Software Predicts Movie Success

scheming daemons writes "TechNewsWorld has an article about software that predicts whether a movie will be successful or not by factoring in its rating by censors (e.g. G, PG, R), strength of the cast, genre, competition from other films at the time of release, special effects, whether it is a sequel, and the number of theaters in which it will show."

Comments

Re:Hollywood has used this formula for years:

[grub]

More explosions, chases and boobs done with a *smaller budget* == more profit. :)

Re:The singularity is near...

[qubex]

I expect it's less of an AI and more of a simple collection of linear statistical models (linear regressions and general linear models) using parameters gleaned from the performance of past films.

Unless yours was a reference to the awful film A.I.

Re:Program??

[erikus]

actually, they did use excel. google cache

Re:More data than they need

[b0rk b0rk b0rk]

The article says as much:

On the other hand, predicting the potential success of a film based on a set of factors is what movie studio executives are paid to do, Peterson acknowledged. It may be interesting to automate the process, he said, but the software would only support industry behavior that already exists. "Film executives look at things like star power, film release dates, target audiences and soundtracks," Peterson said. "This software is going to give you information that is probably already known."

citation + main (unimpressive) results

[dankelley]

The citation for the research paper is Ramesh Sharda and Dursun Delen, Predicting box-office success of motion pictures with neural networks, Expert Systems with Applications, Volume 30, Issue 2, February 2006, Pages 243-254. (http://www.sciencedirect.com/science/article/B6V0 3-4GV2PCH-1/2/35524bc2ff6fd852c98d8c9f3c3dc8c9). This is not a free journal, but if you're at a university it is quite likely that you have a library subscription. The paper is an interesting read, whether you're keen on film or on neural nets.

The main result is that the method (neural net) works a little better than other methods on the same data (Table 4 of paper). It scores 75% in a test; conventional regression scores 71%. As they say in the statistical literature, "big woop"; the fancy new thing is marginally better than the simple old thing.

As for the practical side of things, the main predictive variable is the number of screens on which the film was initially shown. The next-highest predictive variables are a variable representing the use of technical effects and a variable represengint the actors' reputation. Well, none of these indicates that this tool (or others discussed in the paper) is of any real use to the industry. The suggested use of the tool is to predict movie success. But the main predictive variables all represent things the industry already knew, when the film was being made and promoted. It's like asking a patient if they have a cold, and then charging them to tell them they have a cold.

Classifier and statistics

[HuguesT]

If this were 1990, the title would read "neural network predicts movie success" and the discussion would be about the impending success of strong AI.

Reading TFA, it's impossible to know whether this study has any value without seing a proper article, as submited to a reputable stats journal.

First of all this sounds like simple statistical classification with pretty obvious variables. However making classification work is not always trivial.

Methodology is the key here. The sample of 800 movies is rather small, and the details on the chosen explanatory variables is sketchy. With enough variables, even meaningless ones, one can explain anything on a training sample. However with proper classification techniques, using for example jacknife/resubstitution/cross-validation one can find out if the classification model has any actual predictive values.

As someone said "anybody can predict the past", and someone else "prediction is rather difficult, especially about the future".