OM 352 Reminders

• HW 2 is due at 11:59 pm on Wednesday 20

September
• Lab 3 video has been posted to eClass
• HW 1 solution has been posted to eClass on Thursday
14 September

19 Sep 2023 OM 352 1

 HW1, Q6
6. (1 pt.) The “April population in Years 1-29”
of the lake in the fish model is:
a) An input
b) An output
The “April population in Year 0” is actually an
input.

19 Sep 2023 OM 352 2

 This is just one example of many good policies you could have come up

HW1, Q9: One with:

Q9:

possible approach Let the April population exceeds 250,000 (half of capacity) and then
harvest the growth every year, except last year; in which we harvest all the
Aug population minus 50,000.

PAug = PApr + a  (1 – PApr / capacity)  PApr

Therefore, Growth = a  (1 – PApr / capacity)  PApr
Approach A: Visual Optimization Approach B: Calculus

19 Sep 2023 OM 352 3

 OM 352:
Lecture 5: Forecasting
DES = Double Exponential Smoothing
Seasonality
TES = Triple Exponential Smoothing
Prediction intervals
 DES‒Recap
• Initialization
– Level, Trend
• Learning (one-step ahead forecast)
L t LS Dt  (1  LS) (L t  1  Tt  1)
Tt TS (L t  L t  1)  (1  TS) Tt  1
F t = Lt + T t
• Prediction (k-step ahead forecast)
F t = Lt + k × T t
19 Sep 2023 OM 352 5
 Learning

L t LS Dt  (1  LS) (L t  1  Tt  1)

Tt TS (L t  L t  1)  (1  TS) Tt  1

In general: UPDATED = S  NEW + (1 – S)  OLD

19 Sep 2023 OM 352 6

 Active Learning
Which forecasting method
was used to compute the
forecasts shown with a
dashed curve?
a) SES w/LS = 0.5
b) DES w/LS = TS = 0.5

19 Sep 2023 OM 352 7

Finding the Best Values for LS and TS
 Pg. 34

Finding “the best”

parameter values
• Approach 1:
– Organize spreadsheet so that you can see
parameters, performance measures, and
graph of data and forecasts
– Experiment! Vary parameters to search for
better values We used this approach in
3_Performance_Measures_12_Sep_23.xlsx
• Approach 2:
– Use Solver to optimize a performance measure by
varying parameters

19 Sep 2023 OM 352 9

 What’s This Solver Thing?
To bring it up in
Excel:
Data  Solver
Optimize something
(maximize profit,
minimize cost, etc.)
By varying some
decision variables
(“changing cells”)

Keeping in mind any

restrictions
(“constraints”) on the
decision variables
19 Sep 2023 OM 352 10
 Pg. 34

Using Solver to Choose LS and TS

• What to optimize: minimize RMSE
– Could minimize MAD or MAPE, but solver
works more reliably with RMSE
• For the geeks: Because RMSE is a smooth function
• Decision variables: LS, TS
• Constraints:
Something a bit Something a bit
LS
bigger than zero
(e.g.: 0.01, 0.05)
≤ TS ≤ smaller than one
(e.g.: 0.99, 0.95)
19 Sep 2023 OM 352 11
4_DES_Solver_14_Sep_23.xlsx
 Solver Might Not Always Give
the Same Solution!!!
Everywhere I look is uphill!
RMSE
I must have reached the lowest point.

local optimum

global optimum
LS
19 Sep 2023
OM 352 12
 Non-linear Models
• Example: forecasting performance measures
• Solution method in Solver: GRG-Nonlinear
• Always check Mutlistart option

19 Sep 2023
OM 352 13
 Marking optimization problems in OM352

 Feasibility
• Your solution should satisfy all constraints (in which
case we call it a feasible solution)
 Consistency
• Your performance measure (objective) should have been
calculated correctly
 Optimality
• Your solution should be feasible and result in the
optimal (maximum or minimum) value for the
performance measure (objective)

19 Sep 2023 OM 352 14

Seasonality
 What does Seasonality Mean?
• Seasonality: Any predictable pattern that
repeats over a cycle of fixed length
• Active learning: Think of two examples of
seasonality
Variable Number of time periods per cycle
(p)

19 Sep 2023 OM 352 16

 How to measure Seasonality?
• Seasonality Index (SI)
• Informal definition: SI = actual / level
• Example:
– Average monthly sales = $100M
– July sales = $150M
– July SI = 150/100 = 1.5
• “SI = actual / level” means:
– Actual = level  SI
– Level = actual / SI
19 Sep 2023 OM 352 17
 Level, Trend, Seasonality
• Additive trend, multiplicative seasonality
• (Level + Trend)
 seasonality index
• Example:
– Level = 800
– Trend = 200
– Seasonality index: 1.3
– Forecast: (800 + 200)  1.3 = 1,300
19 Sep 2023 OM 352 18
Triple Exponential Smoothing
(TES)
 Triple Exponential Smoothing
(TES)
Works in three
phases
– Initialization Tracks three
– Learning
components
– Level
– Prediction
– Trend
– Seasonality

Let us see how:

5_TES_19_Sep_23.xlsx

19 Sep 2023 OM 352 20

 Pg. 30

TES – Learning
p = # of periods in each seasonal cycle
Dt
L t LS  (1  LS)(L t  1  Tt  1)
St  p

Tt TS(L t  L t  1)  (1  TS)Tt  1

Dt
St SS  (1  SS)S t  p
Lt

All exponential smoothing methods:

UPDATED = S  NEW + (1-S)  OLD

19 Sep 2023 OM 352 21

 Level: Learning phase

L(t) = LS * D(t) / S(t-p) + (1 - LS)*( L(t-1) + T(t-1) )

• NEW: D(t) / S(t-p) = de-seasonalize data for period t using
seasonality of corresponding previous season  level = actual / SI
• OLD: L(t-1) + T(t-1) = best previous estimate of level for period t
19 Sep 2023 OM 352 22
 Trend: Learning phase

T(t) = TS * ( L(t) - L(t-1) ) + (1 - TS) * T(t-1)

• NEW: L(t) - L(t-1) = growth from period t-1 to period t
• OLD: T(t-1) = best previous estimate for trend for period t

19 Sep 2023 OM 352 23

 Seasonality: Learning phase

S(t) = SS * D(t) / L(t) + ( 1 - SS ) * S(t-p)

• NEW: D(t) / L(t) = actual / level  SI = actual / level
• OLD: S(t-p) = previous SI estimate for corresponding season

19 Sep 2023 OM 352 24

 One-step forecasting: The past

F(t+1) = [L(t) + T(t)] * S(t+1-p)

"To forecast one step into the future, take the previous period’s level,
add the previous period’s trend, and multiply the sum with the
corresponding seasonality index from one cycle ago."

19 Sep 2023 OM 352 25

 Pg. 31

k-step forecasting: The future

(“real” forecast)

F(t+k) = [L(t) + k*T(t)] * S(t+k-p)

Active learning: Translate the formula into English

19 Sep 2023 OM 352 26

 Seasonality Index Lookup
The table where we look
The column we go to
The quarter we are looking for (Column 2) when we find
the quarter we are looking
for

19 Sep 2023 OM 352 27

Comparing SES,
DES, and TES
Performance Comparison
for Years 2-4
SES DES TES
BIAS 194.5 -136.0 43.1
MAD 280.5 378.7 105.2
RMSE 497.5 425.3 133.7
MAPE 22.2% 48.7% 14.4%

Active Learning:
Is the ranking of the methods the
same based on all performance
measures?

19 Sep 2023 OM 352 28

Review
Question

Which method was used to compute the red forecast?

a) Last point method
b) Simple exponential smoothing
c) Double exponential smoothing
d) Triple exponential smoothing

19 Sep 2023 OM 352 29

 Summary: How Different Methods
Predict
• Simple models: Ft+k = Ft+1, k = 2, 3, …
– Horizontal line
• DES: Ft+k = Lt + k  Tt , k = 1, 2, 3, …
– Linear trend
• TES: Ft+k = (Lt + k  Tt)  (Seas. Index)
k = 1, 2, 3, …
– Linear trend × seasonality

19 Sep 2023 OM 352 30

 Is TES it?
• Certainly not—many other methods exist
• Could have multiple seasonal cycles
• For example, both daily and weekly
• Seasonality could be additive instead of
multiplicative
• Seasonal Differences instead of Seasonal Indices
• Trend could be multiplicative instead of
additive
• L(t + k) = L(t) × (1 + T(t))k  Advantage: Will never
become negative
instead of
L(t + k) = L(t) + T(t) × k
19 Sep 2023 OM 352 31
Prediction Intervals
 Pg. 39

How Good are the Forecasts?

• TES (optimized):
Year 5, Quarter 1 sales = 1466.4
– Are you willing to bet on it?
• Put limits around a “point forecast”
– “Prediction Interval”
– 95% sure sales will be between low and high
– How do we compute low and high?

19 Sep 2023 OM 352 33

 Forecast Error Distribution
Optimized TES One-quarter-ahead
forecast errors
Greeting Cards Triple Exponential Smoothing
er period sales k forecast level trend quarter seasonal. error abs.error %error
1 1 222 1 0.50
2 2 339 2 0.76 Histogram
3 3 336 3 0.76
4 4 878 443.75 55.25 4 1.98
1 5 443 249.64 507.74 63.90 1 0.85 193.36 193.36 43.6%
2 6 413 436.70 570.94 63.21 2 0.73 -23.70 23.70 5.7%
3 7 398 480.16 631.69 60.78 3 0.64 -82.16 82.16 20.6%
4 8 1143 1370.11 689.87 58.21 4 1.68 -227.11 227.11 19.9%
1 9 695 635.33 749.67 59.78 1 0.92 59.67 59.67 8.6%
2 10 698 587.58 812.89 63.19 2 0.85 110.42 110.42 15.8%
3 11 737 558.92 882.39 69.43 3 0.82 178.08 178.08 24.2%
4 12 1648 1596.10 952.52 70.13 4 1.73 51.90 51.90 3.1%
1 13 1141 943.10 1027.49 74.93 1 1.10 197.90 197.90 17.3%
2 14 1036 937.48 1105.04 77.52 2 0.93 98.52 98.52 9.5%
3 15 938 973.17 1181.60 76.56 3 0.80 -35.17 35.17 3.7%
4 16 2168 2172.63 1258.10 76.50 4 1.72 -4.63 4.63 0.2%
1 17 1 1466.36 1 1.10
2 18 2 1315.27 2 0.93 LS= 0.022608 BIAS= 43.09
3 19 3 1183.63 3 0.80 TS= 0.99 MAD= 105.22
4 20 4 2695.69 4 1.72 SS= 0.937617 MSE= 17869.4
19 Sep 2023 OM 352 34
1 21 5 1802.59 1 1.10 RMSE= 133.6765
2 22 6 1600.51 2 0.93 MAPE= 14.4%
 Closer Look at the Histogram

19 Sep 2023 OM 352 35

Approximate with a Normal Distribution
Normal distribution with mean = 0 and
standard deviation = RMSE of forecast errors
± 2 standard
deviations

19 Sep 2023 OM 352 36

 95% Prediction Interval
• Point Forecast  2  RMSE
• Year 5, Quarter 1 sales
= 1466.4  2  133.7
= 1466.4  267.4
=[1199.0, 1733.7]
=[low, high]

19 Sep 2023 OM 352 37