We developed a pre match, machine learning prediction algorithm which scored a 63.4% accuracy rate in predicting the standard WIN (W), DRAW(D) and LOSS(L) outcomes of English Premier League (EPL) matches in general over the… · More 2016/2017 EPL season.
We addressed the problem of deciphering the standard W/D/L outcomes in the EPL by tackling the problem as a multi-label classification one. Via the agency of a comprehensive team and player training data set & feature vectors, we successively trained a modified support vector machine (SVM) algorithm before deploying it on line-up information roughly an hour preceeding to each match kickoff.
We understand that the different team combinations of 11 mean different winning probabilities. It is therefore no surprise that all teams deploy their best 11 possible players (10 outfield + 1 goalkeeper) each match day. Out of the standard 25-man EPL squad size, only 10 outfield players + 1 goal keeper can begin on the pitch each game.
Given the huge potential player combinations that can be generated out of each 25-man squad size (Outfield players: 22C11=705,432 combinations; Goalkeeper position: 3C1=3; Total of 705,435 combination), it makes less sense to attempt to predict matches before team line-ups are announced(approx. an hour before the game) given the inability to account for stochastic factors/occurences.
Various stochastic factors such as player fitness levels/injuries, disciplinary problems, coach decisions, unexpected events etc affect the certainty of each successive match day line up so we endeavoured to make predictions as soon as the line-ups where announced an hour before the game.
We automated our system such that we relied on instant line-up information from the Twitter API so we essentially had the algorithm automatically make predictions without having to manually enter in the player line-up details.
We are currently focused on error optimization, improving our feature vector set and ensuring the algorithm has the best possible fit to establish an appreciable percentage point difference (+20PP-->70%+ accuracy rate) compared to the mainstream betting industry (averaging ~50% accuracy rate).
Using the EPL as a paradigm (arguably due to its massive global appeal, availability of comprehensive data and higher reputability/officiating levels), we envisage standardizing the pre-match algorithm and applying it to other leagues such as La Liga, Serie A, Bundesliga, Ligue 1 etc).
Our pre-match algorithm will be of greater utility to;
1. Professional sports bettors 2. Quantitative traders 3. Sports pundits 4. News media agencies