Вы находитесь на странице: 1из 26

ARIMA Modeling:

B-J Procedure

A k Chauhan
v_akc@rediffmail.com
9811216905

Structural models multivariate in nature, and


Y = F (movements in current or past values of other
(explanatory) variables).
Univariate time series models
Yt = F (own past values, Current & past values of an
error term)
useful when a structural model is inappropriate.
when other variables are not observable or not
measurable
Explanatory variables are measured at a lower
frequency of observation than yt. For eg, yt might be a
series of daily stock returns, where possible explanatory
variables could be macroeconomic indicators that are
available monthly.

Moving average processes:


Let ut be a white noise process with zero mean & constant
variance 2. Then

yt = + ut + 1ut-1 + 2 ut-2 + . . . +
q ut-q

is a MA(q) model.

Autoregressive processes:
An AR model is one where the current value of a variable
y, depends upon only the values that the variable took
in previous periods plus an error term. An AR model of
order p, denoted as AR(p), can be expressed as

Yt = u + 1 yt-1 + 2 yt-2
yt-p + ut

+ .. + p

where ut is a white noise disturbance term. The


expression of AR(p) model can be written more
compactly using sigma notation

ARMA (p,q) processes


obtained by combining the AR(p) & MA(q) models.
States that the current value of some series y depends
linearly on its own previous values plus a combination of
current and previous values of a white noise error term.
The model could be written

Yt = u + 1 yt-1 + 2 yt-2 + ..+ p ytp +


1 ut-1 + 2 ut-2 + ..+ q ut-q + ut
Is a ARMA (p, q)m model

An AR process has:

a geometrically decaying ACF


a number of non-zero points of PACF = AR
order.

A MA process has:

number of non-zero points of ACF = MA order


a geometrically decaying PACF.

A combination ARMA process has:

a geometrically decaying ACF


a geometrically decaying PACF.

Process of ARIMA
Modeling

Identification Stage
Estimation Stage
Diagnostic Checking
Forecasting

Identification

First identify the order of integration on the


basis of visual inspection of time series plot,
correlogram and formal unit root testing.
Detrend and Differentiate the series if required to
obtain a stationary series.
Identify the undrlaying AR (p) and MA (q)
processes in the stationary series based on the
behaviour of the correlogram.
In order to achieve parsimony keep the number
of AR and MA lag orders, p and q as small as
possible; provided the model still satisfactorily
forecast the series.
It is better to used mixed ARMA (p, q) model
rather than pure AR (p) or MA (q) model.

Estimation

involves estimation of parameters of model


specified in step 1.
This can be done using least squares or
another technique, known as Max likelihood,
depending on the model.

Model checking
determining whether the model specified and
estimated is adequate.
Suggested methods are:
Overfitting -- fitting a larger model than that required.
If the model specified at step 1 is adequate, any extra
terms added to ARMA model would be insignificant.
Residual diagnostics. Residual should be a white
noise series.
Information Criteria: SBIC, AIC HQ should have least
values.

Negative MA Coeff

Negative MA Coeff

Estimation Stage
A pure AR model can be estimated
using OLS estimator but if the model
also includes MA terms ML estimator
is used.

Diagnostic Checking

How good the model fits the


data? How reliable will be the
prediction based on it?
Use some indicators of goodness of fit
such as R2 and ANOVA.

Is the model sufficiently specified


or it requires including some more
AR or MA terms?
Examine the residuals. Their
correlogram can give a clue if
residuals are pure white noise or at
least free from serial correlation. Q
and LB statistics of residuals is useful
for detecting serial correlation. Use LR
test for AR in residuals.

Does the model include some


unnecessary parameters (AR or MA
terms) which can be dropped out
without a significant loss in predictive
efficiency of the model?

t-statistic of individual coefficient can


give an initial (but not conclusive)
clue.
.

Forecasting
In ARMA (p, q) the prediction equation
is simply a linear equation that refers
to past values of original time series
and past values of the errors.
In an ARIMA (p,d,q) model where d>0,
the forecasting can be made at two
different levels at the level of differenced stationary time
series, and
at original integrated time series.

Вам также может понравиться