Scribd
Upload
32 views18 pages

Sta457 Week 2 Notes

This lecture covers Exponential Smoothing (ETS) models and white noise in time series analysis. It discusses various smoothing techniques, including simple, linear trend, and Holt-Winters exponential smoothing, along with their forecasting equations. Additionally, it introduces the concept of state-space models and definitions related to L2 and white noise series.

Uploaded by

easyacemt
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
32 views18 pages

Sta457 Week 2 Notes

This lecture covers Exponential Smoothing (ETS) models and white noise in time series analysis. It discusses various smoothing techniques, including simple, linear trend, and Holt-Winters exponential smoothing, along with their forecasting equations. Additionally, it introduces the concept of state-space models and definitions related to L2 and white noise series.

Uploaded by

easyacemt
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

TA457: Time Series Analysis

Lecture 3

Lijia Wang

Department of Statistical Sciences


University of Toronto

Lijia Wang (UofT) TA457: Time Series Analysis 1 / 18


Overview

Last Time:
1 Detrending
2 Di!erencing
Today:
1 ETS Models
2 White noise

Lijia Wang (UofT) TA457: Time Series Analysis 2 / 18


Error Function
By detrending/di!erencing, we turn a ts into an approximate mean zero
noise series. Big question: What is the estimated function g ?

Figure: Temperature di!erences over time.


Lijia Wang (UofT) TA457: Time Series Analysis 3 / 18
Monthly Mortality

Figure: Monthly mortality Data

Lijia Wang (UofT) TA457: Time Series Analysis 4 / 18


Exponential Smoothing

Introduction: Exponential smoothing was introduced in the late 1950s.


The basic underlying principle is that for a time series x1 , . . . , xn , two
extreme forecasts are:
x̂n+1 = xn (random walk prediction)
1 !n
x̂n+1 = x̄ = n j=1 xj (iid noise prediction)
For general time series, the optimal prediction might fall between these
extremes.

Lijia Wang (UofT) TA457: Time Series Analysis 5 / 18


Simple Exponential Smoothing

Simple exponential smoothing equation:

x̂n+1 = ωxn + ω(1 → ω)xn→1 + ω(1 → ω)2 xn→2 + · · ·

0 ↑ ω ↑ 1. This prediction may be approximately stated as

Smoothing Equation: εn = ωxn + (1 → ω)εn→1


Initial Condition: ε0
where ω and ε0 can be estimated via least squares:
n
"
(ω̂, ε̂0 ) = argmin (xi → εi (ω, ε0 ))2
0↑ω↑1,ε0 i=2

In terms of forecasting
Forecast/Prediction Equation: x̂n+1 = εn

Lijia Wang (UofT) TA457: Time Series Analysis 6 / 18


Linear Trend Exponential Smoothing

Sometimes called double exponential smoothing, this considers the


following equations:

Level equation: εn = ωxn + (1 → ω)(εn→1 + bn→1 )


Trend/Slope equation: bn = ϑ(εn → εn→1 ) + (1 → ϑ)bn→1
Scalar Parameters: ω, ϑ, ε0 , b0 can be estimated using least squares.

In terms of forecasting
Forecast equation: x̂n+m = εn + mbn

Lijia Wang (UofT) TA457: Time Series Analysis 7 / 18


Trend + Seasonal Exponential Smoothing

In Holt-Winters exponential smoothing (or triple exponential smoothing),


suppose m is the forecast horizon and the data has seasonality p units,
define k = ↓(m → 1)/p↔, then we have

Level equation: εn = ω(xn → sn→p ) + (1 → ω)(εn→1 + bn→1 )


Trend equation: bn = ϑ(εn → εn→1 ) + (1 → ϑ)bn→1
Seasonal equation: sn = ϖ(xn → εn→1 → bn→1 ) + (1 → ϖ)sn→p
Scalar Parameters: ω, ϑ, ϖ, ε0 , b0 , s0 , . . . , s→p+1 can be estimated using
least squares.

In terms of forecasting
Forecast equation: x̂n+m = εn + mbn + sn+m→p(k+1)

Lijia Wang (UofT) TA457: Time Series Analysis 8 / 18


Mortality Decomposed

Figure: Holt-Winters Trend and Seasonal Decomposition.

Lijia Wang (UofT) TA457: Time Series Analysis 9 / 18


Mortality Holt-Winters

Figure: Holt-Winters 10 month ahead forecast.

Lijia Wang (UofT) TA457: Time Series Analysis 10 / 18


Another View

Exponential smoothing models produce forecasts which are weighted


averages of past observations, with the weights decaying exponentially as
the observations get older. (Note: Holt-Winters sometimes overpredicts,
so damping is common.)

Alternatively, the model is built from a measurement equation that


describes the observed data, and some equations for the “state” which
describe how the unobserved components or states (level, trend, seasonal)
change over time.

Lijia Wang (UofT) TA457: Time Series Analysis 11 / 18


State-space Model

Exponential Smoothing Models can be written as State-Space Models.


This is sometimes also referred to as the “Innovation Form”:
Observation equation: xt = εt→1 + ϱt
State equation: εt = εt→1 + ωϱt
where ϱt is an innovation sequence.

The innovations can be estimated by the residuals xt → εt→1 . This allows


for data generative models of xt , and the calculation of prediction MSE
Pnn+m , either by simulation or, in some cases, analytically.

Lijia Wang (UofT) TA457: Time Series Analysis 12 / 18


ETS Models

The state space ETS models consider equations for the following:
Error
Trend
Seasonality
Each can be considered as missing (N), additive (A), or multiplicative (M)
relationships.

When ETS components are all additive, the result is the Holt-Winters.

Lijia Wang (UofT) TA457: Time Series Analysis 13 / 18


Theories

Starting here, we concentrate on the theoretical parts:


L2 Definition
white noise
autocorrelation
stationary

Lijia Wang (UofT) TA457: Time Series Analysis 14 / 18


L2 Definition

Definition: A real valued time series {xt , t ↗ Z} is said to be in L2 if


E [xt2 ] < ↘ for each t ↗ Z.

X ↗ L2 (”, F, P) ≃⇐ E (X 2 ) < ↘.
In other words, var(xt ) < ↘ for each t ↗ Z.
Note: In general, we assume that time series are in L2 .

Lijia Wang (UofT) TA457: Time Series Analysis 15 / 18


White Noise

Definition: Suppose that {wt , t ↗ Z} is a sequence of independent and


identically distributed random variables with mean zero, and variance
E [wt2 ] = ςw2 < ↘.
xt = w t
is called a strong white noise series in L2 (abbreviated xt ⇒ iid(0, ςw2 )).

If in addition, we assume wt ⇒ N(0, ςw2 ), then we say xt is a Gaussian


iid
white noise, abbreviated xt ⇒ N(0, ςw2 ).

Lijia Wang (UofT) TA457: Time Series Analysis 16 / 18


Gaussian Noise

Figure: Gaussian white noise series. The term “white” comes from spectral
analysis, where a white noise series shares the same spectral properties as white
light: all periodicities occur with equal strength.

Lijia Wang (UofT) TA457: Time Series Analysis 17 / 18


Gaussian Noise And First Di!erence

If these two series behave in the same way, then it stands to reason
iid # 2
$
g (φt , φt→1 , . . .) = φt ⇒ N 0, ςtemp
Lijia Wang (UofT) TA457: Time Series Analysis 18 / 18

You might also like