Forecasting Gate Receipts Using Neural Networks And Rough Sets

Forecasting Gate Receipts Using
Neural Networks And Rough Sets




Ramesh Sharda

Oklahoma State University

Edith Meany

Henry Amato

University of Nevada
-

Reno

Niketu Mithani

Oklahoma State University

Contact e
-
mail: sharda@okstate.edu

Forecasting Gate Receipts:

A Tough Problem

Previous Work


-

Litman (1997)



-

Need Marketing Estimates


-

Sawhney & Eliashberg (1996)



-

Need initial audience data

A Different Approach

•
Treat the problem as a classification
problem


•
Neural Networks and Rough Sets are
appropriate for classification

Neural Network Description

•
Inputs:



Estimated release date (season)



Intensity of competition rating



Rating





Star power



Genre



Technical Effects



Sequel ?



Estimated Screens at opening

•
Output:



Box office gross receipts:


flop blockbuster


Description


Neural Networks


Test 1

•
43 Input neurons

•
9 Output neurons

•
2 hidden layers (18 and 16 neurons)

•
tanh transfer function

•
Backpropagation


Rough Sets

•
Automatic configuration by H. Amato

Data

•
120 films for 1997


•
Twenty pairs of 20 films for training and
testing


Results




Results





Further

Tests


Data

•

1998
-

255 films used

•

Predictors added
-

1 additional genre of


romance/musical

•

Predictors removed
-

perceived competition, star


quality and technical effects

•

Film budget




Test 2



Twenty pairs of training(155) and test(100)


films.




Results






Further

Tests


Test 3



Twenty pairs of training(155) and test(100) films


with no hidden layers used.



Test 4


Twenty pairs of training(80) and test(64) films used


with the budget information included.










Results




Results















Issues

•
Performance Measurement

•
Misclassification Costs

•
Comparison with other models/techniques

Conclusion

•
Neural networks offer a reasonable
predictive capability

•
Rough sets offer a comparative performance

•
Performance measurement issues

•
Inclusion of misclassification costs

•
Comparison with other models/techniques