计算机代考 QBUS6840 Lecture 6 Exponential Smoothing (Seasonal)

QBUS6840 Lecture 6 Exponential Smoothing (Seasonal)

QBUS6840 Lecture 6

Copyright By PowCoder代写 加微信 powcoder

Exponential Smoothing (Seasonal)

The University of School

Exponential smoothing

Holt-Winters smoothing

Exponential smoothing methods for seasonal data.
Additive seasonality.
Multiplicative seasonality.

Damped Trend Exponential Smoothing
Damped Trend Seasonal

Online Textbook Sections 7.3-7.4 and 7.6:
https://otexts.org/fpp2/expsmooth.html and/or

BOK Sec 8.4-8.5

https://otexts.org/fpp2/expsmooth.html

Objectives

Be able to distinguish between Additive seasonality and
Multiplicative seasonality Exponential smoothing methods for
seasonal data is an Exponential smoothing methods for
seasonal data.

Fully understand the statistical model for exponential
smoothing methods for seasonal data

Be able to use the error correction form to derive forecast and
variance formula

Understand the Damped Trend Exponential Smoothing and
be able to work with its model

Be able to explain Damped Trend Seasonal ES model

Visitor arrivals in Australia
Original series (2006-2015), SES and TCES

Smoothing by SES (black) and TCES (red). They are not suitable
for seasonal data

Holt-Winters smoothing

is an Exponential smoothing methods for seasonal data.

used for both Additive seasonality and Multiplicative
seasonality.

Additive seasonality: the seasonal variation is slowly changing
along the trend as an additive component
Multiplicative seasonality: the seasonal variation is slowly
changing along the trend as a multiplicative component

Introducing Additive Holt-Winters smoothing

The ideal scenario

Yt = ω0 + ω1t + St + εt

Additive decomposition model: assuming ω0, ω1 and St (M
different values) are fixed constants.

Simple exponential method: modelling the case where St = 0,
ω1 = 0 (or constant) and ω0 changes with time. Denoted by

Trend corrected exponential method: modelling the case
where St = 0, both ω1 and ω0 are (slowly) changing

How to model the data if the level, the level growth rate (the
trend), and seasonal patterns are changing?

Additive Holt-Winters smoothing

Suppose the time series {Y1,Y2, …,YT} has an additive
seasonality with seasonal frequency M.

The basic idea of the Holt-Winters method is to use
exponential smoothing for all the levels, trends and seasonal
components.

lt = α(Yt − St−M) + (1− α)(lt−1 + bt−1), 0 ≤ α ≤ 1
bt = β(lt − lt−1) + (1− β)bt−1, 0 ≤ β ≤ 1
St = γ(Yt − lt−1 − bt−1) + (1− γ)St−M , 0 ≤ γ ≤ 1

with the forecast:

Ŷt+1|1:t = lt + bt + St+1−M = lt + bt×1 + St+1−M .

Some people write the seasonal update as

St = γ(Yt − lt) + (1− γ)St−M , 0 ≤ γ ≤ 1

Additive Holt-Winters smoothing

Suppose the time series {Y1,Y2, …,YT} has an additive
seasonality with seasonal frequency M.

The basic idea of the Holt-Winters method is to use
exponential smoothing for all the levels, trends and seasonal
components.

lt = α(Yt − St−M) + (1− α)(lt−1 + bt−1), 0 ≤ α ≤ 1
bt = β(lt − lt−1) + (1− β)bt−1, 0 ≤ β ≤ 1
St = γ(Yt − lt−1 − bt−1) + (1− γ)St−M , 0 ≤ γ ≤ 1

with the forecast:

Ŷt+1|1:t = lt + bt + St+1−M = lt + bt×1 + St+1−M .

Some people write the seasonal update as (not used in our unit)

St = γ(Yt − lt) + (1− γ)St−M , 0 ≤ γ ≤ 1

Explanation
Additive Holt-Winters smoothing

t −M t + 1−M t − 1 t t + 1

Yt − lt−1 − bt−1
St = γ(Yt − lt−1 − bt−1) + (1 − γ)St−MSt

lt = α(Yt − St−M ) + (1 − α)(lt−1 + bt−1)

bt = β(lt − lt−1) + (1 − β)bt−1

Explanation
Additive Holt-Winters smoothing

t −M t + 1−M t − 1 t t + 1

Yt − lt−1 − bt−1
St = γ(Yt − lt−1 − bt−1) + (1 − γ)St−MSt

lt = α(Yt − St−M ) + (1 − α)(lt−1 + bt−1)

bt = β(lt − lt−1) + (1 − β)bt−1

Explanation
Additive Holt-Winters smoothing

t −M t + 1−M t − 1 t t + 1

Yt − lt−1 − bt−1
St = γ(Yt − lt−1 − bt−1) + (1 − γ)St−MSt

lt = α(Yt − St−M ) + (1 − α)(lt−1 + bt−1)

bt = β(lt − lt−1) + (1 − β)bt−1

Explanation
Additive Holt-Winters smoothing

t −M t + 1−M t − 1 t t + 1

Yt − lt−1 − bt−1
St = γ(Yt − lt−1 − bt−1) + (1 − γ)St−MSt

lt = α(Yt − St−M ) + (1 − α)(lt−1 + bt−1)

bt = β(lt − lt−1) + (1 − β)bt−1

Explanation
Additive Holt-Winters smoothing

t −M t + 1−M t − 1 t t + 1

Yt − lt−1 − bt−1
St = γ(Yt − lt−1 − bt−1) + (1 − γ)St−MSt

lt = α(Yt − St−M ) + (1 − α)(lt−1 + bt−1)

bt = β(lt − lt−1) + (1 − β)bt−1

Visitor arrivals in Australia: Lecture06 Example01.py
Additive Holt-Winters method

Visitor arrivals in Australia
Additive Holt-Winters level component estimate

Visitor arrivals in Australia
Additive Holt-Winters seasonal factors

Additive Holt-Winters smoothing
Choice of initial values

How should we set the initial values l0, b0, s0, s−1, . . ., s2−M ,

Suggested Method

1 Do a linear least square regression over the data Y1, . . . ,YT
to find out

Ŷt = ω̂0 + ω̂1t

2 Take l0 = ω̂0 and b0 = ω̂1
3 Find out ŝt = Yt − Ŷt , then take the (seasonal) average of ŝt

as one of s0, s−1, . . ., s2−M , s1−M according to each season.

Additive Holt-Winters smoothing
Choice of initial values

How should we set the initial values l0, b0, s0, s−1, . . ., s2−M ,
Suggested Method

1 Do a linear least square regression over the data Y1, . . . ,YT
to find out

Ŷt = ω̂0 + ω̂1t

2 Take l0 = ω̂0 and b0 = ω̂1
3 Find out ŝt = Yt − Ŷt , then take the (seasonal) average of ŝt

as one of s0, s−1, . . ., s2−M , s1−M according to each season.

Additive Holt-Winters smoothing
Some notes

Useful when level and/or trend and seasonal variation is not
changing much

Choice of initial seasonal indices can be important.

Visitor arrivals in Australia
Additive Holt-Winters forecast

Visitor arrivals in Australia
Additive Holt-Winters forecast

Visitor arrivals in Australia
Additive Holt-Winters fit

Visitor arrivals in Australia
Additive Holt-Winters residuals

Visitor arrivals in Australia
Additive Holt-Winters residual autocorrelations

We will talk about this in Week 7

Alcohol related assaults in NSW
Additive Holt-Winters fit

Additive Holt-Winters smoothing

lt = α(Yt − St−M) + (1− α)(lt−1 + bt−1),
bt = β(lt − lt−1) + (1− β)bt−1,
St = γ(Yt − lt−1 − bt−1) + (1− γ)St−M ,

Yt+1 = lt + bt + St+1−M + εt+1, εt+1 ∼ N(0, σ2).

We can chose the parameters α, β and γ by minimising

(Yt − lt−1 − bt−1 − St−M)2

Additive Holt-Winters smoothing
Error correction formulation for level

lt = α(Yt − St−M) + (1− α)(lt−1 + bt−1)
= lt−1 + bt−1 + α(Yt − lt−1 − bt−1 − St−M)
= lt−1 + bt−1 + αεt

Additive Holt-Winters smoothing
Error correction formulation for trend

bt = β(lt − lt−1) + (1− β)bt−1
= bt−1 + β(lt − lt−1 − bt−1)
= bt−1 + βαεt

The second equation comes from the error form of lt on the
previous slide.

Additive Holt-Winters smoothing
Error correction formulation for seasonality

St = γ(Yt − lt−1 − bt−1) + (1− γ)St−M
= St−M + γ(Yt − lt−1 − bt−1 − St−M)
= St−M + γεt

The second equation comes from the error form of Yt in the model
definition.

Additive Holt-Winters smoothing
Error correction formulation

lt = lt−1 + bt−1 + αεt

bt = bt−1 + αβεt

St = St−M + γεt

Yt = lt−1 + bt−1 + St−M + εt

e.g . Yt+1 = lt + bt + St+1−M + εt+1

= (lt−1 + bt−1 + αεt) + (bt−1 + αβεt) + St−M+1 + εt+1

= lt−1 + 2bt−1 + St−M+1 + α(1 + β)εt + εt+1

Additive Holt-Winters smoothing
Forecasting equations

Ŷt+1|1:t = E(lt + bt + St−M+1 + εt+1|Y1:t)
= lt + bt + St−M+1

(= lt−1 + 2bt−1 + St−M+1 + α(1 + β)εt)

We cannot get rid of the last term with εt . Why?

Ŷt+2|1:t = E(lt+1 + bt+1 + St−M+2 + εt+2|Y1:t)
= E(lt + 2bt + St−M+2 + α(1 + β)εt+1 + εt+2|Y1:t)
= lt + 2bt + St−M+2

Ŷt+h|1:t = lt + hbt + St−M+(h mod M) Or

Ŷt+h|1:t = lt + hbt + St+h−M(k+1)

where k is the integer part of (h − 1)/M.

Additive Holt-Winters smoothing
Variance for interval forecasts

Yt+1 =lt + bt + St+1−M + εt+1

=lt−1 + 2bt−1 + St−M+1 + α(1 + β)εt + εt+1

V(Yt+1|Y1:t) = V(lt + bt + St−M+1 + εt+1|Y1:t)

V(Yt+2|Y1:t) = V(lt+1 + bt+1 + St−M+2 + εt+2|Y1:t)
= V(lt + 2bt + St−M+2 + α(1 + β)εt+1 + εt+2|Y1:t)
= σ2(1 + α2(1 + β)2)

Additive Holt-Winters smoothing
Variance for interval forecasts

Yt+1 =lt + bt + St+1−M + εt+1

=lt−1 + 2bt−1 + St−M+1 + α(1 + β)εt + εt+1

V(Yt+3|Y1:t) = V(lt+2 + bt+2 + St−M+3 + εt+3|Y1:t)
= V(lt+1 + 2bt+1 + St−M+3 + α(1 + β)εt+2 + εt+3|Y1:t)
= V(lt + 3bt + St−M+3 + α(1 + 2β)εt+1 + α(1 + β)εt+2 + εt+3)
= σ2(1 + α2(1 + β)2 + α2(1 + 2β)2)

V(Yt+h|Y1:t) = V

lt + hbt + St−M+h + α

(1 + (h − i)β)εt+i + εt+h|Y1:t

(1 + (h − i)β)2
, for h ≤ M only.

Additive Holt-Winters smoothing (optional)
Variance for interval forecasts

For h > M,

V(Yt+h|Y1:t) = V

lt + hbt + St−M+h + α

(1 + (h − i)β)εt+i + εt+h|Y1:t

lt + hbt + St−2M+h + γεt−M+h

(1 + (h − i)β)εt+i + εt+h|Y1:t

[α(1 + iβ) + Ii,Mγ]

where Ii,M = 1 if i is an integer multiple of M and 0 otherwise.

Additive Holt-Winters smoothing
Forecasting: collecting the results

Ŷt+h|1:t = lt + hbt + St−M+(h mod M).

V(Yt+h|Y1:t) = σ2

[α(1 + iβ) + Ii ,Mγ]

Alcohol related assaults in NSW
Additive Holt-Winters forecast

Alcohol related assaults in NSW
Additive Holt-Winters forecast

Multiplicative Holt-Winters smoothing

Most useful when the seasonal pattern changes in a strong
pattern and is proportional to the level of the series.

Multiplicative Holt-Winters smoothing

lt = α(Yt/St−M) + (1− α)(lt−1 + bt−1),
bt = β(lt − lt−1) + (1− β)bt−1,
St = γ(Yt/(lt−1 + bt−1)) + (1− γ)St−M ,

Yt+1 = (lt + bt)× St+1−M + εt+1, εt+1 ∼ N(0, σ2).

We can chose the parameters α, β and γ by minimising

(Yt − (lt−1 + bt−1)St−M)2

Multiplicative Holt-Winters smoothing
Error correction formulation for level

lt = α(Yt/St−M) + (1− α)(lt−1 + bt−1)
= lt−1 + bt−1 + α(Yt/St−M − lt−1 − bt−1)

= lt−1 + bt−1 + α

Yt − (lt−1+bt−1)St−M

= lt−1 + bt−1 + α

Multiplicative Holt-Winters smoothing
Error correction formulation for trend

bt = β(lt − lt−1) + (1− β)bt−1

= bt−1 + βα

Yt − (lt−1 + bt−1)St−M

see previous slide

= bt−1 + αβ

Multiplicative Holt-Winters smoothing
Error correction formulation for seasonality

St = γ(Yt/(lt−1 + bt−1)) + (1− γ)St−M

= St−M + γ
Yt − (lt−1 + bt−1)St−M

lt−1 + bt−1

= St−M + γ

lt−1 + bt−1

Multiplicative Holt-Winters smoothing
Error correction formulation

lt = lt−1 + bt−1 + α

bt = bt−1 + αβ

St = St−M + γ

lt−1 + bt−1

Yt = (lt−1 + bt−1)× St−M + εt

Multiplicative Holt-Winters smoothing
Forecasting equations

Ŷt+1|1:t = E((lt + bt)St−M+1 + εt+1|Y1:t)
= (lt + bt)St−M+1

Ŷt+2|1:t = E((lt+1 + bt+1)St−M+2 + εt+2|Y1:t)

lt + 2bt + α(1 + β)

St−M+2 + εt+2

= (lt + 2bt)St−M+2

Ŷt+h|1:t = (lt + hbt)St−M+h (h ≤ M)

For h > M, the formula is too complicated. We omit it!

Multiplicative Holt-Winters smoothing
Variance for interval forecasts

lt−1 + 2bt−1 + α(1 + β)

St−M+1 + εt+1

V(Yt+1|Y1:t) = V((lt + bt)St−M+1 + εt+1|Y1:t)

V(Yt+2|Y1:t) = V((lt+1 + bt+1)St−M+2 + εt+2|Y1:t)

lt + 2bt + α(1 + β)

St−M+2 + εt+2|Y1:t

= σ2(1 + α2(1 + β)2(S2t−M+2/S

Multiplicative Holt-Winters smoothing
Forecasting formula

Ŷt+h|1:t = (lt + hbt)× St+h−M .

where h ≤ M.

Dampened trend ES

Extrapolating trends indefinitely into the future can be problematic.

Dampened trend exponential smoothing aims to deal with this

Dampened trend ES
Illustration

Dampened trend ES

lt = αYt + (1− α)(lt−1 + φbt−1),
bt = β(lt − lt−1) + (1− β)φbt−1,

Yt+1 = lt + φbt + εt+1,

where φ is the dampening factor, with 0 ≤ φ ≤ 1.

Dampened trend ES
Forecasting and variance equations

Yt+1 = lt + φbt + εt+1

Ŷt+1|1:t = lt + φbt

V(Yt+1|Y1:t) = σ2

Dampened trend ES
Forecasting and variance equations

Yt+2 = lt+1 + φbt+1 + εt+2

= lt + φbt + φ
2bt + α(1 + φβ)εt+1 + εt+2

Ŷt+2|1:t = lt + bt(φ+ φ

V(Yt+2|Y1:t) = σ2(1 + α2(1 + φβ)2)

Dampened trend ES
Forecasting formula

Ŷt+h|1:t = lt + φbt + φ

3bt + . . .+ φ

Compared with the forecast of the trend correct exponential

Ŷt+h|1:t = lt + h × bt
What happens as h gets larger?

For the dampened forecast Ŷt+h|1:t → lt +

For the trend corrected forecast

Ŷt+h|1:t →∞

Dampened trend ES
Forecasting formula

Ŷt+h|1:t = lt + φbt + φ

3bt + . . .+ φ

Compared with the forecast of the trend correct exponential

Ŷt+h|1:t = lt + h × bt
What happens as h gets larger?
For the dampened forecast Ŷt+h|1:t → lt +

For the trend corrected forecast

Ŷt+h|1:t →∞

Dampened trend seasonal

lt = α(Yt − St−M) + (1− α)(lt−1 + φbt−1),
bt = β(lt − lt−1) + (1− β)φbt−1,
St = γ(Yt − lt) + (1− γ)St−M ,

(or St = γ(Yt − lt−1 − φbt−1) + (1− γ)St−M , )

Yt+1 = lt + φbt + St−M+1 + εt+1,

where φ is the dampening factor, with 0 ≤ φ ≤ 1.

Dampened trend seasonal
Forecasting formula

Ŷt+h|1:t = lt + φbt + φ

3bt + . . .+ φ
hbt + St+h−M

程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com