APBRmetrics

Sorry if I missed it but is how the predictions are made revealed for the winner? Is it just voluntary posting of scripts and that is it?

You post the expected win probability for every possible match-up. Winner is determined by the smallest of amount of log-loss. Even if your ratings for a team might change throughout the tournament, you would already have submitted a win probability for the title game before the start of the tournament.

"A smaller log loss is better. Games which are not played are ignored in the scoring. Play-in games are also ignored (only the games among the final 64 teams are scored). The use of the logarithm provides extreme punishments for being both confident and wrong."

That explains how the entries are presented and scored.

How the probabilities are established was my question. But I may have answered it myself. Either the script is revealed or it isn't. Contest probably doesn't require the model to be revealed, just the probabilities.

tarrazu, i think he was asking if the methodology for the winning predictions is posted.

crow, i doubt anything too specific is posted, since people will probably want to keep their edge for the next year's contest, but i may be mistaken.
here is a a q&a with the 2014 winners(@statsbylopez and @statsinthewild on twitter) - http://blog.kaggle.com/2014/04/21/qa-wi ... -ml-mania/

Crow wrote:...
How the probabilities are established was my question. But I may have answered it myself. Either the script is revealed or it isn't. Contest probably doesn't require the model to be revealed, just the probabilities.

Contest only requires the predictions - not the algorithm or the data.

For $25k, it might be reasonable to require reveal but I guess they don't.

2014 winners used a blended model. Seems like a decent idea in this situation.

I believe that in general, Kaggle asks winners and other high-ranking teams to write a blog post describing their approach to all contests. But I don't think they require that to be anything super-specific.