Demand Forecasting

• Introduction
• Demand Forecasting Techniques
– Qualitative Forecasting Methods
– Quantitative Forecasting Models
• Forecast Error
• Designing a Demand Forecasting System for PPC
• Case: Yankee Fork and Hoe Company
 Forecasting
 Effective planning requires matching product requirements of
customers with the capacity of operations

 Forecasting: Predicting future events, such as, customer

demand
 Demand forecast is generally daily/weekly/monthly/quarterly
figure for each product/product group in a geographic region
Product Group Product Zone Week 1 Week 2 Week 3 Week 4
Electronics Mobiles North India 120 135 150 135
Electronics Mobiles West India 100 115 130 115
Electronics Mobiles South India 140 155 170 155
Electronics Laptops North India 90 105 110 105
Electronics Laptops West India 80 95 100 95
Electronics Laptops South India 110 125 140 125
FMCG Snacks North India 200 215 230 215
FMCG Snacks West India 180 195 210 195
FMCG Snacks South India 220 235 250 235
 Forecasting Techniques
 Forecasting is both an art & a science. The
science part deals with mathematical
models, whereas the art part deals with
judgment, experience and intuition. Thus,
two broad methods

 Qualitative methods
 Based on subjective methods

 Quantitative methods
 Based on mathematical formulas
 Qualitative Forecasting Methods
(subjective)
Characteristics:
– Based on experts’ intuition, experience, and opinions
– Not repeatable by others
– Useful when:
• Forecast has a long time horizon
• Environment is unstable during the forecast period
• Little or no historical data is available
Common Methods:
– Jury of Executive Opinion
– Sales Force Composite
– Delphi Method
– Consumer/Market Survey
 Qualitative Forecasting Methods
Jury of Executive Opinion
– A qualitative forecasting method in which a group of senior
executives reviews historical sales data, market
intelligence, and internal reports.
– Each executive provides an estimate of future demand; these
estimates are discussed and consolidated into a single
forecast.

Main disadvantages-
– Senior executives may be removed from day-to-
day customer interactions.
– Forecasts may be influenced by organizational politics
or power dynamics rather than data.
 Qualitative Forecasting Methods
Sales Force Composite
– A qualitative forecasting method in which each
salesperson estimates future sales for their
assigned territory.
– Territory-level forecasts are aggregated to obtain
the overall demand forecast.
– Often called a grass-roots approach, as salespeople
are close to customers and market conditions.

Main disadvantages-
– Salespeople may overestimate demand to appear
aggressive or protect their territory/job.
– Salespeople may underestimate demand to secure
lower sales quotas and earn higher bonuses.
 Qualitative Forecasting Methods
Delphi Method
– A structured expert-based forecasting method.
– A coordinator gathers forecasts from a group of
external experts.
– A statistical summary of all responses is sent back to
the experts.
– Experts can revise their estimates anonymously, aiming to
reach consensus.
– Participant identities remain confidential to reduce bias.

Main Disadvantages
– Process can be time-consuming and lengthy.
Qualitative Forecasting Methods
Market Research
– A systematic approach to measure
consumer interest in a product or service.
– Involves developing hypotheses and
testing them through surveys,
questionnaires, or interviews.
– Helps in understanding demand,
preferences, and market trends before
forecasting.
Main Disadvantage
– Results may be biased or limited due to
poor response rates.
 Quantitative methods
(numeric, objective)
Characteristics:
• Based on mathematical relationships between variables
• Repeatable by others using the same data and method
• Useful when:
• Historical data is available
• Forecasting short- to medium-term demand
Types of Quantitative Models:
1. Time Series Methods
Analyze historical demand patterns to project future demand
Examples: Moving Average, Exponential Smoothing, Holt–Winters

2. Causal Models (Regression / Econometric Models)

Forecasts based on other independent variables affecting
demand Examples: Sales predicted from advertising spend, price,
GDP, etc.
 Quantitative models: Time series models

• A time series is a set of numbers where

the order or sequence of the numbers is
important, e.g., historical demand
• Analysis of the time series identifies patterns
• Once the patterns are identified, they can be
used to develop a forecast
 Patterns (Components) of Demand
1. Base/permanent (Horizontal)

Quantity
• Stable demand with no clear upward or downward trend
• Values fluctuate (randomness) around a constant average

2. Trend Time

Quantity
• Gradual increase or decrease
over time
• Indicates long-term growth or
decline

Time

3. Seasonal
• Predictable relative highs and lows within a fixed,

Quantity
repeating period of less than one year (e.g., year, month,
Year 1
or week).
• Same month/week tends to have similar relative demand
period after period Year 2
• Example (yearly): Higher ice cream sales in May-June
• Example (weekly): Higher soda sales on weekends

4. Cyclical
• Fluctuations over multiple years
• Influenced by economic or industry cycles
• Unlike seasonal patterns, cycle length is variable and not fixed
 Components of Demand in Time Series
• Individual components
– Permanent (Base) component, B
– Trend, T
– Seasonal, S
– Cyclic, C
– Promotion, P
– Random, et
• Combining different components
Yt = (B + T) + S + et , additive seasonality

Yt = (B + T) x S + et , multiplicative seasonality
 Time-series method
Naive forecasting
Method 1A – Naive Method Method 1B – Naive Method with Trend
(Basic) Adjustment
Forecast = Last actual value Forecast = Last actual + change from
Formula: previous period
Ft+1=At
Formula:
Example:
Ft+1=At+(At−At−1)
August actual = 120 This includes the last period's trend in the
→ Forecast for September = forecast.
120 (Aug. actual) Example:
✅ Very simple
• July = 110
❌ Ignores trend and
• August = 120
seasonality • Trend = 120 (Aug)−110(Jul) =10
→ Forecast for September =
120 (Aug) + 10 (Trend)=130
✅ Simple trend-aware
❌ Assumes trend is linear and constant
 Method 1 – Simple Moving Average
SMA(n)
Idea:
Forecast = Average of last n periods
Formula:
Ft+1 = [At + At-1 + ... + At-n+1 ]/n
Example (3-period SMA):
Jan = 100, Feb = 110, Mar = 150
→ Forecast for April = (100+110+150)/3 =120
✅ Smooths out noise
❌ Slow to react to changes
• Small n, more responsive to changes in data
• Large n, more stable forecasts over time
• n is typically between 3 & 10
 Method 2 – Weighted Moving
Average, WMA(n)
Idea: Moving averages are unresponsive/sluggish to change.
Recent data is more important.
Ft+1 = w1 At + w2 At-1 + ... + wn At-n+1
Example (weights: 0.5, 0.3, 0.2):
Jan = 100, Feb = 110, Mar = 150
→ Forecast for April = 0.5(150)+0.3(110)+0.2(100)= 128
✅ More responsive
❌ Must select weights
One method to assign weights (not the only method):
– w1 > w2 > ... > wn > 0, weights sum to 1
– Sum-of-digits weights
S= 1+2+...+n
w1 = n/S, w2 = (n-1)/S, ……, wn = 1/S
 Method 3 –Simple Exponential Smoothing,
SES (
SES reduces data needs by implicitly weighting past values — a
smarter form of moving average.
[Note: Ft+1 =  At +   At-1 +   At-2 +…… +  t F0 ]
Idea:
Forecast based on the last forecast and last actual value, with
smoothing factor α (0 < α ≤ 1)
Ft+1 =  At + (1-  Ft
Or Ft+1 = Ft + ( At - Ft), Here At - Ft = forecasting error in period t
Example:
α=0.3 March actual = 110 March forecast = 100
→ April forecast = 0.3(110)+0.7(100)=103
✅ Simple, smooths data
✅ Adapts to change (with right α)
• Larger  values make forecast more responsive
• If =1, naive model
• In practice, 0.05    0.30
❌ Needs initial forecast and parameter tuning (can optimize using excel solver)
Time-Series Methods
Exponential Smoothing

450 — 6-week MA
3-week MA
forecast forecast
430 —
Patient arrivals

410 —

390 —

370 — Exponential
smoothing
 = 0.10
| | | | | |
0 5 10 15 20 25 30
Week
Choosing a Method
FORECAST ACCURACY

FORECAST ACCURACY refers to the

difference between forecasts and
corresponding actual sales.
Choosing a Method
Forecast Error

Measures of Forecast Error

E t = A t – Ft

CFE = Et
E 2
t
MSE =
n
|Et | [|Et | /At *(100)]
MAD MAPE =
n n

=
Choosing a Method
Forecast Error

Absolute
Error Absolute Percent
Month, Demand, Forecast, Error, Squared, Error, Error,
t At Ft Et E t2 |Et| (|Et|/At)(100)
1 200 225 -25 625 25 12.5%
2 240 220 20 400 20 8.3
3 300 285 15 225 15 5.0
4 270 290 –20 400 20 7.4
5 230 250 –20 400 20 8.7
6 260 240 20 400 20 7.7
7 210 250 –40 1600 40 19.0
8 275 240 35 1225 35 12.7
Total –15 5275 195 81.3%
Choosing a Method
Forecast Error
Measures of Error
Absolute
CFE = – 15 Error Absolute Percent
– 15Demand, Forecast,
Month, Error, Squared, Error, Error,
E =t =D–t 1.875 Ft |Et| (|Et|/At)(100)
Et E t2
8
1 200 225 –25 625 25 12.5%
2 5275240 220 20 400 20 8.3
3 300 285 15 225 15 5.0
MSE 4= = 659.4 –20 400 20 7.4
8 270 290
5 230 250 –20 400 20 8.7
6 260 240 20 400 20 7.7
7 210 250 –40 1600 40 19.0
195 35 1225 35 12.7
8 275 240
MAD = = 24.4 Total –15 5275 195 81.3%
8
81.3%
MAPE = = 10.2%
8
 Tracking Signal (TS) – Monitoring Forecast Bias
Definition:
• A Tracking Signal (TS) is a measure used in demand forecasting to detect bias in
forecasts.
• It indicates whether a forecast consistently overestimates or underestimates actual
demand. Cumulative Forecast Error (CFE)
Formula: TSt = Mean Absolute Deviation (MAD)
Purpose:
• Helps monitor forecast accuracy over time.
• Alerts when corrective action is needed (adjust forecast method, parameters, or
model).
Interpretation:
• TS ≈ 0: Forecast is unbiased (errors cancel out over time).
• TS > +4 or TS < -4 (common rule of thumb): Forecast is biased → corrective
action needed.
• Positive TS: Forecast consistently underestimates demand.
• Negative TS: Forecast consistently overestimates demand.
TS = Observation:
t At Ft Et=At-Ft CFE MAD CFE/MAD
• TS = 4 →
Example

1 100 95 5 5 5/1=5 5/5=1.0 Approaching bias

2 105 98 7 5+7=12 (5+7)/2=6 12/6=2.0 threshold and
3 103 100
⇒ systematic under-
3 12+3=15 (12+3)/3=5 15/5=3.0 • Errors Et on one side → TS > 0
4 107 102 5 15+5=20 (15+5)/4=5 20/5=4.0
estimation.
Key Points:
• Tracking Signal is used alongside other forecast accuracy metrics (MAD, MAPE,
RMSE).
• Provides an early warning for systematic forecast errors.
 Method 4- Forecasting
Seasonality and Trend
Understanding Trend: Example: Four year quarterly demand data
Quarter Year 1 Year 2 Year 3 Year 4
1 45 70 100 100
2 335 370 585 725
3 520 590 830 1160
4 100 170 285 215
Yearly Total 1000 1200 1800 2200
Step 1 Average per quarter 250 300 450 550
TREND: Average demand increase per year or trend = (2200-1000)/3 = 400 units per year
Projected annual demand for Year 5 incorporating trend: 2200 (year 4) + 400 (trend) =2600 units

Seasonality: Quarterly demand consistently deviates from the annual average in

a predictable pattern. (trick to identify: you can rank quarterly demand and may
notice same ranking each year, e.g., Q1 < Q4 < Q2 < Q3).
Seasonality Adjustment – Steps (Customized for This Example):
1. Calculate the average demand per season (i.e., per quarter).
2. Compute the seasonal index for each quarter:
Seasonal Index = Actual demand for the quarter ÷ Average demand across the four
seasons (quarters) of the same year.
3. Compute base quarterly forecast for Year 5 after adjusting for the trend
4. Forecast quarterly demand for year 5 by adjusting the base quarterly forecast using the
corresponding seasonal index to account for seasonal variation.
Trend+Seasonality-Adjusted Forecast =Trend adjusted Base forecast× Seasonal Index
Time-Series Methods
Step 2: Compute Seasonality Indices
Quarter Year 1 Year 2 Year 3 Year 4
1 45/250 = 0.18 70/300 = 0.23 100/450 = 0.22 100/550 = 0.18
2 335/250 = 1.34 370/300 = 1.23 585/450 = 1.30 725/550 = 1.32
3 520/250 = 2.08 590/300 = 1.97 830/450 = 1.84 1160/550 = 2.11
4 100/250 = 0.40 170/300 = 0.57 285/450 = 0.63 215/550 = 0.39

Quarter Average Seasonal Index

1 (0.18 + 0.23 + 0.22 + 0.18)/4 = 0.20
2 (1.34 + 1.23 + 1.30 + 1.32)/4 = 1.30
3 (2.08 + 1.97 + 1.84 + 2.11)/4 = 2.00
4 (0.40 + 0.57 + 0.63 + 0.39)/4 = 0.50
Sum = 4.00
Note: If the seasonal indices do not sum exactly to 4, adjust
them proportionally so that their total equals 4. If you are
doing monthly then the sum should be 12.
Time-Series Methods
Steps 3 & 4-Forecast: Trend+ Seasonal Influences

Step 3: Projected Annual Demand = 2600

Average Quarterly Demand = 2600/4 =
650

Step 4
Quarter Average Seasonal Index Forecast
1 (0.18 + 0.23 + 0.22 + 0.18)/4 = 0.20 650(0.20) = 130
2 (1.34 + 1.23 + 1.30 + 1.32)/4 = 1.30 650(1.30) = 845
3 (2.08 + 1.97 + 1.84 + 2.11)/4 = 2.00 650(2.00) = 1300
4 (0.40 + 0.57 + 0.63 + 0.39)/4 = 0.50 650(0.50) = 325
Linear Trend With Seasonality
A more refined method 1 2 170 160.1 9.9
1 3 140 172.3 -
370 32.3

A different carpet example

1 4 150 184.5 -
34.5
2 1 230 196.7
320 33.3

• Suppose you forecast 2 2

31.1
Carpe
240 108.9

270 2 3 t Fcst 180 221.1 -

using linear

Sales
41.1
2 4 200 233.3 -
33.3
220
trend model, Ft = a + bt +
3 1 320 245.5
64.5
3 2 310 257.5

et
52.3
3 3 230 269.9 -
39.9
3 4 260 282.1 -
Ft = 135.7 + 12.2 170 22.1

t Carpet Fcst
MAD = 33.9

– 1st & 2nd quarter errors positive 160

120170
147.9
160.1

– 3rd & 4th quarter errors Year 2

negative Year 1 Year 3
Time
Seasonality present
• Use a correct model, say model
Multiplicative seasonality Ft = [a + bt] cs + et
 260 282.1

cs = quarterly seasonal factor

Note: Linear trend (without seasonality) applied to carpet data
At Ft
t Yr Qtr Act. Fcst. Errors Error2
1 1 1 160 147.9 +12.1 146.4
2 1 2 170 160.1 +9.9 98.0
3 1 3 140 172.3 -32.3 1043.3
4 1 4 150 184.5 -34.5 1190.3
5 2 1 230 196.7 +33.3 1108.9
6 2 2 240 108.9 +31.1 967.2
7 2 3 180 221.1 -41.1 1689.2
8 2 4 200 233.3 -33.3 1108.9
9 3 1 310 245.5 +64.5 4160.3
10 3 2 310 257.5 +52.3 2735.3
11 3 3 230 269.9 -39.9 1592.0
12 3 4 260 282.1 -22.1 488.4

MAD = 33.9 MSE = 1360.7

Detailed Procedure for Seasonality Inclusion – Steps:
1. Compute 4-quarter moving averages (MA)
For quarterly data, take the average of 4 consecutive quarters:
MA1=(Q1 + Q2 + Q3 + Q4)/ 4, MA2=(Q2 + Q3 + Q4 + Q5)/4
With 3 years (12 quarters), you will get 9 moving averages (quarters 1–4, 2–5, …, 9–12).
2. Compute centered moving averages (CMA)
Since a 4-quarter MA is not centered, take the average of two consecutive MAs:
CMA=(MAt+MAt+1)2
3. Compute seasonal ratios
For each quarter where CMA is available:
Seasonal Ratio = Actual Demand /CMA
4. Estimate seasonal indices (cq)
For each quarter (Q1–Q4), average the seasonal ratios across the years:
cq=Average of seasonal ratios for quarter q
(Normalize seasonal indices: If the sum of the four quarterly indices ≠ 4, adjust
proportionally so the total equals 4.)
5. Deseasonalize the data
Remove seasonality:
Deseasonalized Demand=Actual Demand/ cq
6. Estimate trend parameters (a and b) for deseasonilzed forecast
Fit a trend model (e.g., Y= a +b t) to the deseasonalized data using regression. Here
t will be number consecutively
7. Forecast deseasonalized demand
Use the trend equation to predict future deseasonalized values.
8. Apply seasonality to obtain final forecast
Multiply by seasonal index:
Forecast=Deseasonalized Forecast×cq
 Step1 Step 2 Step3 Step4 Step5
t Yr Qtr Act. MA CMA SI cs Deseas.
1 1 1 160 1.191 134 (=160/1.1 91)

2 1 2 170 1.153 147

t=2.5 155.0
3 1 3 140 163.75 140/163.75= 0.855* 0.830* 169
t=3.5 172.5
4 1 4 150 181.25 0.828 0.826 182
190.0
5 2 1 230 195.00 1.179 1.191 193
200.0
6 2 2 240 206.25 1.164 1.153 208
212.5
7 2 3 180 222.50 0.809* 0.830 217
232.5
8 2 4 200 241.25 0.829 0.826 242
250.0
9 3 1 310 256.25 1.210 1.191 260
262.5
10 3 2 310 270.00 1.148 1.153 269
277.5
11 3 3 230 c1 (1.179 + 1.210)/2 = 1.195 x [4/4.011] = 0.830 277
c=2 1.195 +
(1.164 1.191 x
1.148)/2 = 1.156 [4/4.011] =
12 3 4 260 0.826 315
c=3 1.156 +
(0.855 0.809)/2= 1.153 x
0.832 [4/4.011] =
Step4 c= 0.832
(0.828 + 0.829)/2 = 0.830
0.828 x [4/4.011] =
4
=su 0.828 4.0110.826 4.000
m
Step 6 : Determine regression parameters
Use deseasonalized data, run regression using software to compute a & b
b = 15.4, a = 117.6  deasonalized trend forecast, Ft = 117.6 + 15.4 t
Step 7 : Forecast, demonstrated here for year 2, Quarter 2. e.g., period 6 )
Ft = [117.6 + 15.4 t] cs
Ft = [117.6 + 15.4 (6)] 1.153 = 242

At Ft
t Yr Qtr Act. Fcst Error Error2
1 1 1 160 158 +2 4
2 1 2 170 171 -1 1
3 1 3 140 136 +4 16
4 1 4 150 148 +2 4
5 2 1 230 232 -2 4
6 2 2 240 242 -2 4
7 2 3 180 187 -7 49
8 2 4 200 199 +1 1
9 3 1 310 305 +5 25
10 3 2 310 313 -3 9
11 3 3 230 238 -8 64
12 3 4 260 250 +10 100

MAD = 3.9 MSE = 23.4

 Method 5 –Holt’s Linear Exponential Smoothing
(Also called Double or Trend Adjusted Exponential Smoothing)
Use: Time series with trend, no seasonality Aₜ
Lₜ
Idea: Forecastt = Levelt-1 + Trendt-1 Forecastₜ
Formula: (α, βare smoothing constants for level and Tₜ
Lₜ₋₁
-1

trend) 1
t-1 t
Level or Average (Lₜ) = α × Aₜ + (1 – α)(Lₜ₋₁ + Tₜ₋₁)
Trend (Tₜ) = β × (Lₜ - Lₜ₋₁) + (1 – β)Tₜ₋₁ Fₜ
Forecast (h periods ahead): Fₜ₊ₕ = Lₜ + h* Tₜ
Example: Given: α=0.4, β=0.3 and initialization for quarter 1: L1=50,T1=4
Task: Show computation from Quarter 2 onwards and forecast for first 3 qtrs. In next year

t Aₜ Lₜ Tₜ
2 54 0.4(54)+0.6(50+4)=54.0 0.3(54-50)+0.7(4)=4.00
3 58 0.4(58)+0.6(54+4)=58.0 0.3(58-54)+0.7(4)=4.00
4 63 0.4(63)+0.6(58+4)=62.2 0.3(62.2-58)+0.7(4)=4.06

F5 = L4 + 1*T4 = 62.2+ 4.06 =

66.26 F6 = L4 + 2*T4 = 62.2 +
8.12 = 70.32 F7 = L4 + 3*T4 =
62.2 +12.18 = 74.38
 Method 6: Holt-Winter Method-additive
(Triple Exponential Smoothing)
Use: Data with trend + constant seasonality effect.
Idea: Forecast (additive) = Level + Trend +Seasonality
Formula (Additive Seasonality):(α, β, γ are smoothing constants; s = season length)
Level (Lₜ) = α × (Aₜ - Sₜ₋ₛ) + (1 – α)(Lₜ₋₁ + Tₜ₋₁) Aₜ
Trend (Tₜ) = β × (Lₜ - Lₜ₋₁) + (1 – β)Tₜ₋₁ Sₜ
Seasonal (Sₜ) = γ × (Aₜ- Lₜ) + (1 – γ)Sₜ₋ₛ Sₜ=Aₜ- Lₜ
Lₜ
Forecast (h periods ahead): Fₜ₊ₕ = Lₜ + h* Tₜ + Sₜ₋ₛ₊ₕ
t

Example: Given: s=4 α=0.4, β=0.3, γ=0.2; Initial values L4=100,T4=5, S1=−10, S2=−5, S3=5,
S4=10

Period 5 Calculation A5=110, S5−4 = S1= −10

L5 =0.4(110 - (-10)) + 0.6(100 + 5) =111.0

T5 =0.3(111 - 100) + 0.7(5) =6.8
S5 =0.2(110 - 111) + 0.8(-10) =−8.2
Forecasts
F6=L5+1*T5+S2=111+6.8−5=112.8
F7=L5+2*T5+S3=111+13.6+5=129.6
F8=L5+3*T5+S4=111+20.4+10=141.4
 Method 6: Holt-Winter Method-Multiplicative
(Triple Exponential Smoothing)
Use: Data with additive trend + proportional seasonal effect
Idea: Forecast (Multiplicative) = (Level + Trend)*Seasonality
(Note: multiplicative trend is also possible).
Formula (Multiplicative Seasonality):(α, β, γ are smoothing constants; s = season length)
Level (Lₜ) = α × (Aₜ/Sₜ₋ₛ) + (1 – α)(Lₜ₋₁ + Tₜ₋₁) Aₜ
Trend (Tₜ) = β × (Lₜ - Lₜ₋₁) + (1 – β)Tₜ₋₁
Seasonal (Sₜ) = γ × (Aₜ/Lₜ) + (1 – γ)Sₜ₋ₛ Sₜ=Aₜ/Lₜ
Lₜ
Forecast (h periods ahead): Fₜ₊ₕ = (Lₜ + h* Tₜ )* Sₜ₋ₛ₊ₕ
t
Example: Given: s=4 α=0.4, β=0.3, γ=0.2, Initial values L4=100,T4=5, S1=0.9, S2=0.95, S3=1.05,
S4=1.1

Period 5 Calculation A5=110, S5−4 = S1= 0.90

L5 =0.4(110/0.90) + 0.6(100 + 5) =116

T5 =0.3(116 - 100) + 0.7(5) = 8.3
S5 =0.2(110/116) + 0.8(0.9) = 0.916
Forecasts
F6 = (L5+1*T5)*S2 = (116+8.3)*0.95
= 118.6 F7 = (L5+2*T5)*S3 =
(116+16.6)*1.05 =137.6 F8 =
(L5+3*T5)*S4 = (116+24.9)*1.10
=157.3
Forecasting Techniques: Causal/Regression
Model

Forecast by Regression/Causal methods

estimates sales on the basis of values of
other independent factors. Use historical
data on independent variables, such as
promotional campaigns, economic
conditions and competitors actions to
predict demand
Quantitative Approach: Causal
Method/Regression Model
(Linear Regression)
Y Deviation,
Regression
Estimate of or error
Y from equation:
Dependent variable

regression Y = a + bX
equation

{ Actual
value
of Y

Value of X used
to estimate Y

X
Independent variable
 Causal Methods
Linear Regression

Sales Advertising
Month (000 units) (000 $)
a 1 264 2.5 = – 8.137
b 2 116 1.3 = 109.23
3 165 1.4
r = 0.98
4 101 1.0
5 209 2.0 r2 = 0.96
Causal Methods
Linear Regression
300 —
Sales (thousands of units)

250 —
Sales Advertising
Month (000 units) (000 $)
200 —
1 264 2.5 a = – 8.137
2—
150 116 1.3 b = 109.23
3 165 1.4 r = 0.98
4—
100 101 1.0 r2 = 0.96
5 209
Y= – 8.137 +2.0109.23X
50 | | | |
Forecast
1.0 1.5for Month
2.0 6
2.5
Advertising (thousands of dollars)
X = $1750, Y = 183.015, or 183,015 units
Causal Methods
Linear Regression (formula)
Sales, Y Advertising, X
Month (000 units) (000 $) XY X2 Y2
1 264 2.5 660.0 6.25 69,696
2 116 1.3 150.8 1.69 13,456
3 165 1.4 231.0 1.96 27,225
4 101 1.0 101.0 1.00 10,201
5 209 2.0 418.0 4.00 43,681
Total 855 8.2 1560.8 14.90 164,259
Y = 171 X = 1.64
XY – nXY
a = Y – bX b=
X 2 – n(X )2
Causal Methods
Linear Regression
Sales, Y Advertising, X
Month (000 units) (000 $) XY X2 Y2
1 264 2.5 660.0 6.25 69,696
2 116 1.3 150.8 1.69 13,456
3 165 1.4 231.0 1.96 27,225
4 101 1.0 101.0 1.00 10,201
5 209 2.0 418.0 4.00 43,681
Total 855 8.2 1560.8 14.90 164,259
Y = 171 X = 1.64
1560.8 – 5(1.64)(171)
a = Y – bX b=
14.90 – 5(1.64)2
Causal Methods
Linear Regression
Sales, Y Advertising, X
Month (000 units) (000 $) XY X2 Y2
1 264 2.5 660.0 6.25 69,696
2 116 1.3 150.8 1.69 13,456
3 165 1.4 231.0 1.96 27,225
4 101 1.0 101.0 1.00 10,201
5 209 2.0 418.0 4.00 43,681
Total 855 8.2 1560.8 14.90 164,259
Y = 171 X = 1.64

nXY – X Y
r=
[nX 2 – (X) 2][nY 2 – (Y) 2]
 Coefficient of Correlation (r)
Coefficient of Correlation (r)
• r measures the direction and strength of the linear
relationship between x and y.
• Sign of r → direction of the relationship (same sign as
regression
slope b).
• |r| → strength of the relationship.
• Range: −1 to +1.

Interpretation of Common Values

• r = −1 → perfect negative relationship (x ↑, y ↓).
• r = +1 → perfect positive relationship (x ↑, y ↑).
• r = 0 → no linear relationship.
• r = +0.3 → weak positive relationship.
• r = −0.8 → strong negative relationship.
 Designing a Demand Forecasting System:
some considerations
 Time Frame (How far to forecast?)
 Short-range, medium-range, long-range
 Appropriate Variable to Forecast,
Level of aggregation and Units
of measure (What to forecast?)
 Forecasting Technique (How
to forecast?)
 Purpose of forecast and decisions from it
 Time and effort required
 Data availability
 Examples of Production Resource
Forecasts
Item Being
Forecast Time Unit of
Forecasted
Horizon Span Measure
(level of
aggregation)
Long Product Lines, Dollars,
Years
Range Factory Capacities Tons
Medium Product Groups, Units,
Months
Range Depart. Capacities Pounds
Short Days, Specific Products, Units,
Range Weeks Machine Capacities Hours
Demand Forecast Applications
Time Horizon
Medium Long Term
Short Term Term (3 (more than
Application (0–3 months– 2 years)
months) 2 years)
Forecast quantity Individual Total sales Total sales
products or Groups or
services families
of products or
Decision area Inventory services Facility location
management Staff planning Capacity
Final assembly Production planning
scheduling planning Process
Workforce Master production management
scheduling scheduling
Master production Purchasing
scheduling Distribution
Forecasting Time series Causal
technique Causal Causal Judgment
Judgment Judgment
 Further Comments on Forecasting Practice
1. Advanced models and implementation concepts
 Models to handle nonlinear trends and multiple variables
 ARIMA (Box-Jenkins) models captures autocorrelation and trends in time series
 Machine learning and AI based forecast
 Use training/testing split for model evaluation
 Combine outputs from multiple models (e.g., average or focus forecasting)
o Focus forecasting selects the best model dynamically per period

2. Independent vs. Dependent Demand

 Forecast only independent demand
 Dependent demand is calculated from MRP or BOM structures

3. Hierarchical (Pyramid) Forecasting

$
busines
 Forecasts are created at multiple levels: s
total

o Level 1: Total business Level 1

X
o Level 2: Product lines Z
Force
Roll
o Level 3: Individual items down
up Product
lines
 Two key methods: Level 2
o Top-down (aggregate → disaggregate)
o Bottom-up (component → aggregate) X1 X2 .
Z1 Z9
 Ensures alignment between strategic and Individual items:
Level 3
operational plans via roll-up and force-down
adjustments
 Pyramid Forecasting
19.11x9000+18.24x12
500
$ total business: Level 1
Business $450,000
forecast Roll- $399,990
up forecast Force down
X Z
Roll up
Line forecast (units) 9,500 12,000 Product lines: Level 2
Roll-up forecast 9,000 12,500
(units) Average $19.1 $18.24
price 1
Roll
up Z1 ….
X1
4,00 X2 Z9
Individual items: Level
0 5,00 3
Individual forecast $20 0
(units) Unit price $18
Roll up for X:
4000+5000 = 9000
units
(20x4000+18x5000)/9000=19.11
 Pyramid Forecasting
An agreed upon
reconciled value
between
$450,000 &
$399,990
Manageme $ total business: Level 1
nt
forecast $420,00
0 Force
down

X Z
Forced forecast (units)
945 1312 Product lines: Level 2
0 5 420,000/399990*12500=13125
420,000/399990*9000=

9450

Forced forecast
X1 X2 ….
(units)
420 Individual items: Level
0 Z1 Z9 3
9450/9000*4000=42
00 5250
9450/90000*5000=5250
Thank You