CSU Football Magazines Case Solution

Note that there are many ways to approach this case and that student solution may not match
exactly what is provided here.

1. There is a considerable amount of data available in the file MagazinesCSU, but not all of it
may be useful for your purposes here. Are there variables contained in the file
MagazinesCSU that you would exclude from a forecast model to determine football
magazine sales in Year 10? If so, why? Are there particular observations of football magazine
sales from previous years that you would exclude from your forecasting model? If so, why?

CSU must make their forecast for the upcoming season in July so that they can place orders for
magazines. Therefore, variables for data that are unknown in July when CSU must place their
order are of little help in creating a forecast for magazine sales. Variables such as Game Day
Weather, Total Game Attendance, and Kickoff Temperature will be unknowns in July. Thus, we
would choose to not include these in a forecasting model.

There is also strong intuitive reasons why some of the variables included here could describe the
same information, and thus, would be highly correlated. For instance, an Opponent’s Previous
Season Number of Wins will be highly correlated with an Opponent’s Previous Season Number
of Losses. We can verify this by calculating the correlation coefficient for Opponent’s Previous
Season Number of Wins and Opponent’s Previous Season Number of Losses which is –0.97.
Similarly, the correlation coefficient for CSU’s Previous Season Number of Wins and CSU’s
Previous Season Number of Losses is –0.99. Therefore, we would generally only want to include
one of these variables in any forecasting model. (As an alternative, we could combine the
variables by calculating Previous Season Win Percentage.) Other highly correlated variables are
Week in Season and Home Game Number (correlation coefficient = 0.97); and Kickoff
Temperature and Week in Season (correlation coefficient = –0.78). This latter correlation makes
sense when you consider that games later in the football season are likely in late autumn as
temperatures are decreasing.

Based on this analysis, we will delete Game Day Weather, Total Game Attendance, Kickoff
Temperature, Opponent’s Previous Season Number of Losses, CSU’s Previous Season Number
of Losses, and Home Game Number.

There are also several individual observations that we might consider. Season 8 Game 1 versus
Ohio A&M is different than all other data points because it was the CSU Throwback Jersey
Game. There are no similar observations to this game and no expectations for such a game in the
upcoming season. Therefore, we would have strong reason to discard this observation.

A box plot of Magazine Sales data is shown below. Here we see that there are several data points
that could be considered outliers. The most extreme is Season 3 Week 1 versus Lincoln
University. This game has Magazine Sales of 6,463. No other game had sales greater than 5,200.
And since this is the only game that was played against Lincoln University and CSU does not
play this team in the upcoming season, we would have strong reason to remove this data point
also.
The other two data points with extremely high magazine sales are Season 2, Week 2 versus
University of Ames and Season 1 Week 3 versus Urbana College. CSU plays the University of
Ames each year, including in the upcoming season. Therefore, it would be inadvisable to throw
out all data points where CSU plays University of Ames. Instead, we could create a new
categorical variable to indicate a game versus the University of Ames as it seems this might
affect magazine sales. However, CSU has only played Urbana College once and will not play the
team in the upcoming season. Therefore, we can probably remove this data point if we think it is
an outlier.

Based on our findings here, we will delete the observations for Season 8 Week 1, Season 3 Week
1, and Season 1 Week 3. Also, we will add a categorical variable for games where CSU plays the
University of Ames.

2. Based on the data in the file MagazinesCSU, develop a regression model to forecast the
average sales of football magazines for each of the seven home games in the upcoming
season (Year 10). That is, you should construct a single regression model and use it to
estimate the average demand for the seven home games in Year 10. In addition to the
variables provided, you may create new variables based on these variables or based on
observations of your analysis. Be sure to provide a thorough analysis of your final model
(residual diagnostics) and provide assessments of its accuracy. What insights are available
based on your regression model?
After removing the variables in Part 1, the predictor variables we are left to consider are Year,
Week in Season, Opponent Preseason Rank, Preseason Ticket Sales, CSU Preseason Rank,
Conference Game, Homecoming Game, Opponent’s Previous Season Number of Wins, CSU’s
Previous Season Number of Wins, and our new categorical variable for Playing University of
Ames. Below we create scatter charts to show the relation between each of these possible
predictor variables and Magazine Sales.

Magazine Sales (Units)

6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 1 2 3 4 5 6 7 8 9 10
Year

Magazine Sales (Units)

6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 2 4 6 8 10 12 14
Week in Season
 6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 20 40 60 80 100 120 140
Opponent Preseason Rank

6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
35,000 40,000 45,000 50,000 55,000 60,000
Preseason Ticket Sales
 6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 10 20 30 40 50 60 70 80
CSU Preseason Rank

6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 0.2 0.4 0.6 0.8 1 1.2
Conference Game (1= Yes; 0 = No)
 6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 0.2 0.4 0.6 0.8 1 1.2
Homecoming (1= Yes; 0 = No)
 6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 2 4 6 8 10 12 14
Opponent's Previous Season Number of Wins

6,000

5,000
Magazine Sales (units)

4,000

3,000

2,000

1,000

0
0 2 4 6 8 10 12
CSU's Previous Season Number of Wins

Based on these scatter charts, we see that there are several possible predictor variables that do
not seem to have much direct correlation with Magazine Sales such as Year, Homecoming, and
Opponent’s Previous Season Number of Wins. We will still consider these variables in our first,
“full” regression model, but we will want to closely examine the regression model output to see
if these variables are helping us predict magazine sales. The following output results from fitting
the “full” regression model with all the variables.
From this regression output, we see that the p values for Year, Opponent’s Previous Season
Number of Wins, and Homecoming are all quite high, indicating that these are not statistically
significant at the .05 (or even much higher) levels. This agrees with our scatter chart analysis that
showed little linear relation between each of these variables and Magazine Sales. Additionally,
the p values for Opponent Preseason Rank, CSU Preseason Rank, and Conference Game are also
all >0.2. Unless, we have strong theoretical reason to keep any of these variables, we would
likely want to remove these and see what effect removing the variables has on the overall
predictive power of our regression model. Note that the p value for the Intercept is also >0.05,
but because this is not within the range of our observations, we will keep the intercept in our
regression model.

Next we will attempt to fit a regression model that includes only the variables for Week in
Season, Preseason Ticket Sales, CSU’s Previous Season Number of Wins, and Playing
University of Ames. The output from this regression model appears below.
From the output of this regression model with fewer variables, we see that all included variables
are statistically significant at the 0.05 level. We also see that the R-squared value for this model
is .6058, compared to .6541 for the regression model that included many more variables. Given
that this model is much more parsimonious and the decrease in goodness-of-fit is relatively
small, it is preferred because it preserves inference and interpretability of coefficients.

<insert margin note:

A more direct method to evaluating a regression model’s predictive ability is to measure its
accuracy in estimating values of the dependent variable for observations that were not used to fit
the model. In this case, however, we have relatively few observations on which to train the
regression model, so holding out any years of data on which to validate the regression model is
not a feasible option.
>

We should also examine the residual plots for our suggested model to check for any possible
violations of model assumptions in our residual analysis. The residual plots appear below.
 Week In Season Residual Plot
1500

1000

500
Residuals

0
0 2 4 6 8 10 12 14
-500

-1000

-1500
Week In Season

Preseason Ticket Sales Residual Plot

1500

1000

500
Residuals

0
30,000 35,000 40,000 45,000 50,000 55,000 60,000
-500

-1000

-1500
Preseason Ticket Sales

CSU's Previous Season Number of Wins Residual Plot

1500
1000
500
Residuals

0
0 2 4 6 8 10 12
-500
-1000
-1500
CSU's Previous Season Number of Wins
 Playing U of Ames (1 = Yes; 0 = No) Residual Plot
1500

1000

500
Residuals

0
0 0 0 1 1 1 1
-500

-1000

-1500
Playing U of Ames (1 = Yes; 0 = No)

Based on these residual plots, we see no obvious violations. Thus, we will use this regression
model for our forecasting of Magazine Sales:

Magazine Sales = 1617.0325 – 108.0959 * (Week in Season) + 0.0307 * (Preseason Ticket Sales
+ 44.2507 * (CSU’s Previous Season Number of Wins) + 1080.4055 * (Playing U of Ames)

Several interesting interpretations are possible from this model. The model suggests that, holding
all other independent variables constant, Magazine Sales decrease by about 108 units each week
into the football season. We also see that, again holding all other independent variables constant,
one more win in CSU’s previous season leads to an increase of about 44 magazine sales per
game in the following season and that Playing the U of Ames increases magazine sales by more
than 1000.
3. Use the forecasting model developed in Part 2 to create a simulation model that Kris can use
to estimate the total football magazine sales amounts in Year 10. Your simulation model
should have seven uncertain inputs: one input for football magazine sales at each CSU game
in Year 10. Then you should sum these sales amounts for each individual game to create a
total football magazine sales amount for Year 10.

To construct a simulation model, we first need to determine which values are random, and then
how best to represent these random variables. The random variable of interest here is Game Day
Magazine Sales. From parts 1 and 2, we have developed a forecasting model, but this forecast
only provides an estimate of actual Magazine Sales. One way to estimate the distribution of the
Magazine Sales random variable is to examine the distribution of forecast errors using our
forecast model from Part 2 and the actual sales in years 1 through 9. For each game in years 1
through 9, we can create a forecast using our model:

Magazine Sales = 1617.0325 – 108.0959 * (Week in Season) + 0.0307 * (Preseason Ticket Sales)
+ 44.2507 * (CSU’s Previous Season Number of Wins) + 1080.4055 * (Playing U of Ames)

and then compare this forecast amount to the actual forecast to create an error value for each
game. We show these forecast predictions and error values below.
Based on this output, we see that the average error value is 0 and the standard deviation of the
error values is 495.61. The average error value of 0 indicates that our forecast model is unbiased.
The standard deviation value indicates that the forecast model is not, of course, a perfect
predictor of Magazine Sales. Based on past sales data, this suggests that our forecast values will
have an error distribution that has a standard deviation around 495.61. However, we still don’t
know the actual distribution of these errors. We could create an empirical distribution using the
observed error values in Years 1 through 9. However, we have less than 60 data points here,
which is not a lot to create an empirical distribution for errors. Other alternatives are that we can
attempt to fit a distribution to the error values or we can assume an error distribution if we have
theoretical evidence to support such an assumption. A histogram based on the observed error
values in Years 1 through 9 appears below.

Histogram of Error Values

14

12

10

0
< -1200 -1200 - 900 -900 - -600 -600 - -300 -300 - 0 0 - 300 300 - 600 600 - 900 900 - 1200

This distribution is clearly not symmetric, but again 60 observations is a fairly small sample for
distribution fitting. A normal distribution is commonly used to represent forecast model errors,
since the underlying assumption of regression models is that the errors (residuals) should be
normally distributed. Therefore, since we have relatively few observations here, a natural
assumption would be that our forecast errors are normally distributed with a mean of 0 and
standard deviation of 495.61. We also note that regardless of the distribution chosen for
magazine demand for each individual game, the total demand for Year 10 will tend toward a
normal distribution due to the central limit theorem.

This means that we can use this error distribution to create Magazine Sales distributions for each
week in Year 10 using the following steps:
1. Determine forecast for each week t in Year 10 ( Ft ), using the formula:

Ft =1617.0325 – 108.0959∗(Week ∈ Season)+ 0.0307∗(PreseasonTicket Sales)+ 44.2507∗(CSU ’ s Previous Sea

2. The Magazine Sales Distribution in Week t of Year 10 is then assumed to follow a

normal distribution with mean Ft and standard deviation 495.61.

As an illustration, we calculate the forecast value for Week 1 of Year 10 as:

F1=1617.0325 – 108.0959∗( 1 ) +0.0307∗( 54,584 ) + 44.2507∗ ( 7 ) +1080.4055∗( 0 )=3494.69

Therefore, the Magazine Sales for Week 1 of Year 10 is assumed to follow a normal distribution
with mean 3494.69 and standard deviation 495.61.

A simulation model created in Excel using 5,000 trials appears below.

The output of this simulation model shows that the total sales of football magazines in Year 10 is
estimated to have an average value of 21,432 and standard deviation of 1,302. Note that the
values will change each time the simulation model is executed due to the use of random
numbers. Therefore, student solutions will not match the values shown here exactly.

Note that it is possible to determine the distribution of total sales in Year 10 without using a
simulation model. Because the sum of random variables that are normally distributed is a random
variable that also follows a normal distribution, we know that the total sales in Year 10 will be
normally distributed with a mean of 3494.69 + 4358.90 + 3062.30 + 2954.21 + 2629.92 +
2521.83 + 2413.73 = 21,435.58 and standard deviation 495.61 √7=1,311.26 which are very
close to the values found from our simulation. The benefit of a simulation model is that this
approach can be used for any input random variables, even if they are not normally distributed.

4. Kris has noticed that of the typical 60 pages in a football magazine, 45 of those 60 pages are
the same for every game in a season. Only the 15 pages that discuss the weekly opponent
change from week-to-week. CSU’s publisher has indicated that it is possible for CSU to
order generic game-day football magazines in the July preceding the season. This generic
magazine contains the 45 pages of material that is the same for each game. Closer to the
week of each game, CSU could then tailor the generic magazine with inserts specific to that
week’s game, along with a book jacket cover displaying players and coaches from the two
teams playing that week. The number of game-specific inserts and book jacket covers can be
determined closer to the actual games in order to allow for a more accurate forecast.
 Thus, the simulation model developed in Part 3 effectively represents the sales amount for
the generic magazine, and then CSU would order the game-specific inserts and book jacket
covers much closer to the actual games when their have a much more accurate forecast of
attendance and sales. However, Kris still is not sure how many generic magazines he should
order. Should he order exactly the forecasted amount from Part 3? More? Less? Why? Based
on the cost values described from the publishing contract, if Kris orders 21,500 generic
magazines in July, what are the estimated expected costs of lost sales (football magazines
that CSU does not sell because they run out) and unsold magazines (football magazines that
CSU must send back to the publisher at the end of the season)?

The cost to CSU for a lost sale is equal to the revenue of selling a magazine ($25.00) minus the
cost of the magazine ($14.00) and minus the amount paid to the vendor ($2.50): $25.00 - $14.00
- $2.50 = $8.50. The cost to CSU for unsold magazines is equal to the cost of a magazine
($14.00) minus the amount the publisher pays CSU to buyback the magazine ($11.50): $14.00 -
$11.50 = $2.50.

Therefore, we can use the simulation model from Part 3 to calculate the Cost of Lost Sales and
Cost of Unsold Magazines by adding an order amount as our decision variable and adding two
additional output calculations in our simulation model as shown below.

The output from this simulation model shows that the average Lost Sales Cost is $4,154 and the
average Unsold Magazines Cost is $1,397 when Kris orders 21,500 magazines.

5. Assuming that CSU can tailor the specific magazines for each game in Year 10 at a later
date, what is the optimal order amount for Kris to place in July prior to Year 10 for the
generic magazines? The optimal order amount should minimize the total expected lost sales
 and unsold magazines cost in Year 10. Assume that Kris must order in batches of 500
magazines.

We can answer this question using the simulation model in Part 4 and adjusting the order amount
in increments of 500 units. Below we show the average and standard deviation of the Total Cost
values for order amounts of 20,000 through 24,000 in increments of 500.

These results show that the optimal order amount is about 22,500 units because this minimizes
the average total cost value. We also see that this is a greater than 23% improvement over
ordering the total forecasted amount of 21,500. The graph below shows the relation between the
Total Cost and the Order Amount.

Note that this part of the case is actually a newsvendor calculation. It is possible to find the
cu
optimal order amount using the newsvendor formula of , where c u is the
cu + c o
understocking cost (cost of not having enough magazines leading to lost sales) and c o is the
overstocking cost (the cost of ordering too many magazines). Thus,
cu 8.50
= =0.77 .
cu +c o 8.50+2.50
This indicates that the optimal order amount is found by calculating the value for 0.77 fractile on
a normal distribution with mean 3494.69 + 4358.90 + 3062.30 + 2954.21 + 2629.92 + 2521.83 +
2413.73 = 21,435.58 and standard deviation 495.61 √ 7=1,311.26 . Using Excel,
NORM.INV(0.77, 21,435.58, 1311.26) = 22,404.40, which is closest to 22,500 when we must
order in batches of 500.

The benefit of using a simulation model here is that we can use this approach for any input
distribution, not just a normal or other distribution with easily calculated values. A simulation
model can also be used for more complex settings that do not follow all the assumptions of a
newsvendor model.