Adaptive Forecasting

Prof. Suresh K Jakhar Indian Institute of Management

Lucknow

Introduction:

 ​The estimates of level, trend,

and seasonality ​are updated
after each demand observation.
  ​The
main advantage of this
technique is that ​estimates
incorporate all new data that are
observed.
Adaptive Forecasting
)(​ +=​ STLF
11​ ++​ tt t t ​

Where​L​t ​= estimate of level at the end of Period ​t

​
T​t ​= estimate of trend at the end of Period ​t ​St+1
=estimate of seasonal factor for Period ​t+1 ​F​t+1
=forecast of demand for Period ​t+q ​(made
Period ​t ​or earlier) ​Dt+1
​ =actual
​ demand
​ =​
observed in Period ​t Et+1 ​ ​– ​D​t+1 =
​ Ft+1 ​ forecast
error in Period ​t
Steps in Adaptive Forecasting
 ​Initialize
◦ ​Compute initial estimates of level (​L​0​),
trend (​T​0​), and ​seasonal factors (​S​1​,...,​S​p​)
 ​Forecast
◦ ​Given the estimates in Period t, forecast
​ ​ ​Estimate
demand for period ​t + 1
error
​ ​= ​F​t+1 ​– ​D​t+1 ​
◦ ​Compute error ​Et+1
Modify estimates
◦ ​Modify the estimates of level (​Lt​ ​+1​), trend
(​Tt​ ​+1​), and seasonal factor (​St+p
​ ​+1​), given the
error ​E​t​+1
Methods
 Average ​2) ​Simple
1) ​Moving
Exponential Smoothing ​3)
Trend-Corrected Exponential
Smoothing
(​Holt’s Model​) ​4) ​Trend- and
Seasonality-Corrected
Exponential
Smoothing (​Winter’s Model​)
Moving Average

 ​Usedwhen demand has ​no

observable trend or
seasonality
 Systematic component of demand =
level ​ ​The level in period ​t ​is the
average demand over the
​ eriods
last ​N p
L​t ​= (​D​t ​+ ​D​t-​ 1 +
​ ... + ​D​t– ​ 1​) / N ​Ft​ ​+1 =
​ ​N+ ​ ​Lt​
and ​Ft​ +​ ​n =
​ ​L​t 
​ ​After observing the
demand for period ​t + ​ 1, revise
the estimates
L​t+​ 1 =
​ (​Dt​ ​+1 +
​ ​D​t ​+ ... + D
​ ​t-​ ​N​+2​) / ​N, F​t​+2 =
​
L​t+​ 1
Moving Average Example

 ​A
supermarket has experienced
weekly demand of
milk of ​D​1 =
​ ​120, ​D2
​ = ​ 27, ​D​3 =
​ 1 ​ ​114,
and ​D4​ =
​ ​122 gallons over the past
four weeks

◦ ​Forecast demand for Period 5 using a

four-period moving average
◦ ​What is the forecast error if demand in
Period 5 turns ​out to be 125 gallons?
Moving Average Example

L​4 =
​ ​(​D4
​ + ​ +
​ ​D3 ​ ​D​2 + ​ )/4
​ ​D1​

= (​ 122 + 114 + 127 + 120)/4 = 120.75

 ​Forecast demand for Period 5

F5​ = ​ =
​ ​L4 ​ 120.75 gallons
  ​Errorif demand in Period 5 = 125
​ ​5 =
gallons E ​ ​5 –
​ F ​ =
​ ​D5 ​ 120.75 – 125 = –
4.25

• ​Revised demand
L5​ = ​ +
​ (​D5 ​ +
​ ​D4 ​ ​D​3 + ​ )/4
​ ​D2​

= (125 + 122 + 114 + 127)/4 = 122

Simple Exponential
Smoothing
 ​Used when demand has ​no
observable trend
or seasonality
Systematic component of demand =
level
 estimate of level, ​L​0​,
 ​Initial
assumed to be
the average of all historical data
Simple Exponential
Smoothing
Given data for Periods 1 to ​n

1​ D​
​ ​ n ​å
L​0 = ​
i​=1​
n​
i

Current forecast ​F​t+​ 1 =

​ ​Lt​ and
​ ​F​t​+​n =
​ ​L​t
Revised forecast using ​smoothing constant
​ 1)
(0 < ​α <
​ a​D​t​+1​+(1–​a​)​L​t
L​t+​ 1 =​
Simple Exponential
Smoothing
 ​Supermarket data
L​0 =
​ ​Di​

​ 1​
i=
4​
å ​/ 4 =120.75
F​1 =​ ​L0​ =120.75
​

E1​ = ​ ​F1​ – ​ ​D​1 =

​ 120.75 –120 = 0.75

​ a​D1​
L​1 =​ ​ +(1–​a​)​L​0
=0.1 ́ 120+0.9 ́ 120.75=120.68
Likewise...
Trend-Corrected Exponential
Smoothing (​Holt’s Model)​

 ​Appropriate when the demand

is assumed to
have a ​level and trend i​ n the
systematic ​component of
demand but ​no seasonality

Systematic component of
demand =
level + trend
Trend-Corrected Exponential
Smoothing (​Holt’s Model​)

 ​Obtain initial estimate of level and

trend by
running a linear regression
Dt​ ​= L0​ + ​ 
​ t T0 ​ ​In Period ​t​, the
forecast for next period is
 Ft​ +​ 1 =
​ ​Lt​ ​+ ​T​t ​Revised estimates for
Period ​t+1 ​Lt​ ​+1 =
​ ​αD​t+ ​ (1 – ​α​)(​Lt​ ​+
​ 1+
Tt​ ​) ​Tt​ ​+1 =
​ ​β(​ ​L​t​+1 –
​ ​L​t​) + (1 – ​β​)​T​t
Example:Holt’s Model ​ ​An
electronics manufacturer has
seen demand for its latest MP3
player increase over the ​past six
months. Observed demand (in
thousands) has been D​1 =​ 8,415;
D​2 =
​ 8,732; D​3 =
​ 9,014; D​4 =
​

9,808; D​5 =
​ 10,413; and D​6 =
​

11,961. Forecast demand for

Period 7 using trend-corrected
​
exponential smoothing with α =
0.1, β = 0.2.
Trend-Corrected Exponential
Smoothing (​Holt’s Model​)

 ​MP3 player demand

D1​ =
​ 8,415, ​D​2 =
​ 8,732, ​D​3 =
​ 9,014, ​D​4
= 9,808, ​D5​ =​ 10,413, ​D6​ =
​ 11,961, ​α
= 0.1, ​β =​ 0.2 ​ ​Using regression
analysis
L​0 =
​ 7,367 and ​T​0 =
​ 673 ​ ​Forecast
for Period 1
F1​ =
​ ​L​0 + ​ =
​ ​T0 ​ 7,367 + 673 = 8,040 ​
Period 1 error
E​1 = ​ –
​ ​F1 ​ =
​ ​D1 ​ 8,040 – 8,415 = –375

Trend-Corrected Exponential
Smoothing (Holt’s Model)

 ​Revised estimate
L​1 =
​ ​αD​1 +
​ (1 – ​α​)(​L0
​ +
​ ​T​0​)
= 0.1 x 8,415 + 0.9 x 8,040 = 8,078

T1​ =
​ ​β(​ ​L​1 – ​ ) + (1 – ​β)​ ​T​0 ​= 0.2 x
​ ​L0​

(8,078 – 7,367) + 0.8 x 673 = 681 

​
With new ​L1​ ​F2​ =
​ ​L​1 + ​ =
​ ​T1 ​ 8,078 +

681 = 8,759 ​ ​Continuing

F7​ = ​ +
​ ​L6 ​ ​T​6 =
​ 11,399 + 673 = 12,072
Trend- and
Seasonality-Corrected
Exponential Smoothing

 ​Appropriate when the

systematic component ​of
demand has a level, trend, and
seasonal ​factor
Systematic component = (level +
trend) x ​seasonal factor

F​t+​ 1 =
​ (​L​t ​+ ​Tt​ ​)​St​ ​+1
Trend- and Seasonality-Corrected
Exponential Smoothing

 ​Afterobserving demand for period ​t

+ 1, revise ​estimates for level, trend,
and seasonal factors

L​t+​ 1 =
​ ​α(​ ​Dt​ ​+1​/​St​ ​+1​) + (1 – ​α)​ (​Lt​ +
​ ​T​t)​

Tt​ ​+1 =
​ ​β​(​Lt​ ​+1 –
​ ​L​t)​ + (1 – ​β​)​Tt​

St​ ​+​p+​ 1 =
​ ​γ​(​Dt​ +
​ 1​/​L​t+
​ 1​) + (1 – ​γ)​ ​St​ ​+1

α ​= smoothing constant for level ​β = ​

​ smoothing
smoothing constant for trend ​γ =
constant for seasonal factor
Example: Winter’s Model
L0​ = ​ =
​ 18,439 ​T0 ​ = 0.47, ​S​2 =
​ 524 ​S1​ ​ 0.68, ​S​3
= 1.17, ​S​4 = ​ =
​ 1.67 ​F1 ​ (​L​0 +
​ ​T​0​)​S1
​ =
​ (18,439
+ 524)(0.47) = 8,913 The observed demand
for Period 1 = ​D​1 =
​ 8,000 Forecast error for
Period 1

= ​E1​ = ​ –
​ ​F1 ​ ​D​1 =
​ 8,913 – 8,000 = 913

Winter’s Model

 ​Assume ​ =
α ​ 0.1, ​β ​= 0.2, ​γ ​= 0.1;
revise estimates for ​level and trend
for period 1 and for seasonal factor
for Period 5
L1​ = ​ ​D1​​ /​S​1​) + (1 – ​α)​ (​L0​ +
​ ​α( ​)
​ ​T0​

= 0.1 x (8,000/0.47) + 0.9 x (18,439 + 524)

= 18,769 ​T​1 =
​ ​β​(​L​1 – ​ ) + (1 – ​β)​ ​T0​
​ ​L0​
= 0.2 x (18,769 – 18,439) + 0.8 x 524 = 485
S​5 =
​ ​γ​(​D1​
​ /​L1​​ ) + (1 – ​γ​)​S​1
= 0.1 x (8,000/18,769) + 0.9 x 0.47 = 0.47
F2​ = ​ +
​ (​L1 ​ ​T​1​)​S2
​ =
​ (18,769 + 485)0.68 =
13,093

Time Series Models

Forecasting Method Applicability ​Moving
average No trend or seasonality Simple
exponential smoothing
No trend or seasonality
Holt’s model Trend but no seasonality
Winter’s model Trend and seasonality