Forecasting

ISE 182106 – Perencanaan dan Pengendalian Produksi

Reference
R1. Jacobs, F. Robert; Berry, William; Whybark, D. Clay; and Vollmann,
Thomas. Manufacturing Planning and Control for Supply Chain
Management. US: McGraw-Hill Professional, 2011. Chapter 3.
Outline

Forecast Definition

Regression Analysis

Decomposition of a Time Series

Short-Term Forecasting Techniques

Using the Forecasts

 R5-1, p. 77: Forecasting

Definitions
q Forecast
Is an estimation of future value of characteristic.

q Forecasting
Is the act of making such estimation.

q Prediction
Is estimation that includes subjective evaluation.
Forecasting in MPC
q In demand management, forecasts of the quantities and timing of
customer demand are developed.
q Forecast are estimates of what might occur in the marketplace.
q Manufacturing plans that specify how the firm will respond are
based on these forecasts.
Distinction between forecasts and plans
q Forecast à Manager cannot be held responsible for not getting a
forecast right.
Plans à managers responsible for making their plans.

q Forecast à The demands of customers are independent demands.

Plan à dependent demand
MPC System
Framework
Providing Appropriate Forecast Information
q Forecast for long term decision are highly aggregated estimates of
general business trends over the long term à too general for SOP.
q The forecast needs for both these applications (strategic
considerations, SOP, and short-term scheduling and execution
decision making) are different.
q The basis of the forecast, investment in forecasting, the nature of
the techniques used, and the frequency is different for each of
these applications.
A Framework for Forecasting
Outline

Forecast Definition

Regression Analysis

Decomposition of a Time Series

Short-Term Forecasting Techniques

Using the Forecasts

Regression Analysis
q A functional relationship between two or more correlated variables.
q Used to predict one variable given the other.
q The relationship is usually developed from observed data.
q The data should be plotted first to see if they appear linear or if at
least parts of the data are linear.
q Linear regression refers to the special class of regression where the
relationship between variables forms a straight line.
Linear Regression Forecasting
q Useful for long-term forecasting of major occurrences and
aggregate planning.
q Used both for time series forecasting and for causal relationship
forecasting.
q Assumption: past data and future projections to fall about a straight
line.
Linear Regression Forecasting (2)

𝑌 = 𝑎 + 𝑏𝑋
𝑌 = 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒
𝑎 = 𝑖𝑛𝑡𝑒𝑟𝑐𝑒𝑝𝑡
𝑏 = 𝑠𝑙𝑜𝑝𝑒
𝑋 = 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒

q When the dependent variable changes as a result of time (plotted as the

horizontal axis) à time series analysis.
q If one variable changes because of the change in another variable, this is a
causal relationship.
Least Squares Method
q The least squares equation for linear regression is Y = a + bx
𝒂 = 𝒚# − 𝒃𝒙#

∑ 𝒙𝒚 − 𝒏𝒙# * 𝒚#
𝒃=
∑ 𝒙𝟐 − 𝒏𝒙# 𝟐

q The line minimizes the sum of the squares of the vertical distance
between each data point and its corresponding point on the line.
𝑴𝒊𝒏𝒊𝒎𝒊𝒛𝒆 = 𝒚𝟏 − 𝒀𝟏 𝟐 + 𝒚𝟐 − 𝒀𝟐 𝟐 + 𝒚𝟑 − 𝒀𝟑 𝟐 + … … … + 𝒚 𝒏 − 𝒀𝒏 𝟐
Example- Least Squares Method

q A firm’s sales for a product line during the 12 quarters of the past three years
were as follows:

q The firm wants to forecast each quarter of the fourth year—that is, quarters
13, 14, 15,and 16.
Example- Least Squares Method(2)
q Fit the line to the data that
minimizes the sum of the
squares of the vertical
distance.
q The standard error of
estimate, or how well the
line fits the data, is

∑𝒏
𝐢"𝟏 𝒚𝒊 (𝒀𝒊
𝟐
𝑺𝒚𝒙 = 𝐧(𝟐
qThe standard error of estimate, or how well the line fits the data, is

#$$%&$'.) !* '++$%''#$., !*⋯* .,$$%./+/.' !

𝑆!" = = 363,9
'0%0
Example- Least Squares Method(4)
q Forecasts for periods 13 through 16 would be Yt = 441.6 + 359.6 t

q Y13 = 441.6 + 359.6 (13) = 5,116.4

q Y14 = 441.6 + 359.6 (14) = 5,476.0
q Y15 = 441.6 + 359.6 (15) = 5,835.6
q Y16 = 441.6 + 359.6 (16) = 6,195.2

q The standard error of estimate, or how well the line fits the data, is

-../0.1.3 ! 4 155./11-..6 ! 4⋯4 86../8959.1 !

𝑆+, = 1:/:
= 363,9
Outline

Forecast Definition

Regression Analysis

Decomposition of a Time Series

Short-Term Forecasting Techniques

Using the Forecasts

Time Series
q Is a chronological sequence of observations on a particular variable.
Is a time-sequenced history of activity. (Fogarty, 1991).
q Chronologically ordered data that may contain one or more
components of demand: trend, seasonal, cyclical, autocorrelation,
and random.
Trend Component
q Refers to the upward or downward movement that characterized a
time series over a period of time.
Seasonal Component
q Periodic patterns in a time series that are repeated on a yearly
basis.
Cycle Component
q Recurring up and down movement around trend levels.
q Refers to long range trends.
Decomposition of a Time Series
q Identifying and separating the time series data into trend, seasonal,
cyclical, autocorrelation, and random components.
q Easy to identify the trend (plot and see the direction of movement)
and the seasonal component (by comparing the same period year
to year).
q Difficult to identify the cycles (these may be many months or years
long), autocorrelation, and random components.
 Types of Seasonal Variation
Additive seasonal variation Multiplicative seasonal variation
q Simply assumes that the seasonal q The trend is multiplied by the
amount is a constant no matter what seasonal factors.
the trend or average amount is. q Forecast including trend and
q Forecast including trend and seasonal = seasonal = trend x seasonal
Trend + Seasonal
Seasonal Factor (or Index)
q A seasonal factor is the amount of correction needed in a time series to
adjust for the season of the year.
q Seasonal usually associate with a period of the year characterized by
some particular activity.
q Cyclical to indicate other than annual recurrent periods of repetitive
activity.
q Determining seasonal indexes:
1. Simple calculation based on past seasonal data
2. The trend and seasonal index from a hand-fit regression line.
Example: Simple Proportion
Assume that in past years, a firm sold an average of 1,000 units of a
particular product line each year. On the average, 200 units were sold
in the spring, 350 in the summer, 300 in the fall, and 150 in the winter.
The seasonal factor (or index) is the ratio of the amount sold during
each season divided by the average for all seasons.
 Average Sales for Seasonal
Past Sales
each season Factor
Spring 200 250 0,8
Summer 350 250 1,4
Fall 300 250 1,2
Winter 150 250 0,6
Total 1000

Using these factors, if we expected demand for next year to be 1,100 units. Forecast the
demand to occur as
Expected Demand Average Sales for Seasonal Next Year Seasonal
for next year each season Factor Forecast
Spring 275 0,8 220
Summer 275 1,4 385
Fall 275 1,2 330
Winter 275 0,6 165
Total 1100
Decomposition Using Least Squares
Regression
q Decomposition of a time series means finding the series’ basic
components of trend, seasonal, and cyclical.
q Indexes are calculated for seasons and cycles.
q The forecasting procedure then reverses the process by projecting
the trend and adjusting it by the seasonal and cyclical indexes,
which were determined in the decomposition process.
Process Decomposition Using Least Squares
Regression
1. Decompose the time series into its components.
a. Find seasonal component.
b. Deseasonalize the demand.
c. Find trend component.

2. Forecast future values of each component.

a. Project trend component into the future.
b. Multiply trend component by seasonal component.
 Step 1. Determine the seasonal factor (or index).
Example:
an average for the same quarters in the
three-year period.
average of the same Seasonal A seasonal factor derived by
Period Sales dividing that average by the general
quarter of each year factor
1 600 2266,67 0,82 average for all 12 quarters
2 1550 3050 1,10
𝑔𝑒𝑛𝑒𝑟𝑎𝑙 𝑎𝑣𝑒𝑟𝑎𝑔𝑒𝑓𝑜𝑟 𝑎𝑙𝑙 12 𝑞𝑢𝑎𝑟𝑡𝑒𝑟 =
3 1500 2700 0,97 33350
4 1500 3100 1,12 12
= 2,779
5 2400 0,82
6 3100 1,10
7 2600 0,97
8 2900 1,12
9 3800 0,82
10 4500 1,10 The seasonal factors are
11 4000 0,97 identical for similar quarters in
12 4900 1,12 each year.
33350
 Step 2. Deseasonalize the original data.

The deseasonalization of demand à remove the seasonal effect on the data

Deseasonalized demand = actual demand ÷ seasonal factor.

average of the same Seasonal

Period Sales Deseasonalized demand
quarter of each year factor
1 600 2266,67 0,82 735,7
2 1550 3050 1,10 1412,4
3 1500 2700 0,97 1544,0
4 1500 3100 1,12 1344,8
5 2400 0,82 2942,6
6 3100 1,10 2824,7
7 2600 0,97 2676,2
8 2900 1,12 2599,9
9 3800 0,82 4659,2
10 4500 1,10 4100,4
11 4000 0,97 4117,3
12 4900 1,12 4392,9
33350
 Step 3. Develop a least squares regression line for the deseasonalized data.

𝑌 = 554.9 + 342.2𝑋
Period Deseasonalized demand
1 735,7 5000,0
2 1412,4 4500,0
3 1544,0 4000,0
4 1344,8 3500,0
5 2942,6 3000,0
6 2824,7 2500,0
7 2676,2 2000,0
8 2599,9 1500,0
9 4659,2 1000,0 Deseasonalized demand
10 4100,4 500,0
Linear (Deseasonalized demand)
11 4117,3 0,0
12 4392,9 0 2 4 6 8 10 12 14
 Step 4. Project the regression line through the period to be forecast.

Step 5. Create the final forecast by adjusting the regression line by the seasonalfactor.

q 𝑌 = 554.9 + 342.2𝑋

Y from Seasonal
Period Quarter regression factor Forecast
line
13 1 5003,5 0,82 4080,816
14 2 5345,7 1,10 5866,645
15 3 5687,9 0,97 5525,876
16 4 6030,1 1,12 6726,228
Error Range
q When a straight line is fitted
through data points and then
used for forecasting, errors can
come from two sources:
q The usual errors similar to the
standard deviation of any set of
data.
q Errors that arise because the line is
wrong.
Outline

Forecast Definition

Regression Analysis

Decomposition of a Time Series

Short-Term Forecasting Techniques

Using the Forecasts

Short-term Forecasting Techniques
q Commonly available in commercial software and meet the criteria
of low cost and little management involvement:
1. Moving averages
2. exponential smoothing
q The techniques are simple mathematical means for converting past
information into forecasts.
q Simple methods do better than the more sophisticated models for
detailed forecasts, especially over short periods.
Moving-Average Forecasting
q Averaging past demand to project a forecast for future demand.
q The moving average procedure uses only a few of the most recent
demand observations, because:
q There may be so many periods of past data that storing them all is an issue.
q Often the most recent history is most relevant in forecasting short-term
demand in the near future.
Moving-Average Forecasting(2)
q Moving average forecast (MAF) at the end of period t:
∑86789:;< #$!%&' ()*&+(
q 𝑀𝐴𝐹! = +
i = period number
t = current period (the period for which the most recent actual demand is
known)
n = number of periods in the moving average

q We are at the end of period t; we know the demand in period t and

forecasts are made for periods t + 1, or t + X periods into the future.
q Forecasts are not made for period t since that period’s demand is known.
 If there were requests for shipments that couldn’t be fulfilled à
Example: capture that information to get better estimates of the actual demand.

∑)0
)0%#*' 𝐴𝑐𝑡𝑢𝑎𝑙 𝑑𝑒𝑚𝑎𝑛𝑑
6 − 𝑝𝑒𝑟𝑖𝑜𝑑𝑒 𝑀𝐴𝐹 𝑚𝑎𝑑𝑒 𝑎𝑡 𝑡ℎ𝑒 𝑒𝑛𝑑 𝑜𝑓 𝑝𝑒𝑟𝑖𝑜𝑑 32 =
6
900 + 1100 + 1500 + 1400 + 1700 + 1200
= = 1300
6
∑)0
)0%)*' 𝐴𝑐𝑡𝑢𝑎𝑙 𝑑𝑒𝑚𝑎𝑛𝑑
3 − 𝑝𝑒𝑟𝑖𝑜𝑑𝑒 𝑀𝐴𝐹 𝑚𝑎𝑑𝑒 𝑎𝑡 𝑡ℎ𝑒 𝑒𝑛𝑑 𝑜𝑓 𝑝𝑒𝑟𝑖𝑜𝑑 32 =
3
1400 + 1700 + 1200
= = 1433
3
q The moving-average model smooth the historical data, with an equal weight on each piece of
historical information.
q Adjusting by incorporating different weights on past periods à the weighted-moving-average
model
Exponential Smoothing Forecasting
q Adjusts the weights given to past data that older data get increasingly
less weight.
q The basic: each new forecast is based on an average that’s adjusted each
time there’s a new forecast error.
q For example:
forecast 90 units and actual demand turns out to be 100 units à increase
the forecast by some portion of the 10-unit error in making the next
period’s forecast.
q If the error indicated demand was changing, we would begin to change
the forecast.
Exponential Smoothing Forecasting (2)
q The proportion of the error that will be incorporated into the forecast à
the exponential smoothing constant (∝).

𝐸𝑆𝐹! = 𝐸𝑆𝐹!,- + ∝ 𝑎𝑐𝑡𝑢𝑎𝑙 𝑑𝑒𝑚𝑎𝑛𝑑! − 𝐸𝑆𝐹!,-

= ∝ 𝑎𝑐𝑡𝑢𝑎𝑙 𝑑𝑒𝑚𝑎𝑛𝑑! + 1− ∝ 𝐸𝑆𝐹!,-

∝ = the smoothing constant (0 - 1)

t = current period (the period for which the most recent actual demand
is known)
𝐸𝑆𝐹@AB = exponential smoothing forecast made one period previously (at the
end of period t − 1)
 𝑬𝑺𝑭𝒕 = 𝑬𝑺𝑭𝒕#𝟏 + ∝ 𝒂𝒄𝒕𝒖𝒂𝒍 𝒅𝒆𝒎𝒂𝒏𝒅𝒕 − 𝑬𝑺𝑭𝒕#𝟏
Example: = ∝ 𝒂𝒄𝒕𝒖𝒂𝒍 𝒅𝒆𝒎𝒂𝒏𝒅𝒕 + 𝟏− ∝ 𝑬𝑺𝑭𝒕#𝟏

Assume:
q ESF26 = exponential smoothing forecast made at the end of period 26 = 1,000,
∝ = 0.1
q ESF27 (made at the end of period 27 when actual demand for period 27 is
known but actual demand in period 28 is not known).
𝐸𝑆𝐹%& = 1000 + 0.1 900 − 1000 = 990
= 0.1 900 + 1 − 0.1 1000 = 990

𝐸𝑆𝐹%' = 0.1 1100 + 1 − 0.1 990 = 1001

Exponential Smoothing Forecasting (4)
q The larger values of ∝ give more weight to recent demands and utilize older
demand data less than is the case for smaller values of ∝.
q The larger values of ∝ provide more responsive forecasts, and smaller values
produce more stable forecasts.
Comparing exponential smoothing and
moving average forecasting procedures
q Trade-off in determining what smoothing constant (or length of
moving average) to use in a forecasting procedure.
q The higher the smoothing constant (∝) or the shorter the moving
average (n), the more responsive forecasts are to changes in
underlying demand, but the more “nervous” they are in the
presence of randomness.
q Similarly, smaller smoothing constants (∝) or longer moving
averages (n) provide stability in the face of randomness but slower
reactions to changes in the underlying demand.
Evaluating Forecasts
q The quality of any forecast is reflected in the quality of the decisions
based on the forecast.
q The ideal comparison of forecasting procedures would be based on
the costs of producing the forecast and the value of the forecast for
the decision à not easy to measure.
q Criterion:
q Honesty, or lack of bias à the procedure should produce forecasts that are
neither consistently high nor consistently low.
q Forecasts shouldn’t be overly optimistic or pessimistic
Mean Error
q To measure bias.
∑:
67< &$!%&' ()*&+(6 ,./0)$&1! ()*&+(6
q M𝑒𝑎𝑛 𝑒𝑟𝑟𝑜𝑟 𝑏𝑖𝑎𝑠 = +
i = period number
n = number of periods of data
Example Bias Calculation:

q When forecast errors tend to cancel one another out, the measure of bias
tends to be low.
q Positive errors in some periods are offset by negative errors in others, which
tends to produce an average error, or bias near zero.
Mean Absolute Deviation (MAD)
q Used measure of forecast error magnitude.

q Expresses the size of the average error irrespective of whether it’s

positive or negative.
q It’s the combination of bias and MAD that allows us to evaluate
forecasting results.

∑+23- 𝑎𝑐𝑡𝑢𝑎𝑙 𝑑𝑒𝑚𝑎𝑛𝑑2 − 𝑓𝑜𝑟𝑒𝑐𝑎𝑠𝑡 𝑑𝑒𝑚𝑎𝑛𝑑2

𝑚𝑒𝑎𝑛 𝑒𝑟𝑟𝑜𝑟 𝑏𝑖𝑎𝑠 =
𝑛
Example MAD Calculation 1:
Example MAD Calculation 2:
Standard Deviation
q When the forecast errors are distributed normally à direct relationship
between MAD and Standard deviation.

q Standard deviation of forecast errors = 1.25 MAD

Outline

Forecast Definition

Regression Analysis

Decomposition of a Time Series

Short-Term Forecasting Techniques

Using the Forecasts

Considerations for Aggregating Forecasts
q There are several reasons for aggregating product items in both time and
level of detail for forecasting purposes.

q Individual products into product lines, geographical areas, or customer

types, for example, must be done in ways that are compatible with the
planning systems.

q Product groupings must also be developed, so that the forecast unit is

sensible to forecasters.
Considerations for Aggregating Forecasts(2)
q A well-known phenomenon that long-term or product-line forecasts are more
accurate than short-term and/or detailed forecasts.
q This merely verbalizes a statistical verity.

The reduction from 20% to 5.8% is due to using a much longer time period.
The same effect can be seen in forecasting demand for product lines instead of for individual items.
Considerations for Aggregating Forecasts(3)
q The total forecast must be consistent with the individual product
forecasts. The whole must be equal to the sum of the parts.
q Very often an individual product’s share of the aggregate product
line totals remains fairly constant.
q That is, there is more uncertainty in the day-to-day demand for the
item than for its share in the demand for the total line.
q We can use this knowledge to disaggregate the aggregate forecasts
and thereby maintain the consistency between the detail and the
totals.
Pyramid Forecasting
q Provides a means of coordinating, integrating, and
assuring consistency between the various sources of
forecasts and any company constraints or goals.
q Procedures:
1. Begins with individual product item forecasts (L3)
2. Rolled up into forecasts for product lines (L2)
3. Aggregate forecasts for product lines into a total
business forecast (in dollars) at level 1
4. Force down (constrain) the product line and
individual item forecasts, so they’re consistent
with the plan
Example: Pyramid Forecasting: Rolled up
Example Pyramid Forecasting: Force Down
q The results are consistent forecasts
throughout the organization, and the
sum of the parts is forced to equal the
whole.
q the process of forcing the consistency
needs to be approached with caution
Incorporating External Information
q Activities that will influence demand and perhaps invalidate the use of a
routine exponential forecasting model à change the forecast directly.
q If technology has prolonged the life of products, need to change the
parameters in the model that relates replacement sales to the average
life of our products in the field.
q There may be circumstances where our knowledge would lead us to
change both the forecast and the forecasting parameters.
q Intelligence must be included in the forecasts, and the forecasts must be
readied for use in preparing and controlling the plans of the firm.
Terima Kasih
Yani Herawati
Program Studi Teknik Industri
UNPAR
[email protected]