FORECASTING GATE RECEIPTS USING NEURAL NETWORK AND ROUGH SETS

glibdoadingAI and Robotics

Oct 20, 2013 (3 years and 10 months ago)

87 views

FORECASTING GATE RECEIPTS USING NEURAL NETWORK AND ROUGH
SETS


Ramesh Sharda, College of Business Administration,

Oklahoma State University, Stillwater, OK 74078

Henry Amato, College of Business Administration,

University of Nevada
-
Reno, Reno, NV 89557

Edith Meany, 408 S. Randall Ave., Madison, WI 53715
-
1652




ABSTRACT


We report on the performance of neural networks
and rough sets in forecasting the performance of
a movie at the box office before it is released. A
movie’s performance is classified in
to one of the
nine categories: flop to blockbuster. The
independent variable used by us can be
determined prior to a movie’s release. Our
results indicate that both approaches achieve
about the same accuracy., but much better than
possible with earlier a
pproaches.


INTRODUCTION


Forecasting the box
-
office results of a particular
feature film has intrigued many scholars as a
difficult and challenging problem. In the past,
researchers concentrated on statistics
-
based
forecasting approaches. Litman summari
zes the
major studies on film box
-
office result
forecasting in his recently published book, “The
Motion Picture Mega
-
Industry” (Litman, 1997).
Studies described in his book are similar in their
statistical nature as well as in their parallelism in
using m
uch of the same independent variables.


A relatively recent study (Sawhney and
Eliashberg 1996) examined 114 films from 1992.
It used many of the same independent variables
applied in the earlier research papers, with the
addition of the size of the audie
nce of a given
film during each week of its release period. The
diffusion model proposed by Sawhney &
Eliashburg resulted in an absolute error of
71.1% before the film’s release to an absolute
error of 1.8 % after the film has been out for
weeks. These
previous studies leave us with a
still unanswered need for a
more accurate film
box
-
office result forecasting method, which,
unlike the previous approaches, could be
accomplished in advance of the film’s theatrical
release.


Our work has explored the use o
f neural
networks and related techniques in forecasting
the performance of a movie at the box office
before it is released. We apply our models
converting the forecasting problem into a
classification problem. Rather than forecasting
the exact amount of
box office receipts, we
classify a movie in one of nine categories ranging
from a “flop” to a “blockbuster.” This paper
reports the comparative results of two
techniques: neural networks and rough sets.


Neural networks have been extensively used as
nonli
near mapping techniques for statistical
modeling. Applications of neural networks have
been reported in many diverse areas and
represent prediction, classification, and
clustering application. The literature is so large
that no one or two references will

suffice. A
classic reference for the concept of neural
networks is Rumelhart and McClelland (1986).
A bibliography of NN application was presented
in Sharda (1994).


Since the concept of rough sets may be
unfamiliar to our readers, we provide a brief
re
view here. Pawlak 1982 introduced the concept
of a rough set in 1982. Rough sets is a rule
-
extraction method for producing a decision table
of if
-
then rules from an observable set of
conditions (inputs) and corresponding decisions
(outputs). Both inputs an
d outputs are
represented by categorical variables with a
limited number of values possible for each
variable. The method basically measures the
degree (discernability) to which the mapping
between the inputs and outputs is consistent.
From that measuremen
t technique the set of
inputs required to maintain the overall
discernible measure is reduced to a minimum set
of inputs (reduct). This set need not be unique.
To a reduct of a data set a rule
-
generation
algorithm is applied. The algorithm generates
rules
based on structural relationships in the data.
This approach is different than curve fitting
methods, such as discriminant analysis or the
backpropagation paradigm for neural networks,
which produces a set of weights used to map the
inputs into the outputs

in a manner which
minimizes the mean squared error between the
observed outputs and the predicted outputs.
[Pawalak & Slowinski 1994, Slowinski 1992.]


To use a roughset it is necessary to first scale all
inputs and outputs into discrete classes. For each

variable the range of values is decomposed into
discrete class if they are not already in that form.
The decomposition rule is an individual choice
but should reflect the precision needed in the
forecast. The purpose of making the data less
detailed is to

better observe the patterns in the
data. Too much detail can obscure the pattern. If
there are any inconsistencies in the data the
rough set analysis will detect them. Sometimes it
is possible to remove the inconsistencies and still
retain the accuracy of

the forecast by effectively
decomposing the data.


METHOD


Our models were built and tested using 1997
data. From 738 films that were made in 1997, a
random sample of 120 films was selected. The
following data were collected, computed, or
generated for
each film: Date of release, Rating,
Intensity of Competition, Star power, Genre,
Technical effects, Sequel, Screens at opening,
and the box office gross. Unfortunately, the
production and marketing budgets of the 1997
films were not available, which, keep
ing in mind
that the film industry’s environment is extremely
competitive, is not surprising at all. In the data
collection process, the database of 1997 films,
provided by ShowBIZData, Inc. was also used.


The majority of the classifications were
convert
ed into binary variables. The
independent variable set therefore consists of 43
independent variables. The output variable
similarly was assigned into ranges. We converted
the box office receipts into nine categories
representing nine binary output variab
les.


The model was tested by building several
training and test files. One hundred films
selected at random for a train file and the other
twenty films selected for a test file. A specific
model is trained and the trained model is applied
to a test file

containing data on twenty other
films..


The neural network model was implemented by
using Neuralware Professional Plus software.
We employ 2 hidden layers, with 18 and 16
processing elements in them, respectively. The
training algorithm is based on the

back
propagation network. A tan h transfer function
was used. Given the description of the variables
above, we had 43 input neurons and 9 output
neurons.


The roughsets model was implemented by the
second author as an add
-
in to MS Excel
spreadsheet.


RE
SULTS AND DISCUSSION


The results presented by us are not directly
comparable to earlier work because our model
predict a class of performance in which the
movie should belong, rather than an actual dollar
figure. A general sense of accuracy is, however,
still relevant. Recall that the model aims to
categorize a film in one of the nine categories.
Accuracy is measured in two ways. The first
metric is the percent correct classification rate.
This measure is a reasonably conservable metric
for accuracy in

this scenario. In reality, the
movie studios might be glad to predict within one
or two categories on either side. We assign an
80% weight to correct prediction within one
category and 60% weight to correct prediction
with in two classes. This is repor
ted as the
Bestpred, Acc 1and Acc 2 respectively.


Table 1 presents the results of the performance
comparisons of a NN model and a rough sets
model. The Bestpred variable denotes the percent
of time when a movie was classified in the
appropriate category.
The Acc1 and Acc2
measures allow for a weighted performance as
described earlier. On an average, neural networks
and rough sets are able to correctly classify over
a third of the test cases with 100% accuracy.
However, the accuracy rate jumps to 70% when
p
rediction within 1 or 2 classes is considered
acceptable. If classification within one category
is considered acceptable the NN model has a
slight edge on accuracy (61.55%) over the RS
model(53.85%). (Again a classification within
one category is considere
d only 80% correct and
in two categories is considered only 60% correct)
It is interesting to note that both rough sets and
neural networks have similar performance in
specific trials. The correlation coefficient of
Bestpred is 0.89.Though our numbers ar
e not
directly comparable to earlier studies because our
model classifies a movie into a category rather
than predict a point estimate of the box office
receipts, this performance is quite good.The
accuracy of prerelease forecasting reported by
Sawhney an
d Eliashberg is reported as a mean
absolute percent error of over 71%. Further,
work is in process to fine
-
tune the models and to
further test the performance in a more complete
experimental design. In any event, these results
are intriguing and point to
the continued value of
neural networks in difficult forecasting problems.


Beyond the attractive accuracy results, these
models could also be adapted for other media
products. The particular parameters within the
testing files of the particular films or o
ther media
products could be altered for testing purposes.
During this alternative testing, the managers of
the given entertainment firm could find out, with
a fairly high accuracy level, how much a specific
actor, a specific release date, or the addition
of
more technical effect, could mean in the financial
success of a film, or television program.

REFERENCES


Eliashbeg, J & M.Sawhney, "A Parsimonious
Model for Forecasting gross Box
-
Office
revenues of Motion Picture",
Management
Science
, Vol 15, No. 2,
pp.113
-
131,1996

Litman , B, "
The Motion Picture Mega
Industry
", 1998

Pawlak, Z., "Rough sets",
International Journal
of Information & Computer Sciences
1, 341
-
356, 1982

Pawlak, Z., and Slowinski, R., "Rough set
approach to multi
-
attribute decision analy
sis",
European Journal of Operational Research
72
443


459,1994


Rumelhart, D.E., and J.L. McCelland,

Parallel Distributed Proceedings
, Vol. 1, MIT
Press, 1986

Sharda, R, "Neural Networks for the OR/MS
Analyst: An Application Bibliography,
"
Interface
s
, Vol 24,No. 2, pp 116
-
130,Mar
-
Apr
1994

Slowinski, R., (ed.),
Intelligent Decision
Support. Handbook of Applications and
Advances of the Rough Sets Theory,
Kluwer
Academic Publishers, Dordrecht, 1992






TABLE 1: Comparitive Performance: Percent
Classification Rate



Neural

Networks


Rough

Sets


Run. No.

BestPred

Acc1

Acc2

BestPred

Acc1

Acc2

1

30%

62%

71%

30%

54%

69%

2

90%

90%

93%

90%

94%

94%

3

85%

85%

91%

75%

75%

84%

4

25%

57%

69%

25%

37%

58%

5

75%

83%

86%

70%

86%

86%

6

30%

66%

72%

10%

3
0%

48%

7

85%

93%

93%

85%

89%

92%

8

75%

87%

90%

70%

78%

84%

9

30%

70%

79%

15%

59%

68%

10

15%

59%

71%

10%

38%

62%

11

5%

29%

53%

25%

53%

65%

12

15%

59%

65%

25%

45%

57%

13

5%

45%

51%

25%

49%

58%

14

20%

48%

60%

20%

52%

61%

15

25%

49%

64%

25%

37%

49%

1
6

50%

70%

76%

10%

46%

64%

17

20%

44%

65%

10%

38%

50%

18

25%

49%

64%

15%

35%

62%

19

15%

35%

53%

10%

30%

51%

20

15%

51%

60%

20%

52%

58%

Mean

36.75%

61.55%

71.30%

33.25%

53.85%

66.00%

StdDev

28.67

18.65

13.58

27.54

20.10

14.42

Median

25%

59%

70%

25%

51
%

62%