Course : Production & Operation Analysis

Effective Period : September 2015

Forecasting

Session 1
D1316 – Production and
Operations Analysis
What is Forecasting

• Forecasting is the art and science of predicting future

events
• Forecasting may involve taking historical data and
projecting them into the future with some sort of
mathematical model
The Time Horizon in Forecasting
Characteristics of Forecast

• They are usually wrong

• A good forecast is more than a single number
• Aggregate forecasts are more accurate
• The longer the forecast horizon, the less accurate the
forecast will be
• Forecasts should not be used to the exclusion of
known information
7 Steps in The Forecasting System

• Determine the use of the forecast

• Select the items to be forecasted
• Determine the time horizon of the forecast
• Select the forecasting models
• Gather the data needed to make the forecast
• Make the forecast
• Validate and implement the results
Forecasting Approaches

• A subjective/qualitative forecasting method is based

on human judgment
• Objective/quantitative forecasting methods are
those in which the forecast is derived from an
analysis of the data
Subjective Forecasting Methods

• Sales force composites

a forecasting technique based on salespersons’ estimates of expected
sales
• Customer surveys
a forecasting method that solicits input from customers or potential
customers regarding future purchasing plans
• Jury of executive opinion
a forecasting technique that takes the opinion of a small group of high
level managers and results in a group estimate of demand
• Delphi method
a forecasting technique using a group process that allows experts to
make forecasts
Objective Forecasting Methods

Causal Models are ones that use data from sources other
than the series being predicted; that is, there may be
other variables with values that are linked in some way
to what is being forecasted

Let Y be the quantity to be forecasted and (X 1, X2, . . . , Xn)

be n variables that have predictive power for Y. A causal
model is Y = f (X1, X2, . . . , Xn)
A typical relationship is a linear one:
Y = α0 + α1X1 + α2X2 +. . . + αnXn
Time Series Methods
• Time series methods are often called naïve methods, as they require
no information other than the past values of the variable being
predicted
• In time series analysis we attempt to isolate the patterns that arise
most often
Decomposition of A Time Series
• Trend
Trend refers to the tendency of a time series to exhibit a stable pattern of growth
or decline
• Seasonality
A seasonal pattern is one that repeats at fixed intervals
• Cycles
Cyclic variation is similar to seasonality, except that the length and the magnitude
of the cycle may vary
• Randomness
A pure random series is one in which there is no recognizable pattern to the data
Evaluation of Forecasts

• The forecast error in period t, et, as the difference between

the forecast value for that period and the actual demand for
that period.
• For multiple step ahead forecast: et = Ft - t, t - Dt
• For one step ahead forecast: et = Ft – Dt
• Let e1, e2, …, en be the forecast errors observed over n periods.
Two common measures of forecast accuracy during these n
periods are the mean absolute deviation (MAD) and the mean
squared error (MSE), given by the following formulas:
n n
MAD  (1 / n) ei MSE  (1 / n) ei2
i 1 i 1
• Note that the MSE is similar to the variance of a
random sample
• The MAD is often the preferred method of
measuring the forecast error because it does not
require squaring
• Although the MAD and the MSE are the 2 most
common measures of forecast accuracy, other
measures are used as well
• One that is not dependent on the magnitude of the
values of demand is known as the mean absolute
percentage error (MAPE) and is given by the formula:
n
MAPE  [(1 / n) ei / Di ]x100
i 1
Artel, a manufacturer of static random access memories
(SRAM), has production plants in Austin, Texas,
Sacramento, California. The managers of these plants are
asked to forecast production yields (measured in %) one
week ahead for their plants. Based on 6 weekly forecasts,
the firm’ management wishes to determine which
manager is more successful at predicting his plant’ yields.
The results of their predictions are given in the following
table:
Week P1 O1 |E1| |E1/O1| P2 O2 |E2| |E2/O2|
1 92 88 4 0.0455 96 91 5 0.0549
2 87 88 1 0.0114 89 89 0 0.0000
3 95 97 2 0.0206 92 90 2 0.0222
4 90 83 7 0.0843 93 90 3 0.0333
5 88 91 3 0.0330 90 86 4 0.0465
6 93 93 0 0.0000 85 89 4 0.0449
• Interpret P1 as the forecast made by the manager of
plant 1 at the beginning of each week, O1 as the yield
observed at the end of each week in plant 1, and E1
as the difference between the predicted and the
observed yields
• The same definitions apply to plant 2
• Let us compare the performance of these managers
using the three measures MAD, MSE and MAPE as
defined previously. To compute the MAD we simply
average the observed absolute errors:
MAD1 = 17/6 = 2.83
MAD2 = 18/6 = 3
• Based on the MAD, the first manager has a slight
edge. To compute the MSE in each case, square the
observed errors and average the results to obtain:
MSE1 = 79/6 = 13.17
MSE2 = 70/6 = 11.67
• The second manager’ forecasts have a lower MSE
than the first, even though the MAD go the other
way.
• Let us now compare their performances based on
the MAPE. To compute the MAPE we average the
ratios of the errors and the observed yields:
MAPE1 = 0.0325
MAPE2 = 0.0336
• Using the MAD or the MAPE, the first manager has a
very slight edge, but using the MSE, the second
manager looks better
• Who is declared the “winner” depends on which
method of evaluation management chooses
Methods for Forecasting Stationary Series

• A stationary time series is one in which each

observation can be represented by a constant plus a
random fluctuation. In symbols,
Dt = µ + εt
where µ is an unknown constant corresponding to
the mean of the series and εt is a random error with
mean zero and variance σ2
Moving Averages

• A moving average of order N is simply the arithmetic

average of the most recent N observations
• For the time being we restrict attention to one step
ahead forecast. Then Ft, the forecast made in period
t-1 for period t, is given by:
t 1
Ft  (1 / N )  Di  (1 / N )( Dt  1  Dt  2  ....  Dt  N )
i t  N
 Moving Averages (cont’d)
Actual
Period Value Three-Month Moving Averages
Januari 10
Februari 12
Maret 16
April 13 10 + 12 + 16 / 3 = 12.67
Mei 17 12 + 16 + 13 / 3 = 13.67
Juni 19 16 + 13 + 17 / 3 = 15.33
Juli 15 13 + 17 + 19 / 3 = 16.33
Agustus 20 17 + 19 + 15 / 3 = 17.00
September 22 19 + 15 + 20 / 3 = 18.00
Oktober 19 15 + 20 + 22 / 3 = 19.00
November 21 20 + 22 + 19 / 3 = 20.33
Desember 19 22 + 19 + 21 / 3 = 20.67
 Moving Averages (cont’d)
Forecast Error Analysis
Absolute Squared Absolute
Period Actual Value Forecast Error error error % error
Januari 10
Februari 12
Maret 16
April 13 12.667 -0.333 0.333 0.111 2.56%
Mei 17 13.667 -3.333 3.333 11.111 19.61%
Juni 19 15.333 -3.667 3.667 13.444 19.30%
Juli 15 16.333 1.333 1.333 1.778 8.89%
Agustus 20 17.000 -3.000 3.000 9.000 15.00%
September 22 18.000 -4.000 4.000 16.000 18.18%
Oktober 19 19.000 0.000 0.000 0.000 0.00%
November 21 20.333 -0.667 0.667 0.444 3.17%
Desember 19 20.667 1.667 1.667 2.778 8.77%
Average -1.333 2.000 6.074 10.61%
Next period 19.667 MAD MSE MAPE
 Moving Averages (cont’d)
3 period moving averages
25

20

15
Actual
Value

Value
Forecast
10

0
1 2 3 4 5 6 7 8 9 10 11 12
Time
 Weighted Moving Average
Three-Month Weighted Moving Averages

Actual
Period Value Weights
Januari 10 0.222
Februari 12 0.593
Maret 16 0.185
April 13 2.2 + 7.1 + 3 / 1 = 12.298
Mei 17 2.7 + 9.5 + 2.4 / 1 = 14.556
Juni 19 3.5 + 7.7 + 3.2 / 1 = 14.407
Juli 15 2.9 + 10 + 3.5 / 1 = 16.484
Agustus 20 3.8 + 11 + 2.8 / 1 = 17.814
September 22 4.2 + 8.9 + 3.7 / 1 = 16.815
Oktober 19 3.3 + 12 + 4.1 / 1 = 19.262
November 21 4.4 + 13 + 3.5 / 1 = 21.000
Desember 19 4.9 + 11 + 3.9 / 1 = 20.036

Next period 20.185

Sum of weights = 1.000

 Weighted Moving Average (cont’d)
Forecast Error Analysis
Absolute Squared Absolute
Period Actual value Weights Forecast Error error error % error
Januari 10 0.222
Februari 12 0.593
Maret 16 0.185
April 13 12.298 -0.702 0.702 0.492 5.40%
Mei 17 14.556 -2.444 2.444 5.971 14.37%
Juni 19 14.407 -4.593 4.593 21.093 24.17%
Juli 15 16.484 1.484 1.484 2.202 9.89%
Agustus 20 17.814 -2.186 2.186 4.776 10.93%
September 22 16.815 -5.185 5.185 26.889 23.57%
Oktober 19 19.262 0.262 0.262 0.069 1.38%
November 21 21.000 0.000 0.000 0.000 0.00%
Desember 19 20.036 1.036 1.036 1.074 5.45%
Average -1.370 1.988 6.952 10.57%
Next period 20.185 MAD MSE MAPE

Sum of weights = 1.000

Exponential Smoothing

• The current forecast is the weighted average of the

last forecast and the current value of demand

where 0 < α ≤ 1 is the smoothing constant, which

determines the relative weight placed on the current
observation of demand. Interpret (1-α) as the weight
placed on past observations of demand
Ft  D t -1  (1   ) Ft 1
• Quarterly data for the failures of certain aircraft
engines at a local military base during the last 2
years are 200, 250, 175, 186, 225, 285, 305, 190. We
will now forecast using exponential smoothing. In
order to get the method started, let us assume that
the forecast for period 1 was 200. Suppose that
α=0.1. The one step ahead forecast for period 2 is
F2 = αD1 + (1- α)F1 = (0.1)(200) + (0.9)(200) = 200
similarly, F3 = αD2 + (1- α)F2 = (0.1)(250) + (0.9)(200) =
205
• Other one step ahead forecasts are computed in the
same fashion. The observed numbers of failures and
the one step ahead forecasts for each quarter are the
following:
Quarter Failures Forecast
1 200 200 (by
assumption)
2 250 200
3 175 205
4 186 202
5 225 201
6 285 203
7 305 211
8 190 220
Trend Based Methods

• Both exponential smoothing and moving average

forecasts will lag behind a trend if one exists
• We will consider 2 forecasting methods that
specifically account for a trend in the data:
regression analysis and holt’s method
 Regression Analysis
• Is a method that fits a straight line to a set of data
• Let (x1, y1), (x2,y2),…,(xn,yn) be n paired data points for
the 2 variables X and Y. Assume that yi is the
observed value of Y when xi is the observed value of
X. Refer to Y as the dependent variable and X as the
independent variable. We believe that a relationship
exists between X and Y that can be represented by
the straight line Y = a + bX
  n 2  (n  1)  (2n  1    n 2   n  1 
2    n  1 
  a  D b 
S xx  
 6   

4 
 2 
   
n
  n * (n  1)   n
S xy
S xy  n iDi   2  *  Di b
i 1  i 1 S xx
 Month # Visitors Month # Visitors
Jan 133 Apr 640
Feb 183 May 1876
Mar 285 Jun 2550
 Sxy=6*(1*133+2*183+3*285+4*640+5*1876+6*2550)
– (6*7/2)(133+183+285+640+1876+2550)] = 52557
 Sxx = [(36*7*13)/6]-[(36*49)/4)]= 105
 b = (52557/105)= 500.54
 a = 944.5 – 500.54*(6+1)/2 = -807.4
Holt’s Method

• Is a type of double exponential smoothing designed to track time

series with linear trend
• The method requires the specification of 2 smoothing constants, α
and β, and uses 2 smoothing equations: one for the value of the
series (the intercept) and one for the trend (the slope)
• St = Dt + (1-)(St-1 + Gt-1)
• Gt = (St – St-1) + (1- )Gt-1
– Dt is observation of demand; St is the intercept at time t;
Gt is the slope at time t; St-1 is the previous intercept; Gt-1 is the previous
slope
• Ft,t+ = St + Gt (the  step ahead forecast made in period t)
• Let us apply Holt’s method to the problem of
developing one step ahead forecasts for the aircraft
engine failure data. Recall that the original series was
200, 250, 175, 186, 225, 285, 305, 190. Assume that
both α and β are equal to 0.1. In order to get the
method started, we need estimates of both the
intercept and the slope at time zero. Suppose that
these are S0 = 200 and G0 = 10. Then we obtain
S1 = (0.1)(200) + (0.9)(200+10) = 209
G1 = (0.1)(209-200) + (0.9)(10) = 9.9
S2 = (0.1)(250) + (0.9)(209+9.9) = 222
G2 = (0.1)(222-209) + (0.9)(9.9) = 10.2
S3 = (0.1)(175) + (0.9)(222+10.2) = 226.5
G3 = (0.1)(226.5-222) + (0.9)(10.2) = 9.6 and so on
• Comparing the one step ahead forecasts to the
actual numbers of failures for periods 4 through 8,
we obtain the following:
Period Actual Forecast I Error I
4 186 236.1 50.1
5 225 240.3 15.3
6 285 247.7 37.3
7 305 260.8 44.2
8 190 275 85
• Suppose you needed to forecast the demand in
period 2 for period 5. This forecast is F2,5 = S2 + (3)G2 =
222 + (3)(10.2) = 252.6
Methods for Seasonal Series

• A seasonal series is one that has a pattern that

repeats every N periods for some value of N (which is
at least 3)
• There are several ways to represent seasonality. The
most common is to assume that there exists a set of
multipliers ct, for 1≤t ≤N, with the property that Σct =
N
• The multiplier ct represents the average amount that
the demand in the tth period of the season is above
or below the overall average. These multiplier are
known as seasonal factors.
Seasonal Factors for Stationary Series

• In this part of the section we present a simple

method of computing seasonal factors for a time
series with seasonal variation and no trend
• This method is as follows:
1. Compute the sample mean of all the data
2. Divide each observation by the sample mean
3. Average the factors for like periods within each
season
• Ex: The county transportation department needs to
determine usage factors by day of the week for a
popular toll bridge connecting different parts of the
city. In the current study, they are considering only
working days. Suppose that the numbers of cars
using the bridge each working day over the past four
weeks were (in thousands of cars)
 Week 1 Week 2 Week 3 Week 4
Monday 16.2 17.3 14.6 16.1
Tuesday 12.2 11.5 13.1 11.8
Wednesday 14.2 15 13 12.9
Thursday 17.3 17.6 16.9 16.6
Friday 22.5 23.5 21.9 24.3

• Find the seasonal factors corresponding to daily

usage of the bridge
1. First we compute the average of all the observations.
You should satisfy yourself that the mean is 16.425
2. Next, we divide each observation by the mean value.
Doing so gives the following:
0.986 1.053 0.889 0.980
0.743 0.700 0.798 0.718
0.865 0.913 0.791 0.785
1.053 1.072 1.029 1.011
1.370 1.431 1.333 1.479
3. Finally, we average factors corresponding to the
same period of the season. That is, average all
factors for Mondays, all factors for Tuesdays and so
on. The resulting 5 seasonal factors for bridge usage
by day of the week are
Monday 0.98

Tuesday 0.74

Wednesday 0.84

Thursday 1.04

Friday 1.4
• You should satisfy yourself that these factors do sum
to 5 as required
• Interpret the results in the following way: Mondays
are an average day because the factor for Mondays
is very close to 1
• The usage is lowest on Tuesdays, about 26% below
the average, and highest on Fridays, about 40%
above the average
• Forecasts for the numbers of cars using the bridge
on any day of the week would be obtained by
multiplying the sample mean of 16.425 by the
appropriate seasonal factor. Hence, usage forecasts
by day of the week (in thousands of cars) would be
Monday 16.1

Tuesday 12.1

Wednesday 13.8

Thursday 17.1

Friday 23
Seasonal Decomposition Using Moving Averages

• A slightly more complex method for estimating

seasonal factors requires the computation of N
period moving averages, where N is the length of the
season
 • Ex: Suppose that the original demand history of
a certain item for the past 8 quarters is given by
10,20,26,17,12,23,30,22. The graph of this
demand series is given below.

35

30

25

20
Demand

15

10

0
1 2 3 4 5 6 7 8
Time

The picture suggests that these data represent 2 seasons, with each season
being 4 periods long. The next step is to compute all 4 period moving
averages. This gives:
 Period Demand MA (4)
1 10
2 20
3 26
4 17 18.25
5 12 18.75
6 23 19.5
7 30 20.5
8 22 21.75

• Next, the moving averages must be centered.

• The moving average 18.25 is computed by averaging the
demand for periods 1,2,3 and 4. The center of these
numbers is 2.5.
• The next average of 18.75 corresponds to periods 2,3,4
and 5. The center of these numbers is 3.5.
• We repeat this process for the remaining averages and
locate them in the centered positions in the following way:
Period Demand MA (4) Centered
1 10
2 20
18.25
3 26
18.75
4 17 18.25
19.5
5 12 18.75
20.5
6 23 19.5
21.75
7 30 20.5
8 22 21.75
• Next, we must get these centered values back on periods. To do so,
we average adjacent values.
• Averaging 18.25 and 18.75 gives 18.5, which corresponds to the
average of periods 2.5 and 3.5, which is 3. Repeating this process for
the other centered values gives centered moving averages for
periods 4,5 and 6 as well.
• The next step is to obtain values for periods 1,2,7 and 8. In order to
obtain values for periods 1 and 2, we average the values for periods 3
and 4, and to obtain values for periods 7 and 8, we average the
values for periods 5 and 6. In this way, we obtain a centered moving
average for each period in which we have an observation of demand.
 (A) (B) (C) (B/C)
Period Demand Centered MA Ratio

1 10 18.81 0.532
2 20 18.81 1.063
3 26 18.5 1.405
4 17 19.125 0.888
5 12 20 0.6
6 23 21.125 1.089
7 30 20.56 1.463
8 22 20.56 1.07

Once we have obtained the centered moving average for each

period, we form the ratio of the demand for that period over the
centered MA. These values, reported in the fourth column
above, are estimates of the seasonal factors for each period.
• The next step is to form the average of the factors that
correspond to the same periods of each season.
• As we have exactly 2 seasons of data, we average
(0.532+0.6)/2=0.566, (1.063+1.089)/2=1.076,
(1.405+1.463)/2=1.434, (0.888+1.07)/2=0.979
• In this case, this procedure yields exactly 4 seasonal factors.
In general, it will result in exactly N seasonal factors.
• Finally, we must be sure that the sum of the seasonal
factors is exactly N, or exactly 4 for this example
• We find 0.566 + 1.076 + 1.434 + 0.979 = 4.055
• In order to make the sum equal to exactly 4, we
multiply each factor by 4/4.055 = 0.9864
• The final seasonal factors are

Period Factor
1 0.558
2 1.061
3 1.415
4 0.966
• The sum of these final seasonal factors is exactly 4.
Based on the given data, these factors tell us that, on
average, the first quarter of each year results in sales
that are about 45% below the yearly average, the
second quarter in sales that are about 6% above the
yearly average and so on.
• The next step in the process is to divide each
observation by the appropriate seasonal factor in
order to obtain the deseasonalized demand
 (A) (B) Deseasonalized
Demand Factor Demand (A/B)
10 0.558 17.92
20 1.061 18.85
26 1.415 18.39
17 0.966 17.6
12 0.558 21.5
23 1.061 21.68
30 1.415 21.22
22 0.966 22.77

A forecast can now be made based on the deseasonalized demand, which must be
reseasonalized by multiplying by the appropriate seasonal factor.
As an example, suppose that we were using a 6 month moving average to forecast
the deseasonalized series. The average of the last 6 observations is 20.52.
This number is then multiplied by the appropriate seasonal factor to obtain
forecasts of future demand. The forecast for period 9 is (20.52)(0.558) = 11.45, for
period 10 is (20.52)(1.061) = 21.77 and so on.
• However, because the deseasonalized series exhibits a trend, it
would be more appropriate to apply a trend based method such as
holt’s method or regression analysis to the deseasonalized series
• If we use regression analysis, we obtain the least squares fit of the
data as Dt=16.8+0.7092t
• Suppose that we are interested in forecasting at the end of period 8
for periods 9 through 12. Substituting t equals 9 through 12 into the
regression equation gives the forecasts of the deseasonalized
series for the following year as 23.18, 23.89, 24.6 and 25.31
• Each of these forecasts is then reseasonalized by multiplying by the
appropriate seasonal factor
• The forecasts obtained by this method for the following year
(periods 9 through 12) are respectively 12.93, 25.35, 34.81 and
24.45
Winter’s Method for Seasonal Problems

• Is a type of triple exponential smoothing and this has

the important advantage of being easy to update as
new data become available
• We assume a model of the form
Dt = (µ + Gt)ct + εt
• Interpret µ as the base signal or intercept at time t =
0 excluding seasonality, Gt as the trend or slope
component, ct as the multiplicative seasonal
component in period t and εt as the error term
• The series: St    Dt / Ct  N   (1   )( St  1  Gt  1)

• The trend: Gt   [ St  St  1]  (1   )Gt  1

• The seasonal ct   ( Dt / St )  (1   )ct  N

factors:
The forecast made in period t for any future period t+ is given
by Ft,t+  = (St + Gt)c t+ -N
Note that this forecasting equation assumes that t≤N
If N<τ≤2N, the appropriate seasonal factor would be ct+τ-2N; if
2N<τ≤3N, the appropriate seasonal factor would be ct+τ-3N; and
so on
• Calculate the sample N
1
means for the 2 separate V1 
N
D j
seasons of data j  2 N 1

0
1
V2 
N
D
j   N 1
j

• Initial slope estimate G 0  (V 2  V 1) / N

 S 0  V 2  G 0[( N  1) / 2]
• The value of the
series at time t =
0 Dt
ct 
Vi  [( N  1) / 2  j ]G 0
• Initial seasonal
factors for eachWhere i = 1 for the first season, i = 2 for the second season and j is the
period period of the season. That is, j = 1 for t = -2N+1 and t = -N+1; j = 2 for t =
-2N+2 and t = -N+2 and so on
• Average the
seasonal factors
c  2N  1  c  N  1 c  N  c0
c  N 1  ,....., c 0 
2 2

• Normalize the  
seasonal factors  cj 
cj    N 1  N
 ci 
 
i 0

Period Obs. Period Obs.
2001 Demand 2002 Demand
1st 10 1st 12

2nd 20 2nd 23

3rd 26 3rd 30

4th 17 4th 22
• V1 = (10+20+26+17)/4 = 18.25
• V2 = (12+23+30+22)/4 = 21.75
• G0 = (21.75-18.25)/4 = 0.875
• S0 = 21.75+(0.875)(1.5) = 23.06
• The initial seasonal factors are computed as follows:

10
C7   0.5904
18.25  (5 / 2  1)( 0.875)

20
C6  1.123
18.25  (5 / 2  2)(0.875)
• The other factors are computed in a similar fashion. They
are
c-5 = 1.391 c-2 = 1.079
c-4 = 0.869 c-1 = 1.352
c-3 = 0.5872 c0 = 0.9539
• We then average C-7 and C-3, C-6 and C-2, and so on, to
obtain the four seasonal factors
c-3 = 0.5888, c-2 = 1.1010, c-1 = 1.3720, c0 = 0.9115
• Finally norming the factors to ensure that the sum is 4 results in
c-3 = 0.5900, c-2 = 1.1100, c-1 = 1.3800, c0 = 0.9200
• Suppose that we wish to forecast the following year’s demand at
time t=0. The forecasting equation is
Ft,t+  = (St + Gt)c t+ -N which results in
F0,1 = (S0+G0)c-3 = (23.06+0.875)(0.59)=14.12
F0,2 = (S0+2G0)c-2 = [23.06+(2)(0.875)](1.11)=27.54
F0,3 = 35.44
F0,4 = 24.38
Now, suppose at that time t=1 we observe a demand of
D1=16
We now need to update our equations. Assume that
α=0.2, β=0.1, γ=0.1
St    Dt / Ct  N   (1   )( St  1  Gt  1)
S1 = 0.2 (D1/c-3)+(1-0.2)(S0+G0)
= 0.2 (16/0.59) + (0.8)(23.06+0.875) = 24.57

Gt   [ St  St  1]  (1   )Gt  1

G1 = 0.1 (S1-S0) + (1-0.1)G0

= 0.1 (24.57-23.06) + 0.9(0.875) = 0.9385

ct   ( Dt / St )  (1   )ct  N
C1 = 0.1(D1/S1)+(1-0.1)C-3
= 0.1(16/24.57)+(0.9)(0.59) = 0.5961
Key Point

Forecast must be measured for accuracy!

The most common means of doing so is by measuring

the either the mean absolute deviation or the
standard deviation of the forecast error
 Measuring Accuracy: Tracking signal
• A tracking signal is a measurement of how well a
forecast is predicting actual values. It is used to
decide when to re-evaluate using a model.
n RSFE
RSFE   ( At  Ft ) TS 
i 1
MAD
Positive tracking signals: most of the time actual values are above our
forecasted values
Negative tracking signal: most of the time actual values are below our
forecasted values
If TS > 4 or < -4, investigate!
 Carlson’ bakery wants to evaluate performance of its croissant forecast

Quarter Actual Forecas Error RSFE |Abs Cum MAD TS

demand t forecas abs
demand t error| forecast
error

1 90 100 -10 -10 10 10 10 -1

2 95 100 -5 -15 5 15 7.5 -2
3 115 100 +15 0 15 30 10 0
4 100 110 -10 -10 10 40 10 -1
5 125 110 +15 +5 15 55 11 +0.5
6 140 110 +30 +35 30 85 14.2 +2.5

MAD = 85/6 = 14.2

TS = 35/14.2 = 2.5
Reference

• Steven Nahmias. (2009). Production and Operations

Analysis. 06. MGH. New York. ISBN: 978-007-126370-
2
D1316
Production and Operations Analysis