Chapter 13 Forecasting
It is difficult to forecast, especially in regards to the future.
MGS3100 Julie Liggett De Jong
It isn’t difficult to forecast, just to forecast correctly.
Economic Forecasts Influence:
Numbers, if tortured enough, will confess to just about anything.
Government policies & business decisions
1
Insurance companies’ investment decisions in mortgages and bonds
Service industries’ forecasts of demand as input for revenue management
Excel Features & Functions FEATURES • Regression • Solver • Sorting FUNCTIONS • SUMPRODUCT( ) • SUMXMY2( ) • YEAR( ) • MONTH( ) • RIGHT( )
Quantitative Forecasting Models
Quantitative vs Qualitative Forecasting Models
Expressed in mathematical notation
2
Quantitative Forecasting Models 1. Causal (Curve Fitting) a. Linear b. Quadratic
2. Moving Averages (Naive) a. Simple n-Period Moving Average b. Weighted n-Period Moving Average
3. Exponential Smoothing a. Basic model b. Holt’s Model (exponential smoothing with trend)
Based on an amazing quantity of data 4. Seasonality
Important Variables
X
Independent variable(s)
Y
True value of dependent variable
Yˆ
Predicted or forecasted dependent variable (Y hat)
Y
Average value of dependent variable (Y bar)
Causal vs Time Series Models
Causal Forecasting Models
Requirements
3
Independent and dependent variables must share a relationship
Curve Fitting
Oil company wants to expand its network of self-service gas stations
We must know the values of the independent variables when we make the forecast
Self Service Gas Stations
We’ll use historical data for five stations to calculate average traffic flow and sales
4
Sales & Traffic Data
Plot the averages in a scatter plot.
Traffic flow: average # of cars / hour Sales: average dollar sales / hour
Figure 1, p274
Method of Least Squares
300
300
250
250
200
200
S ales /hour ($)
S ales /hour ($)
Scatter Plot of Sales & Traffic Data
150 100
100
50
50
-
y = a + bx
150
0
50
100
150
200
250
0
50
100
150
Cars/hour
200
250
Cars/hour Figure 2, p274
Use Regression to fit a Linear Function
Figure 3, p275
Regression computes three types of errors n
^
–
RSS =
(Yi – Y )2 Σ i=1
ESS =
(Yi – Yi )2 Σ i=1
TSS =
– (Yi – Y )2 Σ i=1
n
^
Regression Residual
n
Total
TSS = ESS + RSS
R2 = Figure 5, p277
RSS TSS
5
Should we build a gas station at Buffalo Grove where traffic is 183 cars/hour?
How confident are we in this forecast?
^ y
=
+ -
2 * Standard Error (Se)
+
b
*
x
Sales/hour = 57.104 + 0.92997 * 183 cars/hr
= $227.29
Confidence intervals use the following statistics: 1.00 =68%
1.96 = 95.0%
n
Yˆ
a
Se =
^
(Yi – Yi )2 Σ i=1 n – k -1
=
3.00 = 99.7%
ESS n – k -1
6
Excel calculates the Standard Error for us.
The 95% confidence interval is: [227.29 – 2(44.18); 227.29 + 2(44.18)] [$138.93; $315.65]
Fitting a Quadratic Function Other important information: 9T-statistic and its p-value 9Upper & Lower 95% 9F significance 9R2 and Adjusted R2
Figure 5, p277
Fitting a Quadratic Function
Use Solver to fit a Quadratic Function
300 250
Sales/hour ($)
200
y = a0 + a1x + a2x2 150 100 50 0
50
100
150
200
250
Cars/hour
Figure 10, p283
y = a0 + a1x + a2x2
Figure 7, p281
7
We could create a formula that exactly passes through every data point…..
Which curve to fit?
n
SSE =
^ Σ (Yi – Yi )2
i=1
But, why wouldn’t we want to do that?
Goodness of fit statistics: Sum of Squared Errors (SSE)
Goodness of fit statistics: Mean Squared Error (MSE)
8
MSE =
Sum of Squared Errors
Regression: SSE = 5854 Quadratic: SSE = 4954
(# of points – # of parameters)
Causal Forecasting Models Negative Slope indicates downward trend
Positive Slope indicates upward trend
140
140
120
120
y = 5x + 5
y = -5x + 135 Customers
100
80 60
80 60
40
40
20
20
Time
Time-Series Forecasting Models
21
25
23
15
19
17
13
11
5
9
3
7
25
21
23
15
19
13
17
11
5
9
7
1
1
0
0
3
Regression: MSE = 5854 / (5 - 2) = 1951.3 Quadratic: MSE = 4954 / (5 - 3) = 2477.0
Profit
100
Time
1. Curve Fitting: a) Linear b) Quadratic
Time is the independent variable
9
2. Moving Averages (Naive) a) Simple n-Period Moving Avg b) Weighted n-Period Moving Avg
3. Exponential Smoothing a) Basic model b) Holt’s Model (trend)
Curve Fitting Seasonality
Plot historical values as function of time and draw a linear “trend line”. Use trend line to predict future value.
Moving Averages (Naïve): Use previous period’s actual value to forecast the current period (i.e., use 12th value to predict 13th value). The Bank of Laramie
10
Moving Averages:
Simple n-Period Moving Averages:
Use average of past 12 values as best forecast for 13th value.
Use average of the most recent 6 values to predict 13th value.
Steco: a simple nperiod moving averages forecasting model
MONTH Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec.
ACTUAL SALES ($000s) 20 24 27 31 37 47 53 62 54 36 32 29
THREE-MONTH SIMPLE MOVING AVERAGE FORECAST
(20 (24 (27 (31 (37 (47 (53 (62 (54
+ + + + + + + + +
24 27 31 37 47 53 62 54 36
+ + + + + + + + +
27)/3 31)/3 37)/3 47)/3 53)/3 62)/3 54)/3 36)/3 32)/3
= = = = = = = = =
FOUR-MONTH SIMPLE MOVING AVERAGE FORECAST
23.67 27.33 31.67 38.33 45.67 54.00 56.33 50.67 40.67
(20 + (24 + (27 + (31 + (37 + (47 + (53 + (62 +
24 27 31 37 47 53 62 54
+ + + + + + + +
27 + 31 + 37 + 47 + 53 + 62 + 54 + 36 +
31)/4 37)/4 47)/4 53)/4 62)/4 54)/4 36)/4 32)/4
= = = = = = = =
25.50 29.75 35.50 42.00 49.75 54.00 51.25 46.00
Three- and Four- Month Simple Moving Averages
yˆ
16
Goodness of fit statistics
MAD =
MAPE =
∑
∑
actual sales − forecast sales
all forecasts
all forecasts
number of forecasts
actual sales − forecast sales actual sales number of forecasts
∗ 100 %
=
y
15
+
y
14
+
y
13
+
y
12
4
Table 1, p290
STECO: Simple n-Period Moving Average
MONTH Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec.
ACTUAL SALES ($000s) 20 24 27 31 37 47 53 62 54 36 32 29
THREE-MONTH FOUR-MONTH SIMPLE MOVING SIMPLE MOVING AVERAGE ABSOLUTE AVERAGE FORECAST ERROR FORECAST
$ $ $ $ $ $ $ $ $
23.67 27.33 31.67 38.33 45.67 54.00 56.33 50.67 40.67
7.33 9.67 15.33 14.67 16.33 0.00 20.33 18.67 11.67
SUM = MAD =
114.00 12.67
$ $ $ $ $ $ $ $
ABSOLUTE ERROR
25.50 29.75 35.50 42.00 49.75 54.00 51.25 46.00
11.50 17.25 17.50 20.00 4.25 18.00 19.25 17.00
SUM = MAD =
124.75 15.59
Figure 14, p291
11
Simple n-Period Moving Average forecasting models have two shortcomings
Philosophical Shortcoming Most recent observations receive no more weight or importance than older observations.
Operational Shortcoming All historical data used to make forecast must be stored in some way to calculate the forecast.
Weighted n-Period Moving Averages:
Weighted n-Period Moving Averages:
Recent data is more important than old data
Use weighted average of previous values & assign higher weights to more recent observations
Resolves philosophical shortcoming of simple period moving average forecasting
yˆ = α 0 y6 + α1 y5 + α 2 y4
12
Constraints:
Constraints:
The α' s (weights) are positive numbers
Smaller weights are assigned to older data
alpha2 = alpha1 = alpha0 = SUM OF WTS=
Constraints:
0.167 0.333 0.500 1.00
Month Actual Sales (000) 3month WMA Fcst Absolute Error January 20 February 24 March 27 April 31 24.83 6.17 May 37 28.50 8.50 June 47 33.33 13.67 July 53 41.00 12.00 August 62 48.33 13.67 September 54 56.50 2.50 October 36 56.50 20.50 November 32 46.34 14.34 December 29 37.01 8.01 99.35 11.04
Sum = MAD =
All the weights sum to 1 Use Solver to find the optimal weights Figure 16, p293
Weighted n-Period Moving Averages resolves philosophical shortcoming of simple period moving average forecasting
Exponential Smoothing resolves operational shortcoming of simple period moving average forecasting
13
Exponential Smoothing
Exponential Smoothing Forecast for t + 1
Observed in t
Forecast for t
yˆ t +1 = αy t + (1 − α )yˆ t
α
Where
Resolves operational shortcoming of the Moving Averages Model: Number of 8-period Moving Inventory Items Average Model to Forecast
5,000
0.500
alpha =
Actual Sales (000) Fcst Sales Absolute Error Month January 20 20.00 February 24 20.00 4.00 March 27 22.00 5.00 April 31 24.50 6.50 May 37 27.75 9.25 June 47 32.38 14.63 July 53 39.69 13.31 August 62 46.34 15.66 September 54 54.17 0.17 October 36 54.09 18.09 November 32 45.04 13.04 December 29 38.52 9.52
Exponential Smoothing Model
40,000: 5,000 * 8
is a user-specified constant
10,001: 5,000 yˆ t 5,000 y t 1α
Saving alpha and the last forecasts stores all the previous forecasts. When t = 1,
the expression becomes:
yˆ t +1 = αy t + (1 − α )yˆ t
VARIABLE yt
Does exponential smoothing produce a better forecast?
yˆ 2 = αy t + (1 − α )yˆ t
COEFFICIENT
y t-1
α α(1-α)
y t-2
α(1-α)
2 3
y t-3
α(1-α)
y t-4
α(1-α)
y t-5
α(1-α)
y t-6
α(1-α)
4 5 6 7
y t-7
α(1-α)
y t-8
α(1-α)
y t-9
α(1-α)
y t-10 Sum of the Weights
109.17 9.92
Sum = MAD =
8 9
10
α(1-α)
α = 0.1 0.1
α = 0.3 0.3
α = 0.5 0.5
0.09
0.21
0.25
0.081
0.147
0.125
0.07290
0.10290
0.06250
0.06561
0.07203
0.03125
0.05905
0.05042
0.01563
0.05314
0.03529
0.00781
0.04783
0.02471
0.00391
0.04305
0.01729
0.00195
0.03874
0.01211
0.00098
0.03487 0.68619
0.00847 0.98023
0.00049 0.99951
The value of alpha affects the performance of the model
Case 1: Response to Sudden Change System Change when t = 100 2 1
yt
Response to a Unit Change in y t
1 0 -1 94
95
96
97
98
99
t
100
101
102
103
1.2 1.1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 100
105
110
115
120
125
t
A forecasting system with alpha = 0.5 responds quickly to changes in the data.
14
Case 2: Response to Steady Change
Case 2: Response to Steady Change
Steadily Increasing Values of yt (Linear Ramp)
Steadily Increasing Values of yt (Linear Ramp)
yt
6 5 4 3 2 1 0
yt
0
2
4
6
8
10
6 5 4 3 2 1 0 0
t
2
4
6
8
10
t
Exponential smoothing is not a good forecasting tool in a rapidly growing or a declining market.
But the model can be adjusted (Holt’s model / exponential smoothing w/trend)
Case 3: Response to Seasonal Change Seasonal Pattern in y t 1.5 1.3 1.1 0.9
yt
0.7 0.5 0.3 0.1 -0.1 0
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17
t
Seasonality Exponential smoothing is not a good model to use here because it ignores the seasonal pattern.
Takes into consideration and adjusts for the seasonal patterns in data
1. Look at original data to see seasonal pattern. Examine the data & hypothesize an m-period seasonal pattern.
15
2. Deseasonalize the Data
3. Forecast using deseasonalized data
4. Seasonalize the forecast to account for the seasonal pattern
Gillett Coal Mine
Coal Receipts Over a Nine-Year Period
Deseasonalized Data
3,000
3,000.0 2,500.0 Coal (000 Tons)
Coal (000 Tons)
2,500
2,000
1,500
1,000
2,000.0 1,500.0 1,000.0 500.0 -
500
1- 1- 1- 1- 2- 2- 2- 2- 3- 3- 3- 3- 4- 4- 4- 4- 5- 5- 5- 5- 6- 6- 6- 6- 7- 7- 7- 7- 8- 8- 8- 8- 9- 9- 9- 91 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 9-3
9-1
8-3
8-1
7-3
7-1
6-3
6-1
5-3
5-1
4-3
4-1
3-3
3-1
2-3
2-1
1-3
1-1
0
Tim e (Year & Qtr)
Tim e (Ye ar and Quarte r)
1. Look at original data to see seasonal pattern. Examine the data & hypothesize an m-period seasonal pattern. Figure 27, p303
2. Deseasonalize the Data
Figure 32, p306
16
2.Deseasonalize the Data
Time Coal 4 Period Year-Qtr Receipts Moving Average 1-1 2,159 ----1-2 1,203 ----1-3 1,094 1,613 1-4 1,996 1,594 2-1 2,081 1,626 2-2 1,332 1,721 2-3 1,476 1,856 (2,159+1,203+1,094+1,996)/4 2-4 2, 533 1,898 3-1 2,249 1,948 3-2 1,533 2,063 3-3 1,935 2,060 3-4 2,523 2,050 4-1 2,208 2,066
a) Calculate a series of m-period moving averages, where m is the length of the seasonal pattern. b) Center the moving average in the middle of the data from which it was calculated. c) Divide the actual data at a given point in the series by the centered moving average corresponding to the same point. d) Develop seasonal index e) Divide actual data by the seasonal index
= 1,613
a) Calculate a series of m-period moving averages, where m is the length of the seasonal pattern. Figure 28, p304
4 Period Moving Average
Data & Centered Moving Average
Centered Moving Average 3,000
1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050 2,066 2,061 2,112 2,213
1,603 1,610 1,674 1,788 1,877 1,923 2,005 2,061 2,055 2,058 2,064 2,087 2,163 2,255
Receipts
2,500
(1613 + 1594)/2 = 1603
Coal (000 Tons)
Time Coal Year-Qtr Receipts 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523 4-1 2,208 4-2 1,597 4-3 1,917 4-4 2,726
Centered Moving Average
2,000
1,500
1,000
500
0 1- 1- 1- 1- 2- 2- 2- 2- 3- 3- 3- 3- 4- 4- 4- 4- 5- 5- 5- 5- 6- 6- 6- 6- 7- 7- 7- 7- 8- 8- 8- 8- 9- 9- 9- 91 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
Time (Ye ar & Qtr)
b) Center the moving average in the middle of the data from which it was calculated.
b) Center the moving average in the middle of the data from which it was calculated. Figure 29, p305
Figure 28, p304
Time Coal Year-Qtr Receipts 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523
4 Period Moving Average
Centered Moving Average
Ratio of Coal Receipts to Centered Moving Average
1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050
1,603 1,610 1,674 1,788 1,877 1,923 2,005 2,061 2,055 2,058
0.682 1.240 1.244 0.745 0.787 1.317 1.122 0.744 0.942 1.226
1,094 / 1,603 = 0.682
c) Divide the actual data at a given point in the series by the centered moving average corresponding to the same point. Figure 28, p304
Ti m e C o al Y e ar - Q tr R ec e i pt s 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523
4 P e ri o d Mo v i ng A ve r a ge
1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050
C e n te r ed Ra t i o o f C o al R ec e ip t s t o S e as o n al M o vi n g A v er a g e C e n te r ed M ov i n g A v er a ge I n di c e s 1.112 0.786 1,603 0.682 0.863 1,610 1.240 1.238 1,674 1.244 1.112 1,788 0.745 0.786 1,877 0.787 0.863 1,923 1.317 1.238 2,005 1.122 1.112 2,061 0.744 0.786 2,055 0.942 0.863 2,058 1.226 1.238
d) Develop seasonal index for each quarter • Group ratios by quarter • Average all of the ratios to moving averages quarter by quarter • Add Seasonal Indices data to table • Normalize the seasonal index
17
Deseasonalized Data 3,000.0
2,000.0 1,500.0 Deseasonalized Data
1,000.0 500.0
93
83 91
71 73 81
61 63
51 53
41 43
-
31 33
1,603 1,610 1,674 1,788 1,877 1,923 2,005 2,061 2,055 2,058
Deseasonalized Data 1,941.0 1,529.8 1,267.7 1,611.9 1,870.9 1,693.8 1,710.3 2,045.6 2,021.9 1,949.4 2,242.2 2,037.5
21 23
1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050
Ratio of Coal Receipts to Seasonal Centered Moving Average Indices 1.112 0.786 0.682 0.863 1.240 1.238 1.244 1.112 0.745 0.786 0.787 0.863 1.317 1.238 1.122 1.112 0.744 0.786 0.942 0.863 1.226 1.238
13
Centered Moving Average
11
4 Period Moving Average
Coal (000 Tons)
2,500.0 Time Coal Year-Qtr Receipts 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523
Tim e (Ye ar & Qtr)
e) Divide actual data by the seasonal index
e) Divide actual data by the seasonal index
Figure 31, p306
Time Coal Year-Qtr Receipts 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523
4 Period Moving Average
Centered Moving Average
1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050
1,603 1,610 1,674 1,788 1,877 1,923 2,005 2,061 2,055 2,058
Ratio of Coal Receipts to Seasonal Centered Moving Average Indices 1.108 0.784 0.682 0.860 1.240 1.234 1.244 1.108 0.745 0.784 0.787 0.860 1.317 1.234 1.122 1.108 0.744 0.784 0.942 0.860 1.226 1.234
Deseasonalized Data Forecast 1,948.1 1,948.1 1,535.4 1,948.1 1,272.3 1,678.5 1,617.8 1,413.1 1,877.8 1,546.8 1,700.0 1,763.0 1,716.6 1,721.9 2,053.1 1,718.4 2,029.3 1,937.1 1,956.5 1,997.4 2,250.4 1,970.7 2,045.0 2,153.4
3. Forecast method in deseasonalized terms • Review the graphed deseasonalized data to reveal pattern • Use forecasting method that accounts for the pattern in the deseasonalized data • Use Excel’s Solver to minimize the error
Figure 32, p306
Time Coal Year-Qtr Receipts 1-1 2,159 1-2 1,203 1-3 1,094 1-4 1,996 2-1 2,081 2-2 1,332 2-3 1,476 2-4 2,533 3-1 2,249 3-2 1,533 3-3 1,935 3-4 2,523
4 Period Moving Average --------1,613 1,594 1,626 1,721 1,856 1,898 1,948 2,063 2,060 2,050
Centered Moving Average --------1,603 1,610 1,674 1,788 1,877 1,923 2,005 2,061 2,055 2,058
Ratio of Coal Receipts to Seasonal Centered Moving Average Indices ----1.108 ----0.784 0.682 0.860 1.240 1.234 1.244 1.108 0.745 0.784 0.787 0.860 1.317 1.234 1.122 1.108 0.744 0.784 0.942 0.860 1.226 1.234
Deseasonalized Seasonalize Data Forecast Forecast 1,948.1 1,948.1 2,159.000 1,535.4 1,948.1 1,526.409 1,272.3 1,678.5 1,443.212 1,617.8 1,413.1 1,743.439 1,877.8 1,546.8 1,714.276 1,700.0 1,763.0 1,381.390 1,716.6 1,721.9 1,480.540 2,053.1 1,718.4 2,120.128 2,029.3 1,937.1 2,146.723 1,956.5 1,997.4 1,564.974 2,250.4 1,970.7 1,694.495 2,045.0 2,153.4 2,656.854
4. Reseasonalize the forecast to account for the seasonal pattern • Multiply the deseasonalized forecast by the seasonal index for the appropriate period. • Graph the actual Coal Receipts and Seasonalized Forecast
Figure 33, p307
1. Look at original data to see seasonal pattern. Examine the data & hypothesize an m-period seasonal pattern.
Actual & Forecast
Coal (000 Tons)
3,500 3,000
2. Deseasonalize the data.
2,500 2,000 1,500 1,000 500 0
Coal Receipts Seasonalized Forecast 1- 1- 1- 1- 2- 2- 2- 2- 3- 3- 3- 3- 4- 4- 4- 4- 5- 5- 5- 5- 6- 6- 6- 6- 7- 7- 7- 7- 8- 8- 8- 8- 9- 9- 9- 9- 1 1 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 0- 01 2
Year-Quarter
4. Reseasonalize the forecast to account for the seasonal pattern
a) Calculate a series of m-period moving averages, where m is the length of the seasonal pattern. b) Center the moving average in the middle of the data from which it was calculated. c) Divide the actual data at a given point in the series by the centered moving average corresponding to the same point. d) Develop seasonal index e) Divide actual data by the seasonal index 3. Forecast method in deseasonalized terms. 4. Reseasonalize the forecast to account for the seasonal pattern.
18
SUMXMY2( ) Returns the sum of squares of differences of corresponding values in two arrays. Syntax: SUMXMY2(array_x,array_y), where Array_x is the first array or range of values. Array_y is the second array or range of values. The equation for the sum of squared differences is:
SUMXMY2 =
∑ (x − y )2
Measures of Comparison MAD =
MAPE =
∑
number of forecasts
∑
all forecasts
n
MSE =
actual sales − forecast sales
all forecasts
actual sales − forecast sales actual sales number of forecasts
∑ ( actual
∗ 100 %
sales − forecast sales ) 2
t =1
number of forecasts
Model Validation
Create experience by simulating the past.
Create the model with a portion of the historical data.
Use remaining data to see how well the model would have performed.
19
Qualitative Forecasting Models
Expert Judgment
Consensus Panel
Delphi Method
Coordinator requests forecasts
Delphi Method
Coordinator receives Individual forecasts Coordinator determines (a) Median response (b) Range of middle 50% of answers Coordinator sends to all experts (a) Median response (b) Range of middle 50% (c) Explanations
Coordinator requests explanations from any expert whose estimate is not in the middle 50%
Grassroots Forecasting
20
Market Research
21