The New Yorker has an article by Malcolm Gladwell on “the formula” to figure out which films become hits:
The way the neural network thinks is not that different from the way a Hollywood executive thinks: if you pitch a movie to a studio, the executive uses an ad-hoc algorithmperfected through years of trial and errorto put a value on all the components in the story. Neural networks, though, can handle problems that have a great many variables, and they never play favoriteswhich means (at least in theory) that as long as you can give the neural network the same range of information that a human decision-maker has, it ought to come out ahead. Thats what the University of Arizona computer scientist Hsinchun Chen demonstrated ten years ago, when he built a neural network to predict winners at the dog track. Chen used the ten variables that greyhound experts told him they used in making their betslike fastest time and winning percentage and results for the past seven racesand trained his system with the results of two hundred races. Then he went to the greyhound track in Tucson and challenged three dog-racing handicappers to a contest. Everyone picked winners in a hundred races, at a modest two dollars a bet. The experts lost $71.40, $61.20, and $70.20, respectively. Chen won $124.80. It wasnt close, and one of the main reasons was the special interest the neural network showed in something called race grade: greyhounds are moved up and down through a number of divisions, according to their ability, and dogs have a big edge when theyve just been bumped down a level and a big handicap when theyve just been bumped up. The experts know race grade exists, but they dont weight it sufficiently, Chen said. They are all looking at win percentage, place percentage, or thinking about the dogs times.