Auto-correlation

Regression Flowchart
 Autocorrelation
• Lebanon Example

• Errors are correlated

• Observed in time series data

– A time series is a sequence of observations taken
sequentially in time

– An intrinsic feature of a time series is that, typically

adjacent observations are dependent

– The nature of this dependence among observations of

a time series is of considerable practical interest
 Autocorrelation

– Say, output of a farm regress on capital

and labor on quarterly data and there is
a labor strike on a particular quarter
affecting the output in that particular
quarter  No autocorrelation

– But if the strike affect the output in

other quarters as well  Autocorrelation
 Reasons
• Inertia or sluggishness
– agricultural commodities
Supplyt = B1+ B2Pt-1+ ut
– Say at the end of period t, Pt turns out to be
lower than Pt-1, so the farmer may decode to
produce less in period (t+1) than t
• Data manipulation
– Monthly to quarterly data by averaging,
thereby damping the fluctuating of monthly
data
– Smoothness leads to systematic pattern in
disturbances
 Reasons
• Model Specification errors
– Omitting relevant variable(s)
• Ex: Y=b1+b2X2t+b3X3t+b4X4t+ut but we run the
regression Y=b1+b2X2t+b3X3t+ vt, where
vt=b4X4t+ut
• If we plot v, it will not be random but exhibit a
systematic pattern  creating (false)
autocorrelation
• Lags
consumptiont=b1+b2incomet+b3consumptiont-1+ ut
 Autocorrelation
ut

Time
Order of auto-correlation
 Consequences of autocorrelation

• Least square estimates are linear and

unbiased but they do not have minimum
variance property

• t & F statistics are not reliable

• R2 may be an unreliable measure of

true R2
 Detection
• Plotting OLS residuals against time
– No autocorrelation  errors are randomly
distributed
– Presence of autocorrelation  errors
exhibit a distinct behavior

• Correlogram Q stat

• Serial correlation LM test

• DW statistics (based on estimated
residuals) ***
Durbin-Watson statistic is not appropriate as a test
for serial correlation in this case, since there is a
lagged dependent variable on the right-hand side of
the equation.
 Corrections for Serial Correlation
• Let, Yt = b1+b2X2t+b3X3t+ …+bkXkt+et …….(1)

• et = ρet – 1+ vt,  errors follow AR(1) process

• -1 < ρ <1, vt~ N (0, σv2), independent of other

errors over time

• The error process is generated by a rule which

says that the error in time t is determined by
diminishing the value of the error in previous
period (multiplying by ρ) and then adding the
effect of a random variable
 Corrections for Serial Correlation
• ρ measures correlation coefficient
between errors in time period t and that
in time period (t-1)
• Assuming that equ (1) holds for (t-1) period also
• Yt-1 = b1+b2X2t-1+b3X3t-1+bkXkt-1+et-1
• Multiplying this equation by ρ and subtracting
from equation (1)
• Yt* = b1(1- ρ )+b2X2t* +b3X3t *+…+bkXkt *+vt ..(2)

– whereYt* = Yt- ρYt-1, Xt* = Xt- ρXt-1, (et - ρet – 1) = vt

 Corrections for Serial Correlation
• By construction, the transformed equation has an
error process, which is independently distributed with
0 mean and constant variance
• OLS can give efficient estimates of all the regression
parameters
• Generalized differencing procedure would be useful
when value of ρ was known a priori
• However, this is usually not the case

• Procedures for estimating ρ

– The Cochrane- Orcutt Procedure
– The Hildreth-Lu Procedure
– ML method
 Cochrane- Orcutt Procedure
• It involves a series of iterations, each of which produces a
better estimate of ρ than does the pervious one
• In the first step, OLS is used to estimate original model
[equation (1)]
• The residuals from this equation are then used to perform the
regression

êt  ρ ê t-1 + vt

• The estimated value of ρ can be substituted in equation (2)

and a new regression is run
• Substituting ‘b’s in equ (1), we get new residuals and again run
the regression
êˆt  ρ êˆ t-1 + vt
 Cochrane- Orcutt Procedure
• This second set of residuals can be used to
obtain a new set of ρ
• The iterative process can be carried on for as
many steps as desired
• Standard procedure is to stop iteration when
the new estimates of ρ differ from old ones
by less than 0.01 or 0.005 or after 20-25
estimates of ρ have been obtained
• EViews estimates AR models using nonlinear
regression techniques.
 Residuals are not random

DW is low  Autocorrelation
t, F, R2 are unreliable
Ho: No autocorrelation
Here, 0% chance
So, there is autocorrelation
Add ar(1) as regressor to rectify autocorrelation

AR(1) significant along with others and

coefficient of AR is < 1
Ho of no autocorrelation is accepted
 Forecasting
• Static: Yi =24.45+0.509 Xi
• Dynamic Yt+1 = b1 + b2 Xt+1+ ρut+ vt+1
 Forecasting
• Root-Mean-Squared-Error (RMSE) and Mean Absolute
Error (MAE) criteria depend on the scale of the
variable, while the Mean Absolute Percentage Error
(MAPE) and Theil inequality coefficient are insensitive
to the scale of the variable

• The smaller the error, the better is the forecasting

performance for the series

• The Theil inequality coefficient always lies between

zero and one, where zero indicates a perfect fit

• Any forecast is also said to be “good”, if the bias and

variance proportions are small so that most of the bias
should be concentrated on the covariance proportions.
 Degree Days
• DDs are a practical method for determining cumulative
temperatures over the course of a season. Originally
designed to evaluate energy demand and consumption,
degree days are based on how far the average
temperature departs from a human comfort level of 65 °F
*.

• Simply put, each degree of temperature above 65 °F is

counted as one cooling degree day, and each degree of
temperature below 65 °F is counted as one heating degree
day

• Degree days are determined by subtracting average daily

temperature from 65° F. Example: average daily
temperature is 25° F, (65 – 25 = 40) then we have
accumulated 40 degree days on this date.