Professional Documents
Culture Documents
8/23/02
2:28 PM
Page 91
C H A P T E R
Forecasting Models
11
errors. MAPE is important because it can be used to help increase forecasting accuracy. 11-10. We can draw line plots of the actual and forecast values for each observation (or time period). Such line plots are automatically drawn by most forecasting software including ExcelModules. The line plots can be used to show whether there are sizable errors in the forecast. In addition, they can be used to detect whether the forecasting model does a good job of replicating the pattern of actual values over the past few time periods. For the model to be valid, there should be no consistent under- or over-forecast seen, and forecast errors must be randomly distributed. 11-11. The correlation coefcient is a measure of the strength of the linear relationship between two variables. That is, it measures how one variables value is linearly related to changes in the value of the other variable. It is usually denoted by r and can be any number between and including 1 and 1. A value of 1 indicates the two variables are perfectly correlated in a positive manner (i.e., if either variable increases in value, the other one follows suit). A value of 1 indicates the two variables are perfectly correlated in a negative manner. Finally, a value of 0 indicates the two variables are not linearly correlated. 11-12. To use Solver to determine the optimal weights in the weighted moving average model, we set the weights as the decision variables (or Changing Cells). The objective (or Target Cell) is the measure of forecast error, such as MAD, MSE, or MAPE, which we wish to minimize. If we want to specify that the weights must add up to 1, we must include it as a constraint in the model. The only other constraint is the non-negativity constraint on the decision variables (weights). The Assume Linear Model option should not be checked in solving this problem since the formula for the objective function is nonlinear. 11-13. Y 36 4.3 X1 a. Y b. Y c. Y 36 337 36 380 36 423 (4.3)(90) (4.3)(80) (4.3)(70)
91
6634 CH11 UG
8/23/02
2:28 PM
Page 92
92
CHAPTER 11
FORECASTING MODELS
sum of the weights [(2 4) 6 4]/4 412 [(2 5) 4 6]/4 50 [(2 10) 5 4]/4 714 [(2 8) 10 5]/4 734 [(2 7) 8 10]/4 80 [(2 9) 7 8]/4 814 [(2 12) 9 7]/4 10 [(2 14) 12 9]/4 1214
22.92% 21.17%
The weighted moving average appears to be slightly more accurate in its annual forecasts. 11-15. See le P11-15.XLS
Year 1 2 3 4 5 6 7 8 9 10 11 Demand 4 6 4 5 10 8 7 9 12 14 15 Two-Year Moving Average Four-Year Moving Average
(4 6)/2 (6 4)/2 (4 5)/2 (5 10)/2 (10 8)/2 (8 7)/2 (7 9)/2 (9 12)/2 (12 14)/2 MAPE
22.57%
27.07%
11-16.
MAPE3-MA MAPE2-MA
22.92% 22.57%
MAPE3-WMA MAPE4-MA
21.17% 27.07%
I would use the 3-year weighted moving average since it has the lowest MAPE ( 21.17%). 11-17. 0.3. New forecast for year 2 is last periods forecast (last periods actual demand last periods forecast): new forecast for year 2 5,000 5,000 5,000 (0.3)(4,000 (0.3)( 300 5,000) 1,000)
6634 CH11 UG
8/23/02
2:28 PM
Page 93
CHAPTER 11
FORECASTING MODELS
93
The MAPE can be used to determine which forecasting method is more accurate. MAPE for this model 25.50% The 3-year weighted moving average model continues to appear to be more accurate. Year 1 2 3 4 5 6 11-18. MAPE Forecast 410.0 422.0 443.9 466.1 495.2 521.8 11-19. See le P11-19.XLS.
Year 1 2 3 4 5 6 Sales 450 495 518 563 584 ? Forecast Using 410 (0.6) (450 434 (0.6) (495 470.6 (0.6)(518 499 (0.6) (563 537.4 (0.6)(584 MAPE 10.20% 0.6 434 470.6 499.0 537.4 565.6 Forecast Using 410 446 490.1 515.21 558.221 MAPE (0.9)(450 (0.9)(495 (0.9)(518 (0.9)(563 (0.9)(584 7.05% 0.9 410) 446) 490.1) 515.21) 558.2) 446 490.1 515.21 558.2 581.4
MAPE 0.9 7.05% Because it has the lowest MAPE, the smoothing constant 0.9 gives the most accurate forecast. Note: Using Solver, we nd the optimal to be 1.0 (see le P11-20.XLS). The MAPE with 1.0 is 6.28%. 11-21. See le P11-21.XLS
Year 1 2 3 4 5 6 Sales 450 495 518 563 584 ? Three-Year Moving Average
4 5 6 Appearances
The observations do not form a perfect straight line but approach linearity over the range shown. b. (See le P11-24.XLS)
TV Appearances, X 3 4 7 6 8 35 33 Demand for Drums, Y 3 6 7 5 10 38 39
11.71%
1 through 5.
X2 9 16 49 36 64 125 199
XY 9 24 49 30 80 240 232
X2
XY
Y 421.2 33.6X Projected sales in year 6 (X Y 421 (33.6)(6) 622.80 MAPE 1.08% 11-23. MAPE
0.3
6) is:
15.37% 11.71%
MAPEmoving average
MAPEregression 1.08% Regression is obviously the preferred method because of its low MAPE.
a = Y bX = 39 / 6 (1)( 33 / 6 )
6634 CH11 UG
8/23/02
2:28 PM
Page 94
94
CHAPTER 11
FORECASTING MODELS
1.0X, MAPE
23.21%.
d. There are many other factors to consider, including seasonality and any underlying causal variables such as advertising budget. 11-26. a. See le P11-26.XLS.
Actual Miles 17 21 19 23 18 16 20 18 22 20 15 22 0.2 Forecast (Ft) 17.00 17.00 17.80 18.04 19.03 18.83 18.26 18.61 18.49 19.19 19.35 18.48
11-25. To answer the discussion questions, two forecasting models are required: a three-period moving average and a three-period weighted moving average. Once the actual forecasts have been made, their accuracy can be compared using the MAPE. a, b. (See le P11-25.XLS)
Period 4 5 6 7 8 9 10 11 12 13 Month Apr. May June July Aug. Sept. Oct. Nov. Dec. Jan. Demand 10 15 17 11 14 17 12 14 16 11 Moving Average 13.67 13.33 13.67 14 14.33 14 14 14.33 14.33 14 Weighted Average 14.5 12.67 13.5 15.17 13.67 13.50 15 14 13.83 14.67
Week 1 2 3 4 5 6 7 8 9 10 11 12
Error 4.00 1.20 4.96 1.03 2.83 1.74 0.61 3.51 0.81 4.35 3.52
b.
c. MAPE for moving average is 17.14%. MAPE for weighted average is 21.37. Moving average forecast for February is 13.6667. Weighted moving average forecast for February is 13.1667. Because a three-period average forecasting method is used, forecasts start for period 4. Based on this analysis, the moving average appears to be more accurate. The forecast for February is about 14.
Actual Value, At 50 35 25 40 45 35 20 30 35 20 15 40 55 35 25 55 55 40 35 60 75 50 40 65 Smoothed Value, Ft ( 0.1) 50 50 48 46 45 45 44 42 41 40 38 36 36 38 38 37 38 40 40 40 42 45 45 45 47
11-27. a, b. See le P11-27.XLS. The accompanying table shows a comparison of the calculations for the exponentially smoothed forecasts using constants of 0.1 and 0.6. c. Students should note how stable the smoothed values for the 0.1 smoothing constant are. When compared to actual week 25 calls of 85, the 0.6 smoothing constant appears to do a better job. On the basis of the forecast error, the 0.6 constant is better also. However, other smoothing constants need to be examined. For example, using Solver we nd that the smoothing constant which minimizes MAPE is 0.279. The MAPE in this case is 35.06%. (See le P11-27.XLS)
Smoothed Value, Ft ( 0.6) 50 41 31 37 42 38 27 29 32 25 19 32 46 39 31 45 51 44 39 51 66 56 46 58 MAPE
Week, t 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
6634 CH11 UG
8/23/02
2:28 PM
Page 95
CHAPTER 11
FORECASTING MODELS
95
0.9 (See le
11-31. See le P11-31.XLS a. The regression equation is Y a bX, or Y 1 1X The standard error of the estimate is 1.15.
Week 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
The coefcient of correlation is 0.845 b. 11-32. See le P11-32.XLS a. It appears from the graph that the points scatter around a straight line. b. The regression relationship is Y 5.060 1.593X. c. For the tourist level of 10,000,000, the model predicts that 2,099,000 riders will use the system. That is, Y 5.060 1.593(10) 20.99, or 2,099,000 people d. If X 0 tourists, the model produces a ridership of 506,000 people. This is probably a very erroneous forecast, however, because the value X 0 is outside the range of observed values.
Ridership (100,000s)
of 0.6 seems to
11-33. See le P11-33.XLS. Y a bX, where Y number of patients and X From ExcelModules Y 29.733 3.285X For: (X (X (X 11): Y 12): Y 13): Y 29.733 29.733 29.733 (3.28)(11) (3.28)(12) (3.28)(13)
year.
Forecast 0.1 (70 0.1(68.5 0.1(64.8 0.1(71.7 0.1(71.3 0.1(72.8 65) 65.5) 65.8) 65.7) 66.3) 66.8)
Comparing the trend line with the actual observations (see le P11-33.XLS), we see that the model is reasonably accurate (albeit not exceptionally accurate). The MAPE is 7.01%. 11-34. See le P11-34.XLS. Let, Y a bX, Where Y rate. Y 1.229 0.545X For: X 131.2 : Y 90.6 : Y 1.229 1.229 number of patients and X crime
4.74%
(0.545)(131.2) (0.545)(90.6)
72.7 50.6.
Based on MAPE,
0.1.
6634 CH11 UG
8/23/02
2:28 PM
Page 96
96
CHAPTER 11
FORECASTING MODELS
Therefore crime rate crime rate 131.2 l 72.7 patients 90.6 l 50.6 patients
11-35. To compute a seasonalized or adjusted sales forecast, we just multiply each seasonal index by the appropriate trend forecast. index Y Y
seasonal trend forecast
200 2 11-37. (See le P11-37.XLS) Let, Y Y a bX, Where Y 0.003X GPA and X SAT score 1.028
250
225 250
1200 (0.9) 4
The trend and regression models are now much more precise. 11-39. See le P11-39.XLS for the calculations using centered moving averages. We have data for 16 period, with 4 seasons (quarters) per year. The MAPE is just 1.06% indicating that the multiplicative decomposition model does a very good job of predicting quarterly sales. This is also indicated by the error graph which shows that the forecasted and actual values are nearly identical. Forecasts for 2002 are as follows: Quarter 1 66.1, Quarter 3 72.8, and Quarter 4 98.3. 59.4, Quarter 2
As an indication of the usefulness of this relationship, we can calculate the correlation coefcient: r and r2 0.479 A correlation coefcient of 0.692 is not particularly high. The coefcient of determination, r 2, indicates that the model explains only 47.9% of the overall variation. Therefore, while the model does provide an estimate of GPA, there is considerable variation in GPA which is as yet unexplained. For: X 350: Y 1.028 (0.003)(350) 2.227 X 800: Y 1.028 (0.003)(800) 3.77 11-38. a. See le P11-38.XLS. Comparing the MAPE values for the three models, we see that: MAPEExponential MAPETrend MAPERegression 20.89% 1019.32% 707.88% 0.692
11-40. See le P11-40.XLS for the calculations using centered moving averages. We have data for 48 periods with 12 seasons (months) per year. The MAPE is 4.58% indicating that the multiplicative decomposition model does a reasonable job of predicting quarterly sales. This is also indicated by the error graph which shows that the forecasted and actual values are very close to each other.
The exponential smoothing model seems to be the clear choice for the best model. b. With older data excluded (and a case can easily be made for doing so), the conclusions may reverse. For example, if we include the data for just last 16 years (1986 onwards), the revised MAPE values are (see le P11-38.XLS): MAPEExponential 17.05%
6634 CH11 UG
8/23/02
2:28 PM
Page 97
CHAPTER 11
FORECASTING MODELS
97
Southeast Airlineengine maintenance cost: Cost 0.671 0.0041 airframe age Coefcient of determination 0.460 Coefcient of correlation 0.678 The graphs portray both the actual data and the regression lines for airframe and engine maintenance costs for both airlines. Note that the two graphs have been drawn to the same scale to facilitate comparisons between the two airlines. Northern Airline: There seem to be modest correlations between maintenance costs and airframe age for Northern Airline. There is certainly reason to conclude, however, that airframe age is not the only important factor. Southeast Airline: The relationships between maintenance costs and airframe age for Southeast Airline are much less well dened. It is even more obvious that airframe age is not the only important factorperhaps not even the most important factor. Overall, it would seem that: 1. Northern Airline has the smallest variance in maintenance costs, indicating that the day-to-day management of maintenance is working pretty well. 2. Maintenance costs seem to be more a function of airline than of airframe age. 3. The airframe and engine maintenance costs for Southeast Airline are not only lower but more nearly similar than those for Northern Airline, but, from the graphs at least, appear to be rising slightly more sharply with age. 4. From an overall perspective, it appears that Southeast Airline may perform more efciently on sporadic or emergency repairs, and Northern Airline may place more emphasis on preventive maintenance. Ms. Youngs report should conclude that: 1. There is evidence to suggest that maintenance costs could be made to be a function of airframe age by implementing more effective management practices. 2. The difference between maintenance procedures of the two airlines should be investigated. 3. The data with which she is presently working do not provide conclusive results.
Airframe Engine
Utilizing ExcelModules, we can develop the following regression equations for the variables of interest. See le P11-Airline.XLS for the calculations. Each regression is shown on a separate worksheet.
Northern Airline 90 80 70 Cost ($) 60 50 40 30 20 10 5 7 9 11 13 15 17 Average Airframe Age (Thousands) Southeast Airline 90 80 70 Cost ($) 60 50 40 30 20 10 5 7 9 11 13 15 17 Average Airframe Age (Thousands) 19 19 Airframe Engine
Northern Airlineairframe maintenance cost: Cost 36.097 0.0026 airframe age Coefcient of determination 0.769 Coefcient of correlation 0.877 Northern Airlineengine maintenance cost: Cost 20.571 0.0026 airframe age Coefcient of determination 0.612 Coefcient of correlation 0.783 Southeast Airlineairframe maintenance cost: Cost 4.597 0.0032 airframe age Coefcient of determination 0.390 Coefcient of correlation 0.625
6634 CH11 UG
8/23/02
2:28 PM
Page 98
98
CHAPTER 11
FORECASTING MODELS
b. With a ticket price of $20, revenues in 2003 (games 31 to 35) are expected to be:
Game 1 2 3 4 5 Attendance 48098 55055 51451 36116 49559 Revenue $961,968 $1,101,106 $1,029,023 $722,312 $991,177
50% of childrens fees, plus 15% of group fees. The weighted fees each year (19891998) are $0.975, $0.975, $0.975, $0.975, $1.275, $1.775, $1.775, $2.275, $2.20, $2.875. See sheet #2 in le P11-Akron.XLS. Gate admissions r 31,598.4 39,402.3 given year) 0.848 (average fee in
With a ticket price of $21 (5% increase in 2004), revenues in 2004 (games 36 to 40) are expected to be:
Game 1 2 3 4 5 Attendance 50520 57799 53990 37880 51957 Revenue $1,060,915 $1,213,784 $1,133,793 $795,488 $1,091,100
MAPE 14.36% If we assume that admission fees are not raised in 1999 and 2000, expected gate admissions 144,880 in each year and revenues $416,530. Comparing the earlier time-series model to this second regression, we note that the r is higher and MAPE is lower in the time-series approach. 2. The student should respond that the other factors are the variability of the weather, the special events, the competition, and the role of advertising.
Kwik Lube
c. It appears that the school will run out of space for the homecoming game in 2004. It may be time for Dr. Starr to start to a fund raising campaign for a stadium expansion project. Of course, they could run out of space even earlier if SWU does get the number-one ranking next year. Alternatively, attendance could drop off sharply if Pitterno does not achieve what is expected of him. 1. The relationship between Kwik Lube sales (y), average industry sales (x), and year (t with t 1 corresponding to 1992) is shown in the table below. The x and y values are in thousands of dollars. One could try a multiple regression analysis but the correlation of y with just x is 0.998, leading one to use the simple linear regression equation: y 2.99x 1.42. (See sheet #1 in le P11-Kwik.XLS)
t 1 2 3 4 5 6 7 8 x 22 25 24 26 33 35 39 44 y 68 75 75 78 99 104 120 133
The year 1991 was excluded since the Kwik Lube revenues were not for an entire year. 1999 (t 8) was the last year of Kwik Lube operation without the competition from Speedy Lube. The forecasted sales for 2000 would be estimated using the average industry sales of $47,000 for x: y 2.99(47) 1.42 141.95 and the forecasted sales for 2001 would use the industry sales of $52,000: y 2.99(52) 1.42 156.90 The estimated lost sales is the difference between the forecasted and actual sales: (141,950 $156,900) ($111,000 $111,000) $76,850. 2. Without the questionnaire study, the best estimate of lost sales would be from the regression of y on t: (See sheet #2 in le P11-Kwik.XLS) y 9.38t 51.8 with a somewhat lower correlation. The estimated lost sales would be $59,820, about $20,000 less than the estimate based on average industry sales. Even recovering as little as 10 percent of this difference would pay for the study.
So the forecasts for 1999 and 2000 are 145,523 and 154,720, respectively. Using a weighted average of $2.875 to represent gate receipts per person, revenues for 1999 and 2000 are $418,378 and $444,820, respectively. To complicate the situation further, students may legitimately use a regression model to forecast admission fees for each of the three categories, or for the weighted average fee. This number would then replace $2.875. Here is the result of a linear regression using weighted average admission fees as the predicting (independent) variable. Weights are obtained each year by taking 35% of adult fees, plus
6634 CH11 UG
8/23/02
2:28 PM
Page 99
CHAPTER 11
FORECASTING MODELS
99
3. The lawsuit led by Dick Johnson should discuss two basic areas which will build a sound case for damages being awarded in his favor. The rst factor involves the concept behind setting up a franchise. Franchises are designed so that independent owners can start a business with a well-known name (and consequently, with an already-captured market). This, coupled with proven strategies and expertise given to a franchise purchaser by the franchise seller, reduces the usually high probability of a new business going under in its infancy stage. The franchise fee is the cost paid for the reduced risk of a new enterprise. Naturally, the franchising rm will protect itself against competition in a franchise contract. A franchise holder who violates such clauses has, in essence, gained free proven strategies and has capitalized on them. Thus, the franchising rm has been damaged by the fact that a competitor has gained information without paying for it. This is the case with Kwik Lube. A franchise owner, T. A. Williams, has beneted from Johnsons expertise more than is justied by the monetary gains earned from franchise fees. This is not simply an economic issue, however, for such a situation was thought of before by Johnson. He had sought to protect himself with a noncompetition clause in his franchised contract. Thus, Williams is legally in the wrong for his breach of contract. What this rst area of discussion in the lawsuit does is to determine that there, in fact, has been damage done to the plaintiff, Johnson. The second area to be discussed in the lawsuit should deal with how those damages can be mitigated by the defendant, Williams. Usually in lawsuits, there is a problem with measuring the damage done. Johnson, however, can measure his loss by forecasting sales and then comparing actual sales to predicted sales. In summary, the lawsuit should discuss how damage was incurred to plaintiff, Johnson, and how said damage should and/or could be mitigated. A well-presented lawsuit or petition to the court should result in a favorable judgment for the owner of Kwik Lube.
The coefcient of determination is 78.8% and the standard error is 7.611. Students should be asked to interpret these values. The MAPE is 8.79%, which is slightly more than the MAPE obtained using the decomposition model. The error graph reveals a similar pattern to the earlier model, with sizable differences in some periods.