Transcript Motion Picture Revenue Prediction

Motion Picture Revenue Prediction
An Artificial Neural Network Method
for Predicting Opening Weekend
Box-Office Revenue
ECE 539 – Fall 2001
Final Project
Chad M. Steighner
ID: 253-699-5562
Concept
 Use opening weekend revenues from 1989 through
2000 to train a MLP with back-propagation for
classification into 5 classes.
 Input features include:

Genre, MPAA rating, Date, # Screens, Critical Rating,
Distributor, Run-time and Weekend Length
 The MLP is then tested with motion picture data
from 2001.
Implementation
 Found 473 films with
opening weekend data
(www.boxofficeguru.com)
 Used www.imdb.com to
obtain add’l fields
 Created Parsedata.java to
construct TrainingData and
TestingData for MLP
•
•
432 Training (1989 – 2000)
41 Testing (2001)
 Through testing and 3-way
cross validation found the
best set-up to be:





3 Layer MLP (1 hidden)
6 hidden neurons
Learning rate = 0.1
Momentum = 0.9
1000 Epochs
Results
 45.6% avg. class. of 2001 films. (5.87 St.dev)

$10-$12M, $12-$14M, $14-$17M, $17-$28M, $28M+
 No Exact Replica Baseline Study:

Nat’l Research Group (LA)



- telephone surveys to within 5% of opening weekend revenue.
Moviefone claims to be even closer (movie info website)
Prof. Arthur De Vany (UC-Irvine)


Bose-Einstein distribution of particles falling into urns.
Equally likely particles (audience) will fall into a few urns
(movies) as it is for them to be distributed in any other way.