∗ 𝑦𝑡−1 ]
• 𝛾1 = 𝐸 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 𝑢𝑡−1 + 𝜃1 𝑢𝑡−2 + 𝜃2 𝑢𝑡−3
2 2
• 𝛾1 = 𝐸 𝜃1 𝑢𝑡−1 + 𝜃1 𝜃2 𝑢𝑡−2 + 𝐶𝑟𝑜𝑠𝑠 𝑃𝑟𝑜𝑑𝑢𝑐𝑡 𝑡𝑒𝑟𝑚𝑠
• 𝛾1 = (𝜃1 + 𝜃1 𝜃2 ) 𝜎 2
• 𝛾2 = (𝜃2 ) 𝜎 2
• 𝛾3 = 0
Example MA (2) Processes: Auto-Correlation

𝛾0
• 𝜏0 = =1
𝛾0

𝛾1 𝜃1 +𝜃1 𝜃2 𝜎 2 𝜃1 +𝜃1 𝜃2
• 𝜏1 = = =
𝛾0 (1+𝜃12 +𝜃22 ) 𝜎 2 (1+𝜃12 +𝜃22 )

𝛾𝑠 𝜃2
• 𝜏2 = =
𝛾0 (1+𝜃12 +𝜃22 )
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

ACF and PACF Plots: MA Process

ACF and PACF Plots: MA Process

• Auto-correlation function (ACF) plot and partial ACF plot

• An MA process has
• Number of non-zero points of acf = MA order
• A geometrically decaying pacf.

Chris Brooks; Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Lesson 22: Introduction to ARMA

process

Module: Advanced Time-Series Models

ARMA Process

• By combining the AR(p)and MA(q)models, an ARMA(p, q)model is obtained

• 𝑦𝑡 = 𝜇 + 𝜑1 𝑦𝑡−1 + 𝜑2 𝑦𝑡−2 + ⋯ + 𝜑𝑝 𝑦𝑡−𝑝 + 𝑢𝑡 + 𝜃1 𝑢𝑡−1 + 𝜃2 𝑢𝑡−2 + ⋯ +
𝜃𝑞 𝑢𝑡−𝑞

• 𝜑(𝐿)𝑦𝑡 = 𝜇 + 𝜃 𝐿 𝑢𝑡

• Where 𝜃 𝐿 = 1 + 𝜃1 𝐿 + 𝜃2 𝐿2 + ⋯ + 𝜃𝑞 𝐿𝑞

• 𝜑 𝐿 = 1 − 𝜑1 𝐿 − 𝜑2 𝐿2 − ⋯ − 𝜑𝑝 𝐿𝑝

• Also, 𝐸 𝑢𝑡 = 0; 𝐸 𝑢𝑡2 = 𝜎 2 ; 𝐸 𝑢𝑡 𝑢𝑠 = 0; 𝑡 ≠ 𝑠
ACF and PACF Plots: ARMA Process

• Auto-correlation function
(ACF) plot and partial ACF
plot
• An ARMA process has
• A geometrically decaying acf
• A geometrically decaying pacf

Chris Brooks; Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Building ARMA Model

Building the ARMA Model for Price Prediction

• Identification: Determine the order of the process, that is, p and q

values for ARMA (p,q)
• Estimation: Estimate the parameters with OLS/MLE
• Model Diagnostics: Testing the model. Residual diagnostics:
Clean iid residuals
• Objective: A parsimonious model removes irrelevant lags of AR
and MA terms.
• Information criteria for model/lag selection
Information Criteria for ARMA Model Selection

• Akaike’s (1974) information criterion (AIC)=−2 𝑙𝑜𝑔 𝐿 + 2𝐾

• Schwarz’s (1978) Bayesian information criterion
(SBIC)=−2 𝑙𝑜𝑔 𝐿 + 𝐾 ∗ 𝑙𝑜𝑔(𝑇)
• Hannan–Quinn criterion (HQIC)=−2 𝑙𝑜𝑔 𝐿 + 2𝐾𝑙𝑜𝑔(𝑙𝑜𝑔 𝑇 )
𝑘
• Where 𝐾 = = k is number of parameters, T is the sample size,
𝑇
and - 𝑙𝑜𝑔(L) is log-likelihood of observing the parameters
obtained from the model
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Forecasting with time-series models

Time-series forecasting

• Time-series models use conditional expectations 𝐸(𝑦𝑡+1 |Ω𝑡 ):

expected value of ‘y’ at ‘t+1’ given all the information up to ‘t’ (Ω𝑡 )
• For a zero mean white noise process 𝐸 𝜇𝑡+1 Ω𝑡 = 0 ∀ 𝑠 > 0
• Naïve Forecasting: 𝐸(𝑦𝑡+1 |Ω𝑡 )= 𝑦𝑡 ; or random walk process (no
change forecast
• The unconditional expectation is the unconditional mean of ‘y’
with out any time reference (long-term mean)
• For mean-reverting stationary process, the long-term average
becomes the forecast
Problems with structural models

• 𝑦𝑡 = 𝛽1 + 𝛽2 𝑥2𝑡 + 𝛽3 𝑥3𝑡 + ⋯ + 𝛽𝑘 𝑥𝑘𝑡 + 𝑢𝑡 : conditional forecasts

• 𝐸(𝑦𝑡 |Ω𝑡−1 ) = 𝛽1 + 𝛽2 𝐸(𝑥2𝑡 ) + 𝛽3 𝐸(𝑥3𝑡 ) + ⋯ + 𝛽𝑘 𝐸(𝑥𝑘𝑡 )
• To forecasts conditional expectations for y, one needs forecasts of
x’s, i.e., 𝐸(𝑥𝑘𝑡 )
• This makes the process cumbersome and complex
• One may have to look the historical values of x’s to forecast the
current values; however, this can be directly captured in the
historical time-series values of 𝑦𝑡 itself
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Forecasting with ARMA Models: Part I

Forecasting with ARMA models

• Forecast from ARMA(p,q) model at time ‘t’ for ‘s’ steps into the
future is given as
𝑝 𝑞
• 𝑓𝑡+𝑠 = Σ𝑖=1 𝑎𝑖 𝑓𝑡+𝑠−𝑖 + Σ𝑘=1 𝑏𝑘 𝑢𝑡+𝑠−𝑘
• Here, 𝑎𝑖 and 𝑏𝑘 are the autoregressive and moving average
coefficients
Forecasting MA(q)

• Let us look at MA(3) model: 𝑦𝑡 = 𝜇 + 𝜃1 𝜇𝑡−1 + 𝜃2 𝜇𝑡−2 +

𝜃3 𝜇𝑡−3 + 𝜇𝑡
• Assuming parameter constancy (i.e., the relationship holds), then
• 𝑦𝑡+1 = 𝜇 + 𝜃1 𝜇𝑡 + 𝜃2 𝜇𝑡−1 + 𝜃3 𝜇𝑡−2 + 𝜇𝑡+1
• 𝑓𝑡,1 = 𝐸 𝑦𝑡+1 |Ω𝑡 = 𝐸(𝜇 + 𝜃1 𝜇𝑡 + 𝜃2 𝜇𝑡−1 + 𝜃3 𝜇𝑡−2 + 𝜇𝑡+1 |Ω𝑡 )
• The values of error terms up to time ‘t’ is known, but after that we
have to take their conditional expectation, which is zero
• That is 𝐸 𝜇𝑡+1 |Ω𝑡 =0
Forecasting MA(q)

• That is 𝐸 𝜇𝑡+1 |Ω𝑡 =0; therefore

• 𝑓𝑡,1 = 𝐸 𝑦𝑡+1 |Ω𝑡 = 𝜇 + 𝜃1 𝜇𝑡 + 𝜃2 𝜇𝑡−1 + 𝜃3 𝜇𝑡−2
• 𝑓𝑡,2 = 𝐸 𝑦𝑡+2 |Ω𝑡 = 𝐸(𝜇 + 𝜃1 𝜇𝑡+1 + 𝜃2 𝜇𝑡 + 𝜃3 𝜇𝑡−1 Ω𝑡 = 𝑢 +
𝜃2 𝜇𝑡 + 𝜃3 𝜇𝑡−1
• 𝑓𝑡,3 = 𝐸 𝑦𝑡+3 |Ω𝑡 = 𝐸(𝜇 + 𝜃1 𝜇𝑡+2 + 𝜃2 𝜇𝑡+1 + 𝜃3 𝜇𝑡 Ω𝑡 = 𝑢 +
𝜃3 𝜇𝑡
• 𝑓𝑡,4 = 𝐸 𝑦𝑡+4 |Ω𝑡 = 𝐸(𝜇 + 𝜃1 𝜇𝑡+3 + 𝜃2 𝜇𝑡+2 + 𝜃3 𝜇𝑡+1 Ω𝑡 = 𝑢
Forecasting MA(q)

• Since the MA(3) process has a memory of only three periods, any
forecast of four or more steps ahead converge to long-term
unconditional mean (i.e., the intercept term)
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Forecasting with ARMA Models: Part II

Forecasting AR(p) process

• Consider AR(2) process

• 𝑦𝑡 = 𝜇 + 𝜙1 𝑦𝑡−1 + 𝜙2 𝑦𝑡−2 + 𝑢𝑡 ; unlike MA process, AR process
has infinite memory
• The ‘t+1’ forecast is obtained as:
• 𝑦𝑡+1 = 𝜇 + 𝜙1 𝑦𝑡 + 𝜙2 𝑦𝑡−1 + 𝑢𝑡+1 , then
• 𝑓𝑡,1 = 𝐸 𝑦𝑡+1 Ω𝑡 = 𝐸(𝜇 + 𝜙1 𝑦𝑡 + 𝜙2 𝑦𝑡−1 + 𝑢𝑡+1 |Ω𝑡 )=𝜇 + 𝜙1 𝑦𝑡 +
𝜙2 𝑦𝑡−1 (since actual values of 𝑦𝑡 and 𝑦𝑡−1 are observed)
Forecasting AR(p) process

• 𝑓𝑡,1 = 𝐸 𝑦𝑡+1 Ω𝑡 = 𝐸(𝜇 + 𝜙1 𝑦𝑡 + 𝜙2 𝑦𝑡−1 + 𝑢𝑡+1 |Ω𝑡 )=𝜇 + 𝜙1 𝑦𝑡 +

𝜙2 𝑦𝑡−1 (since actual values of 𝑦𝑡 and 𝑦𝑡−1 are observed)
• Similarly, for next steps 2 and 3
• 𝑓𝑡,2 = 𝐸 𝑦𝑡+2 Ω𝑡 = 𝐸(𝜇 + 𝜙1 𝑦𝑡+1 + 𝜙2 𝑦𝑡 + 𝑢𝑡+2 |Ω𝑡 )=𝜇 +
𝜙1 𝑓𝑡,1 + 𝜙2 𝑦𝑡
• 𝑓𝑡,3 = 𝐸 𝑦𝑡+3 Ω𝑡 = 𝐸(𝜇 + 𝜙1 𝑦𝑡+2 + 𝜙2 𝑦𝑡+1 + 𝑢𝑡+3 |Ω𝑡 )=𝜇 +
𝜙1 𝑓𝑡,2 + 𝜙2 𝑓𝑡,1
Forecasting AR(p) process

• Hence, the generic ‘s’ step ahead forecast becomes

• 𝑓𝑡,𝑠 = 𝜇 + 𝜙1 𝑓𝑡,𝑠−1 + 𝜙2 𝑓𝑡,𝑠−2
• These steps can be used to generate ARMA(p,q) order forecast
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Forecasting with ARMA Models: Part III

One-Step-Ahead vs. Multi-Step-Ahead
Forecasts and Rolling vs. Recursive Samples

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Determining the Accuracy of Forecast

Determining the Accuracy of Forecast

• Conditional expectations: The expression 𝐸 𝑦𝑡+1 𝛺𝑡 states the

expected value of y at ‘t+1’ conditional upon information available
up to time t, i.e., 𝛺𝑡 .
• Naïve forecasting: Forecast or expectation of y, s steps into future
is the current value of y, i.e., 𝐸 𝑦𝑡+1 𝛺𝑡 = 𝑦𝑡 . Then the process
follows random walk. For a mean reverting series some long-term
unconditional average is the best forecast for the series in future.
Forecasting with Time-Series Models
Steps ahead Forecast (F) Actual (A) Squared error Absolute error
(𝑭 − 𝑨)^𝟐 |F-A|
1 0.2000 -0.4000 ? ?
2 0.1500 0.2000 ? ?
3 0.1000 0.1000 ? ?
4 0.0600 -0.1000 ? ?
5 0.0400 -0.0500 ? ?

• Mean Squared Error (MSE)=?

• Mean Absolute Error (MAE)=?
• Root Mean Square Error (RMSE)=?
Forecasting with Time-Series Models
Steps ahead Forecast Actual Squared error Absolute error
1 0.2000 -0.4000 0.3600 0.6000
2 0.1500 0.2000 0.0025 0.0500
3 0.1000 0.1000 0.0000 0.0000
4 0.0600 -0.1000 0.0256 0.1600
5 0.0400 -0.0500 0.0081 0.0900

• MSE=(0.3600+0.0025+0.0000+0.0256+0.0081)/5=0.08
• MAE=(0.6000+0.0500+0.0000+0.1600+0.0900)/5=0.16
• RMSE=SQRT(MSE)=sqrt(0.08)=0.28
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Lesson 23: Non-stationarity,

Cointegration, and Error correction
Models
Module: Advanced Time-Series Models
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Time-Series Stationarity: Recap

Time-Series and Stationarity

• A strictly stationary process

• A time series process {yt} is said to be strictly stationary if the joint probability
distribution of a sequence {yt1 to ytn} is the same as that for {yt1+m to ytn+m}  m: i.e.,
𝐹 𝑦𝑡1 , … , 𝑦𝑡𝑛 = 𝐹 𝑦𝑡1+𝑚 , … , 𝑦𝑡𝑛+𝑚
• A weakly stationary process
• A weakly stationary (or covariance stationary) process satisfies the following three
equations
• 𝐸 𝑦𝑡 = 𝑢, 𝑡 = 1,2,3 … ∞
• 𝐸 𝑦𝑡 − 𝑢 𝑦𝑡 − 𝑢 = 𝜎 2 < ∞
• 𝐸 𝑦𝑡1 − 𝑢 𝑦𝑡2 − 𝑢 = γ𝑡 2−𝑡1 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡1 𝑎𝑛𝑑 𝑡2
Time-Series and Stationarity

• A weakly stationary process

• A weakly stationary (or covariance stationary) process satisfies the following
three equations:
• 𝐸 𝑦𝑡 = 𝑢, 𝑡 = 1,2,3 … ∞
• 𝐸 𝑦𝑡 − 𝑢 𝑦𝑡 − 𝑢 = 𝜎 2 < ∞
• 𝐸 𝑦𝑡1 − 𝑢 𝑦𝑡2 − 𝑢 = γ𝑡 2−𝑡1 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡1 𝑎𝑛𝑑 𝑡2
• To summarize, a weak stationary process should have a constant mean,
constant variance, and a constant auto-covariance structure.
• What is auto-covariance: 𝐸 𝑦𝑡 − 𝑢 𝑦𝑡+𝑠 − 𝑢 = γ𝑠 ; s = 0,1,2, ...
γ𝑠
• Autocorrelation: 𝜏𝑠 = , s = 0,1,2, ... ; 𝜏𝑠 lies in the interval ±1
γ0
Time-Series and Stationarity

• Problems with non-stationary series

• Shocks to non-stationary systems do not die away and are highly persistent
• Estimation with non-stationary data can lead to spurious regressions, i.e., artificially high t-stat.
and adjusted R-sq, even if the variables are not related to each other. t-ratio will not follow t-
distribution
Adj. R sq T-stat

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Types of Non-Stationarity
Types of Non-Stationarity
• Two types of non-stationarity
• In financial econometrics, two kinds of non-stationarity become important
1. Random walk model with drift: y t =  + yt-1 + ut (1)
2. Deterministic trend process: yt =  + t + ut (2)
• Here ut is a white noise iid process distributed as N (0, 𝜎 2 )
Recall: White Noise IID Process

• It can be defined as follows

1. E(yt) = 0 , t = 1,2,...,
2. Var(yt)= 𝜎 2
3. 𝐸 𝑦𝑡1 − 𝑢 𝑦𝑡2 − 𝑢 = γ𝑡 2 −𝑡1 = 0 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡1 𝑎𝑛𝑑 𝑡2 for all t1 ≠ t2
• Thus, white noise has a constant (zero) mean, finite constant
variance, and is uncorrelated across time.
• Often, we assume a fix distribution (e.g., Normal), then it is a
special case of IID process.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Stochastic Non-Stationarity
Stochastic Non-Stationarity

• A generalized model of the autoregressive process with drift.

• yt =  + yt-1 + ut; this is an explosive process if  > 1
• This does not describe real economic data because shocks to the
system persist and explode.
• So, the only non-stationarity case considered in that of  = 1.
Stochastic Non-Stationarity

• Consider the general case of an AR(1) with no drift: 𝑦𝑡 = 𝑦𝑡−1 + µ𝑡

• 𝑦𝑡−1 = 𝑦𝑡−2 + µ𝑡−1 and 𝑦𝑡−2 = 𝑦𝑡−3 + µ𝑡−2
• 𝑦𝑡 =  𝑦𝑡−2 + µ𝑡−1 + µ𝑡 =2 𝑦𝑡−2 + µ𝑡−1 + µ𝑡
• T successive substitutions 𝑦𝑡 = 𝑇 𝑦0 + µ𝑡−1 + 2 µ𝑡−2 + 3 µ𝑡−3 +
⋯ + 𝑇 µ0 + µ𝑡
• <1  𝑇 →0 as T→: Shocks to system die away
• =1  𝑇 =1 T: Shocks persist in the system and never die away: 𝑦𝑡 =
𝑡=0 𝑢𝑡 as T→
𝑦0 + σ∞
Stochastic Non-Stationarity (NS)

• 𝑦𝑡 = u + 𝑦𝑡−1 + µ𝑡 :Random walk non-stationary (NS) process with drift

• NS can be removed using first differences
• 𝛥𝑦𝑡 = 𝑦𝑡 - 𝑦𝑡−1 = 𝑦𝑡 - L 𝑦𝑡 = (1-L) yt = u + µ𝑡
• First differencing has introduced stationarity, so the process is called
integrated to the order of one I(1)
• Here (1-L) is called the characteristic equation of the process. Since
the equation has a root/solution of 1, it is also called unit root process.
 Stochastic Non-Stationarity
• First differencing has introduced stationarity, so the process is called
integrated to the order of one: I(1).
• If a non-stationary series, 𝑦𝑡 must be differenced d times before it becomes
stationary, then it is said to be integrated of order d: I(d).
• This would be written 𝑦𝑡 ~ I(d). So if 𝑦𝑡 ~ I(d) then Δ𝑑 𝑦𝑡 ~ I(0).
• Applying the difference operator, Δ, d times, leads to an I(0) process, i.e., a
process with no unit roots.
• I(1) and I(2) series can wander a long way from their mean value and cross
this mean value rarely, while I(0) series should cross the mean frequently.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Deterministic Non-Stationarity
Deterministic Non-Stationarity

• Deterministic trend process: yt =  + t + ut

• This is a rather simple case of non-stationarity
• Only de-trending is required, and a time ‘t’ term is added to the
model to remove the trend
• Thus, residuals obtained from the series, ut ‘s have their trend
removed
• Thus, any subsequent estimation can be done on these error
terms
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Time-Series and Stationarity:

Visual Examination
White Noise IID Process with Zero Mean
White noise IID process with zero mean

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

Random Walk and Random Walk with Drift
Random walk and Random walk with drift

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

Deterministic Trend Process
Deterministic Trend Process

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

Autoregressive Processes with Differing
Values of (0, 0.8, 1)
Autoregressive Processes with Differing Values of  (0, 0.8, 1)

Chris Brooks. Introductory Econometrics for Finance. 4th Edition. Chapter 6

Time-Series and Stationarity: Visual Inspection

• A white noise process visibly has no trending behavior, and it frequently crosses its
mean value of zero.
• The random walk (thick line) and random walk with drift (faint line) process exhibit
‘long swings’ away from their mean value, which they rarely cross.
• A comparison of the two lines in this graph reveals that the positive drift leads to a
series that is more likely to rise over time than to fall; obviously, the effect of the drift
on the series becomes greater and greater the further the two processes are
tracked.
• Finally, the deterministic trend process clearly does not have a constant mean and
exhibits completely random fluctuations about its upward trend. If the trend were
removed from the series, a plot similar to the white noise process would result.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Testing for Unit-Roots: Part I

Testing for a Unit Root : Simple Dicky–Fuller
Test
• Consider the model: yt = y t-1 + ut
• The basic objective of the test is to test the null hypothesis that 
=1, against the one-sided alternative  <1.
• So, we have: H0: series contains a unit root ; vs. H1: series is
stationary.
• We usually use the regression: yt = yt-1 + ut so that a test of =1
is equivalent to a test of =0 (since -1=).
• Why not original regression.
Testing for a Unit Root: Simple Dicky–Fuller
Test
• The model can be generalized by allowing for a drift (intercept)
and deterministic trend or neither.
• Consider the model: yt = y t-1++t+ut
• First Differencing the series: yt = yt-1++t +ut,
Where yt = yt- yt-1
• Then H0: Series contains a unit root  =1 and H1:  <1
• This is a test for a random walk against a stationary AR(1) with
drift and a time trend.
Testing for a Unit Root: Simple Dicky–Fuller
Test
• Dickey Fuller (DF) tests are also known as  tests: , , 
෢
𝛙
• This test statistic is computed as follows: Tau= ෢
𝑆𝐸(𝛙 )

• The test statistic does not follow the usual t-distribution under the
null since the null is one of non-stationarity but rather follows a
non-standard distribution.
• Tau values are larger (negative) than regular t-stat. indicating
more evidence required against null.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Testing for Unit-Roots: Part II

Testing for a Unit Root: Augmented Dicky–
Fuller (ADF) Test
• The null hypothesis of a unit root is rejected in favor of the
stationary alternative in each case if the test statistic is more
negative than the critical value.
• Also, this test requires ut to be white noise IID.
• ut will be autocorrelated if there was autocorrelation in the
dependent variable of the regression (yt).
• The solution is to “augment” the test using p lags of the
dependent variable.
Testing for a Unit Root: Augmented Dicky–
Fuller (ADF) Test
• The solution is to “augment” the test using p lags of the dependent variable
𝑝

𝛥𝑦𝑡 = 𝜓𝑦𝑡−1 + ෍ 𝛼𝑖 𝛥𝑦𝑡−𝑖 + 𝑢𝑡

𝑖=1

• This is referred to as the augmented DF test, as the lags absorb any

autocorrelation in the structure
• Again, the optimum lag selection can be done through information criteria
[What is the problem of too few and too many lags?].
• There are other tests Phillips Perron (PP) and KPSS test
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Testing for Unit-Roots: Part III

Summary

• ADF/PP: H0 : yt ~I 1 , i. e, nonstationarity and H1 : yt ~I 0 , i. e, stationarity

• KPSS: H0 : yt ~I 0 , i. e. , stationarity and H1 : yt ~I i , i. e, nonstationarity

Outcome ADF/PP KPSS

1 Reject 𝑯𝟎 Do not Reject 𝑯𝟎
2 Do not Reject 𝑯𝟎 Reject 𝑯𝟎
3 Reject 𝑯𝟎 Reject 𝑯𝟎
4 Do not Reject 𝑯𝟎 Do not Reject 𝑯𝟎
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Introduction to Cointegration and

Mean Reversion
Introduction

• A price series or a combination of two-series returns back to its

long-run mean value
• This mean value is driven by some fundamental factor
(cash/future relation)
• This leads to a property called stationarity
• Price series are rarely stationary; it is the returns that are most
often of stationary
Cointegration

• Most of the time if two variables are I(1), then their combination (error
term!) would also be I(1)
• Similarly, if two or more variables with different orders are combined,
then the combination will have the integration order equal I(d) to the
largest order variable I(d)
• Consider a simple regression model: 𝑦1 = 𝛼0 + 𝛼1 ∗ 𝑥1 + 𝛼2 ∗ 𝑥2 + 𝑢𝑡
or alternatively
• 𝑦1 − 𝛼0 − 𝛼1 ∗ 𝑥1 − 𝛼2 ∗ 𝑥2 = 𝑢𝑡 : here the error term is a combination
of three variables y1, x1, and x2.
Cointegration

• If these variables are I(1), chances are that error term would also be I(1).
• It would be desirable to have an error term as I(0): This is called cointegration.
• But what would be these economic and financial scenarios where an error would be
I(0).
• A set of variables are called cointegrated if a linear combination is stationary.
• Many time series (spot vs. futures, exchange rates) in finance and economics move
together.
• While in the short-run, they can go apart, in the long-run fundamental forces of the
market bring them together.
• It is a long-term equilibrium phenomenon.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Error Correction Models (ECM)

Error Correction Models (ECM)

• For univariate series, the approach to deal with non-stationarity is

to use first differences and then use the I(0) series for further
analysis.
• However, in this manner, we lose the original long-term
relationship between the variables at I(1) [Why we are only
discussing I(1)].
• For example, if two variables 𝑦𝑡 𝑎𝑛𝑑 𝑥𝑡 are I(1), then one may
consider estimating the following relation ∆𝑦𝑡 = α + 𝛽∆𝑥𝑡 + 𝑢𝑡
[Think of doing this for cash and futures data].
Error Correction Models (ECM)

• Often long-run is defined as a steady state equilibrium when

values have converged to steady and are not changing, i.e., 𝑦𝑡 =
𝑦𝑡−1 = 𝑦 𝑎𝑛𝑑 𝑥𝑡 = 𝑥𝑡−1 = 𝑥
• Thus, ∆𝑦𝑡 and ∆𝑥𝑡 are not changing and close to 0. So,
estimating equation above is meaningless
• Then the model ∆𝑦𝑡 = α + 𝛽∆𝑥𝑡 + 𝑢𝑡 has no solution
• However, there is a solution to this problem if the variables are
cointegrated.
Error Correction Models (ECM)

• Consider the following class of model, that employs lagged cointegrated

variables as well as their first differences. [𝑦𝑡−1 − ϒ𝑥𝑡−1 =𝑒𝑡 ]
• ∆𝑦𝑡 = α + 𝛽1 ∆𝑥𝑡 + 𝛽2 (𝑦𝑡−1 − ϒ𝑥𝑡−1 ) + 𝑢𝑡 ;
• This is called error correction model (ECM) or vector error correction model
(VECM).
• [𝑦𝑡−1 −ϒ𝑥𝑡−1 ] is the error correction term, provided that 𝑦𝑡 𝑎𝑛𝑑 𝑥𝑡 𝑎𝑟𝑒
𝑐𝑜𝑖𝑛𝑡𝑒𝑔𝑟𝑎𝑡𝑒𝑑, this term would be I(0) even though 𝑦𝑡 𝑎𝑛𝑑 𝑥𝑡 are I(1) [May
an intercept also]
• So, it is perfectly valid to use the OLS procedure here
Error Correction Models (ECM)

• Of course, the error correction term appears with lags (t-1); otherwise,
it would indicate that changes in y from t-1 to t are driven by
disequilibrium in the long-term relation at t.
• Gamma (ϒ) coefficient here defines the long-run relationship between x
and y.
• 𝛽1 describes the short-run relationship (first differences), and 𝛽2
describes the speed of adjustment toward equilibrium.
• This discussion can be extended to a system of more than two
variables.
 INDIAN INSTITUTE OF TECHNOLOGY KANPUR

Engle-Granger Approach to ECM and

Cointegration
Engle-Granger Approach

• Consider the following equilibrium model of k cointegrated

variables
• 𝑦𝑡 = 𝛽1 + 𝛽2 𝑥2𝑡 + 𝛽3 𝑥3𝑡 + ⋯ + 𝛽𝑘 𝑥𝑘𝑡 + 𝑢𝑡
• If the variables 𝑦𝑡 ’s and 𝑥𝑡 ’s are cointegrated, then 𝑢𝑡 can be I(0)
• One can set-up tests of stationarity on ut to confirm this
Error Correction Models (ECM)

Two-step Engle-Granger approach to parameter estimation:

Step 1
• Use the I(1) variables to estimate the cointegrating model and extract the residuals ut
• Test these errors for I(0) stationarity
Step 2
• If the residuals are stationary, then estimate the following VECM model
• ∆yt = α0 + α1 ∆xt + α2 (ut ) + vt
• If the relationship is 𝑦𝑡 − 𝛽2 𝑥2𝑡 − 𝛽3 𝑥3𝑡 , then [1 − 𝛽2 − 𝛽3 ] is the cointegrating vector
• All the problems related to OLS discussed earlier are also applicable here
Thanks!