Beating the bookmaker: rating models and social media
On October 14 another seminar of our Research Group "Machine Learning and Social Computing" was held. Andrey Shelopugin together Alexander Sirotkin presented their work on: "Can I win at the bookmaker: rating models and social media."
First, Andrey talked about rating systems and his topic - improving algorithms for predicting the results of basketball games. Basketball is a team sport, which helped to choose the Glicko-2 algorithm as the basis of the model. Andrey told about the heuristic improvements made to the model: taking in to account the injuries of players and the transitions between the teams, the motivation of the players, the score of the match, and told how each of the improvements is expressed in a probabilistic form. Despite the fact that there were few fans of ordinary sports among listeners, the discussion was still stormy: the idea how to beat the bookmakers in forecasting is amazing.
The main part of the report included a new stage of their study where "honest" forecast based on the model, the coefficients of real bookmakers and online game discussion data were reduced to one model. Strategies for lowering the coefficients become visible through comparison with the "honest" forecast and the activity of the discussions online.
This is a promising area for social algorithms, where classical algorithmic models and Internet data help to make some conclusions about mass behavior, and the first results of Andrey's research confirm this: the mathematical model includes such complex concepts as "interesting game," "strong team," "unexpected result ".