Professional Documents
Culture Documents
Florian Pelgrin
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 1 / 32
Road map
1 Introduction
Overview
2 Identification
Overview
Identifying d
Seasonality
Identifying p and q
4 Diagnostic checking
Overview
Residual diagnostics
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 2 / 32
Introduction Overview
1. Introduction
Step 1 : Identification
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 3 / 32
Introduction Overview
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 4 / 32
Introduction Overview
Least Squares or
Estimation
Maximum Likelihood
No
Model
ok?
Yes
Forecasting
Use the Model Explanation
Control
2. Identification
2.1. Overview
Three objectives
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 6 / 32
Identification Identifying d
2.2. Identifying d
Graphical procedure : this involves plotting the data over time and
the corresponding (partial) autocorrelation function.
If the ACF does not decrease to zero or at a very slow decay : this
suggests non-stationarity (or long-memory effects).
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 7 / 32
Identification Identifying d
2000 30 4
3
0 20
2
10
-2000 1
0 0
-4000
-10 -1
-6000 -2
-20
-3
-8000 -30
-4
-10000 -40 -5
250 500 750 1000 250 500 750 1000 250 500 750 1000
2.3. Seasonality
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 10 / 32
Identification Seasonality
• • •
• •
•• • ••• •
••
• •• •
• ••• • •• ••
• • • ••• ••• ••
• ••• ••
12.6
• •• •• • •• • ••
• • • • •
•••• • •• • •• ••
• ••• • •
•
•••• • •
• •• •• • •• •
rfs
••
12.4
• ••••••
• •
•••• •• •
• • •
••• • • •
•• • ••
• ••• • •• • ••
•••
12.2
•
• •• • •
•
•••• • •
• • •
•••••
• • • •
•••• • • •
12.0
• • ••
•
••
••
3000
Quadrillion Btu
2500
2000
1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993
Year
The ACF ;
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 13 / 32
Identification Seasonality
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 14 / 32
Identification Seasonality
Example: Seasonality
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 16 / 32
Identification Identifying p and q
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 17 / 32
Identification Identifying p and q
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 18 / 32
Identification Identifying p and q
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 19 / 32
Estimation and information criteria Estimation
Etc
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 20 / 32
Estimation and information criteria Information criteria
Definition
The Akaike information criterion (AIC) is defined to be
2
AIC(k) = log σ̂ 2 (k) + (k)
T
where k is the total number of parameters estimated.
Remarks :
AIC may give more than one minimum (p, q) ;
AIC depends on the normality assumption ;
AIC is not consistent : it will deliver on average too large a model
(even with T → ∞)—AIC tends to over-parameterize.
AIC is generally efficient (small average variation in selected model
orders from different samples within a given population).
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 23 / 32
Estimation and information criteria Information criteria
Remarks :
SBIC embodies a much stiffer penalty term than AIC, i.e. SBIC
penalizes larger models more than AIC and thereby tends to select
lower-order models than the AIC.
SBIC is strongly consistent in selecting p and q : if the data are truly
generated by an ARMA(p,q) process, the SBIC picks the true model
with probability 1 as the sample size approaches infinity.
SBIC is less efficient than AIC.
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 24 / 32
Estimation and information criteria Information criteria
Definition
The Hannan-Quinn information criterion (HQIC) is defined to be
2k
HQIC(p, q) = log σ̂ 2 (k) + log (log(T ))
T
where k is the total number of parameters estimated.
Remarks :
HQIC contains a penalty term which is somewhere between AIC and
SBIC.
HQIC is strongly consistent in selecting p and q.
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 25 / 32
Estimation and information criteria Information criteria
General methodology
1 Set upper bounds, pmax and qmax , for the AR and MA order,
respectively ;
2 Fit all possible ARMA(p,q) models for p ≤ pmax and q ≤ qmax using a
common sample size ;
min AIC(p, q)
p≤pmax ,q≤qmax
min SBIC(p, q)
p≤pmax ,q≤qmax
min HQIC(p, q).
p≤pmax ,q≤qmax
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 26 / 32
Estimation and information criteria Information criteria
Even if one decides to only use one of these criteria, several models
that are close to the minimum information criterion value might be
considered (adjacent models)...
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 27 / 32
Diagnostic checking Overview
4. Diagnostic checking
4.1. Overview
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 28 / 32
Diagnostic checking Residual diagnostics
Two methods
Graphical procedure
1 Plot of the residuals : do they exhibit systematic patterns ?
2 Use the SACF and SPACF (as in the identification step) of the
residuals : Do they have significant elements ?
Testing procedure : autocorrelation tests.
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 29 / 32
Diagnostic checking Residual diagnostics
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 30 / 32
Diagnostic checking Residual diagnostics
where ρ̂2ˆ (k) is the k-th order sample autocorrelation of the estimated
residuals, T is the sample size, and K is chosen sufficiently large.
The Q-test has an asymptotic chi-square (χ2 ) distribution with
K − p − q degrees of freedom. The null hypothesis of uncorrelated
(estimated) residuals is rejected if the Q exceeds the tabulated critical
value (for a chosen significance level).
The Q-test is only asymptotically valid and may perform poorly for
small and medium size samples.
Refinements have been introduced in the literature : Ljung-Box test,
McLeod-Li test, Monti’s test, etc.
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 31 / 32
Diagnostic checking Residual diagnostics
(c) Discussion
Florian Pelgrin (HEC) Univariate time series Sept. 2011 - Dec. 2011 32 / 32