Scribd
Upload
40 views22 pages

ARMA and ARIMA

- ARIMA models use past values of a time series to forecast future values. They combine autoregressive (AR) and moving average (MA) components. - The ARIMA process involves identifying whether a time series is stationary and differencing if needed. This is represented by the "I" in ARIMA. - The Box-Jenkins methodology is used to identify, estimate, and diagnose ARIMA models. This includes examining the ACF and PACF and testing residuals. - The best ARIMA model is selected using statistical significance, AIC, BIC, and checking that residuals are independent white noise. Overfitting helps determine the optimal model order.

Uploaded by

Nayli Athirah Azizon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
40 views22 pages

ARMA and ARIMA

- ARIMA models use past values of a time series to forecast future values. They combine autoregressive (AR) and moving average (MA) components. - The ARIMA process involves identifying whether a time series is stationary and differencing if needed. This is represented by the "I" in ARIMA. - The Box-Jenkins methodology is used to identify, estimate, and diagnose ARIMA models. This includes examining the ACF and PACF and testing residuals. - The best ARIMA model is selected using statistical significance, AIC, BIC, and checking that residuals are independent white noise. Overfitting helps determine the optimal model order.

Uploaded by

Nayli Athirah Azizon
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 22

Week 3 & 4

FINC7325
Autoregressive and Moving Average
(ARMA & ARIMA)

Source: Studendmund (2014) and others


Note: The slide is prepared for the purpose of lecture and not to be
distributed to others or for publication purposes
ARIMA Models
• ARIMA is a highly refined curve-fitting device that uses
current and past values of the dependent variable to produce
short-term forecasts of that variable.
• ARIMA ignores independent variables and underlying
theories except those that hypothesize repeating patterns in
the variable under study.
• ARIMA is appropriate when little is known about the
dependent variable.
• ARIMA models can often beat moderately sophisticated
econometric models in forecasting outside a sample.
ARIMA Models (continued)
• The ARIMA approach combines two specifications:
1. Autoregressive process expresses a dependent variable as a
function of past values of the dependent variable. (p)
2. Moving average process expresses a dependent variable as
a function of past values of the error term. (q)
ARIMA Models (continued)
• Before ARIMA can be applied, the time series must be
stationary.
• A nonstationary series can often be converted to stationary
by taking the first difference of the variable:

• If first differences do not produce stationarity, the first


difference of the first differences might:
ARIMA Models (continued)
• The number of differences required for a series to become
stationary is denoted d.
• If a forecast of Y* or Y** is made, the it must be converted
back into terms of Y.
• For example, if d = 1, then:

• This process is represented by the “I” in ARIMA and stands


for Integrated.
• ARIMA is: AutoRegressive Integrated Moving Average
ARIMA Models (continued)
• As a shorthand, an ARIMA model with p, d, and q is denoted
as: ARIMA(p,d,q).

• An ARIMA(2,1,1) would have


2 autoregressive terms,
1 first difference, and
1 moving average term

where:
Box-Jenkins Methodology for time Series Modeling Procedure

(Source: Makrtidakis et al., 1998)


Theoretical ACF & PACF – ARMA Models
Process ACF PACF
AR (p) Dies down/ tails off Cuts off at lag p
MA (q) Cuts off at lag q Dies down/tails off
ARMA (p,q) Tails off/dies down Tails off/dies down

Note: it may not be easy to decide whether ACF/PACF cuts off or tails off. But don’t worry, the
process may help us to improve and develop the new fitted model.

Example:
AR (1) : ACF tails off , PACF cuts off at lag 1
MA (1): ACF cuts off at lag 1, PACF tails off
ARMA (1,1): ACF and PACF both tail off
*IF ACF is nonstationary, PACF is not relevant, the series need a transformation!
Nonstationary series shows ACF tails off slowly in LINEAR fashion. In practice, p and q
are less than 3.
• Hands on using simulated data below:-
Lets do indentification procedures for the following data using EVIEWS2:
DATA1- ACF cuts off at lag 1, while PACF tails off exponentially quickly.. MA(1)
DATA2-ACF cuts off at lag 1, PACF tails off alternatingly
MA(1) or ARMA (0,1)
DATA3-ACF tails off exponentially, PACF cuts off at lag 2
AR(2) or ARMA(2,0)
DATA4-ACF tails off slowly in linear fashion, NONstationary, transformation by
taking the first difference.
Differdata4-ACF tails off alternatingly, PACF cuts off lag 1
ARIMA (1,1,0)
DATA5-ACF tails off slowly in linear fashion-NONstationary-Transformation
Differ5- ACF tails off PACF cuts off at lag 2- ARIMA(2,1,0) both ACF and PACF tails
off ARIMA (p,1,q)
Parameter estimations
• Data1: MA(1) is significant
Based on ACF and PACF residual are independent at earlier lags
LM test indicates residual are independent,

Heteroskedasticity test
Parameter estimations
• Data3: AR(2) component is significant
Based on ACF and PACF residual are independent at earlier lags
LM test indicates residual are independent,

Heteroskedasticity test
Overfit data3-AR(2)-ARMA(2,0)
• AR(3)- the coefficient of AR(3) is not significant-reject
• ARMA(2,1)- MA(1) coefficient is not significant-reject

• Assume AR(2) as the best model to predict data3


Parameter estimations
• Data4: ARIMA(1,1,0) component is significant
Based on ACF and PACF residual are independent at earlier lags
LM test indicates residual are independent,

Heteroskedasticity test, residual homoskedastic


• Overfit data 4- ARIMA(1,1,0)
• ARIMA(2,1,0)- AR(2) coefficient is not significant-reject
• ARIMA(1,1,1)-ma(1) coefficient is not significant-reject
• ARIMA(2,1,1)-all the coefficients are not significant-reject

Assume ARIMA(1,1,0 as the best model


Parameter Estimation and Model Selection Criteria
• Variances of estimates can be obtained from matrix of second-order partial
derivatives of the likelihood. Significance of parameters is tested using
standard t-test,

• AIC and BIC criteria can be applied in choosing the best model. The smaller
value of AIC/BIC indicates a better model.
• There is small chance that the best model based on these criteria does not
have white noise errors, and choosing 2nd best model may result in better
adequacy.
Residual Analysis and Diagnostic checking
• Independent assumption of residual can be checked using ACF plot.
Residuals would be independent and indicates white noise if ACF
values falls inside Bartlett confidence interval.
• Attention is given to residuals at low lag!
• As ACF is not a test, Serial LM test is more often used to check
autocorrelation.
Model improvement for autocorrelation in residuals
Fitted Model ACF and PACF of residuals Residual New Model What Next
Model
AR(1) Independent None None Apply overfitting
AR(1) ACF Tails off, AR(1) AR(2) Check if residuals of
PACF Cuts off at lag 1 AR(2) are independent.
Then apply overfitting
AR(2) ACF cuts off at lag 1, MA(1) ARMA(2,1) Check if residuals of
PACF Tails off ARMA(2,1) are
independent. Then
apply overfitting

MA(1) Not independent but no ARMA(1,1) or ARMA(1,1) and Similar to above.


pattern apply MA(2)
overfitting
Overfitting
• Test whether model with higher degree of p and q is better.
• Used AIC and BIC criteria
Data1= MA(1)
Overfit ARMA(1,1) and MA(2)
ARMA(1,1)- only MA(1) coefficient is significant-reject
MA(2)-only MA(1) coefficient is significant –reject
In this case we going to proceed with MA1)

Next, forecast! But for short-term only.


Using unemployment female data
• Data is stationary at first difference.
ARIMA(p,1,q)- ARIMA(0,1,1)
ACF cuts off at 1, PACF tails off = MA(1)
Data consists of 300 observations
Estimate model using sample 1- 275
Forecast within sample , 276-300
• ARIMA(0,1,1) – MA(1) is significant- residuals are independent and
homos..AIC=10.07 and BIC=10.08
Next: Overfit ARIMA(0,1,1)
1)ARIMA(1,1,1)- AR(1) coefficient not significant-IGNORE
2)ARIMA(0,1,2)-MA(2) coefficient is not significant-IGNORE
3)ARIMA(1,1,2)-AR(1) and MA(2) coefficients are significant-
• residuals are independent and homos..AIC=18.0867 and BIC=18.1807
Based on AIC criteria, we select ARIMA (0,1,1) instead of ARIMA(1,1,2)

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy