Subjective Objective

Judgmental Causal Techniques

• Sales Force Surveys Regression / Econometric Models

• Jury of Experts
• Delphi Techniques
Experimental Time Series Techniques

• Customer Surveys • Naïve Approach

• Focus Group sessions • Cumulative Approach
• Test Marketing • Moving Average (Simple / Weighted)
• Exponential Smoothing
 Arijit Mitra

z z
Objective Approaches
in Forecasting – The
Techniques
z
Time-Series Forecasts

▪ Forecasts that project patterns identified in recent time-series

observations
▪ Time-series - a time-ordered sequence of observations taken at
regular time intervals

▪ Assumption is that the future values of the time-series can be

estimated from past values of the time-series
Time-Series Behaviors: Decomposition of the data
z

▪ Level (a): The base, the constant part of the demand

▪ Trend (b): A long-term upward or downward movement in
data
▪ Population shifts
▪ Changing income
▪ Seasonality (S):
▪ Short-term, fairly regular variations related to the
calendar or time of day
▪ Restaurants, service call centers, and theaters all
experience seasonal demand
▪ Randomness (E): Residual variation that remains after all
other behaviors have been accounted for
z
Additive Model for Decomposition of Demand Behavior

▪ When the fluctuation of demand is almost constant with time,

Demand can be expressed as sum of all the components
▪ D = Level (a) + Trend (b*t) + Seasonal component (S) + Random Variations (E)

▪ This is known as additive model

Additive Decomposition is required to

understand different components of
such demand data
z
Multiplicative Model for Decomposition of Demand Behavior

▪ Multiplicative Model for Decomposition: When the fluctuation

of demand is varying with time, Demand can be expressed as
multiplication of all the components
▪ D = Level (a) X Trend (b*t) X Seasonal component (S) X Random Variations (E)

▪ This is known as multiplicative model

Demand

Time
z
The Mixed Model – The most practical One!

▪ Demand (D) = [𝒂 + 𝒃 ∗ 𝒕 ] ∗ 𝑺 + 𝑬
▪ Somewhere in between the additive and multiplicative models,
but this is way most time things happen!
▪ The Level and trend parts are additive, but this whole term is
multiplied by seasonality and then the errors are additive again
z
Naïve Method Vs Cumulative Approach

▪ Naïve – A forecast for any period that equals the previous

periods’ actual value. Very simple and inaccurate but
widely used.
▪ The last data gets the total importance and NO importance
for the historical values!
▪ Ft, (t+1) = Xt

▪ Cumulative – Where every historical value gets the same

importance, i.e., it is the simple average of all the historical
values (including the last one)
σ𝒕𝒊=𝟏 𝑿𝒊
▪ Ft, (t+1) =
𝒕
z
Time Series Forecasting: Some Methods along the continuum

▪ Naive Approach

▪ Exponential Smoothing

▪ Weighted Moving Average

▪ Simple Moving Average

▪ …………

▪ ……………

▪ …………..

▪ Cumulative Approach
 Naïve Approach
ORDERS
MONTH PER MONTH FORECAST
Jan 120
Feb 90
Mar 100
Apr 75
May 110
June 50
July 75
Aug 130
Sept 110
Oct 90
Nov -
Naïve Approach (Contd.)
ORDERS
MONTH PER MONTH FORECAST
Jan 120 -
Feb 90 120
Mar 100 90
Apr 75 100
May 110 75
June 50 110
July 75 50
Aug 130 75
Sept 110 130
Oct 90 110
Nov - 90
 z
Simple Moving Average

▪ Technique that averages a number of the most

recent actual values in generating a forecast
n
▪ As new data become available, the forecast is 
updated by adding the newest value and
Di
i=1
dropping the oldest and then re-computing the MAn =
average n
▪ The number of data points included in the
average determines the model’s sensitivity
where
▪ Fewer data points used-- more responsive
n = number of periods in the moving average
Di = demand in period i
▪ More data points used-- less responsive
 3-month Simple Moving Average
ORDERS MOVING
MONTH PER MONTH AVERAGE 3

Jan 120 
i=1
Di
Feb 90 MA3 =
Mar 100 3
Apr 75
May 110
June 50
July 75
Aug 130
Sept 110
Oct 90
Nov -
 3-month Simple Moving Average
ORDERS MOVING
MONTH PER MONTH AVERAGE 3

Jan 120 – 
i=1
Di
Feb 90 – MA3 =
Mar 100 – 3
Apr 75 103.3
May 110 88.3 90 + 110 + 130
June 50 95.0
= 3
July 75 78.3
Aug 130 78.3
= 110 orders for Nov
Sept 110 85.0
Oct 90 105.0
Nov - 110.0
 5-month Simple Moving Average
ORDERS MOVING
MONTH PER MONTH AVERAGE 5

Jan 120 
i=1
Di
Feb 90 MA5 =
Mar 100 5
Apr 75
May 110
June 50
July 75
Aug 130
Sept 110
Oct 90
Nov -
 5-month Simple Moving Average
ORDERS MOVING
MONTH PER MONTH AVERAGE 5

Jan 120 – 
i=1
Di
Feb 90 – MA5 =
Mar 100 – 5
Apr 75 –
May 110 – 90 + 110 + 130+75+50
= 5
June 50 99.0
July 75 85.0
Aug 130 82.0 = 91 orders for Nov
Sept 110 88.0
Oct 90 95.0
Nov - 91.0
 z
Weighted Moving Average
n
WMAn = 
i=1
Wi Di

where
Wi = the weight for period i, between 0 and 100 percent
 Wi = 1.00

▪ The most recent values in a time series are given more weight in
computing a forecast
▪ The choice of weights, w, is somewhat arbitrary and involves some trial
and error
Weighted Moving Average Example
MONTH WEIGHT DATA
August 17% 130
September 33% 110
October 50% 90
3
November Forecast WMA3 = 
i=1
W i Di
Weighted Moving Average Example
MONTH WEIGHT DATA
August 17% 130
September 33% 110
October 50% 90
3
November Forecast WMA3 = 
i=1
W i Di

= (0.50)(90) + (0.33)(110) + (0.17)(130)

= 103.4 orders
 z
Exponential Smoothing New Info Forecasted Info
(with Old info)

Ft =  Dt - 1 + (1 - )Ft - 1
▪ A weighted averaging method
that is based on the previous
where:
forecast plus a percentage of Ft +1 = forecast for next period
the forecast error Dt = actual demand for present
period
Ft = previously determined
forecast for present period
 = weighting factor, smoothing
constant
 Exponential Smoothing (α=0.30)

PERIOD MONTH DEMAND F2 = D1 + (1 - )F1

1 Jan 37
2 Feb 40
3 Mar 41
4 Apr 37
F3 = D2 + (1 - )F2
5 May 45
6 Jun 50
7 Jul 43
8 Aug 47
9 Sep 56 F13 = D12 + (1 - )F12
10 Oct 52
11 Nov 55
12 Dec 54
 Exponential Smoothing (α=0.30)

PERIOD MONTH DEMAND F2 = D1 + (1 - )F1

1 Jan 37
= (0.30)(37) + (0.70)(37)
2 Feb 40
3 Mar 41 = 37
4 Apr 37
F3 = D2 + (1 - )F2
5 May 45
6 Jun 50 = (0.30)(40) + (0.70)(37)
7 Jul 43 = 37.9
8 Aug 47
9 Sep 56 F13 = D12 + (1 - )F12
10 Oct 52 = (0.30)(54) + (0.70)(50.84)
11 Nov 55
= 51.79
12 Dec 54
 Exponential Smoothing
FORECAST, Ft + 1
PERIOD MONTH DEMAND ( = 0.3) ( = 0.5)
1 Jan 37 – –
2 Feb 40
3 Mar 41
4 Apr 37
5 May 45
6 Jun 50
7 Jul 43
8 Aug 47
9 Sep 56
10 Oct 52
11 Nov 55
12 Dec 54
13 Jan –
 Exponential Smoothing
FORECAST, Ft + 1
PERIOD MONTH DEMAND ( = 0.3) ( = 0.5)
1 Jan 37 – –
2 Feb 40 37.00 37.00
3 Mar 41 37.90 38.50
4 Apr 37 38.83 39.75
5 May 45 38.28 38.37
6 Jun 50 40.29 41.68
7 Jul 43 43.20 45.84
8 Aug 47 43.14 44.42
9 Sep 56 44.30 45.71
10 Oct 52 47.81 50.85
11 Nov 55 49.06 51.42
12 Dec 54 50.84 53.21
13 Jan – 51.79 53.61
z
Try this dataset for the Cumulative
Approach, it is simple!

▪ Average all the historical values that you have got so far!
 z
Forecasting Accuracy Metrics

MAD =
 Actual t − Forecastt MAD weights all errors evenly
n

 (Actual − Forecast t )
2
MSE weights errors according to
MSE = t
their squared values
n −1
Actual t − Forecast t
 Actual t
100 MAPE weights errors
according to relative error
MAPE =
n
z
Monitoring Forecasting Error: Tracking Signal

σ(𝑨𝒄𝒕𝒖𝒂𝒍𝒕 −𝑭𝒐𝒓𝒆𝒄𝒂𝒔𝒕𝒕 )
▪ Tracking Signal =
𝑴𝑨𝑫𝒕

▪ It is plotted and its deviation from a mean value of 0 is

observed to understand the accuracy of the forecasted
values.

▪ It is a control chart where the deviations from 0 have a lower

and upper limit (usually ± 4 to ± 5). If the deviations are
consistently out of that limits, the forecasting accuracy is
questionable, and method is biased.
 z
Tracking Signal Nomenclatures
σ(𝑨𝒄𝒕𝒖𝒂𝒍𝒕 −𝑭𝒐𝒓𝒆𝒄𝒂𝒔𝒕𝒕 )
▪ Tracking Signal =
𝑴𝑨𝑫𝒕
▪ σ(𝑨𝒄𝒕𝒖𝒂𝒍𝒕 − 𝑭𝒐𝒓𝒆𝒄𝒂𝒔𝒕𝒕 ) = Running sum of errors (with proper signs)

▪ MADt = Updated MAD in the t (each time we need to calculate the

MAD up to time period t)
▪ There are various method for calculating the updated MAD. I am
simply using the Running MAD i.e.
MADt = Average of all the abs(error) till time t
▪ Another approximated method is:

MADt = MAD(t-1) + 0.2*[abs(error) – MAD(t-1) ]

z
Causal Model for Forecasting: Associative Techniques

▪ When the demand data is dependent on some other

variable(s) than time or time-period / along with the
time

▪ If the number of Independent Variable is one, we use

simple regression

▪ If the number of Independent Variable is more than

one, we use multiple regression
z
Choosing a Forecasting Technique
▪ Factors to consider
▪ Cost
▪ Accuracy
▪ Availability of historical data
▪ Availability of forecasting software
▪ Time needed to gather and analyze data and prepare a
forecast
▪ Forecast horizon