Football is, without a doubt, the most followed sport branch. Thousands of matches are played each year in many organizations such as national leagues, world cups, and continental cups. There is also a financial aspect of football. Many football fans do not only support their favorite teams in stadiums but they also bet on football teams to make money. There are many betting companies around the world and there is a serious amount of money circulating every day. This makes predicting match results crucial. The idea is based on the simple assumption that the betting companies do deep analysis to determine winning or drawing ratios for each match. Otherwise, they could not make any profit. Our task is to find a pattern between the results and the ratios using artificial intelligence.
Many artificial intelligence methods have been developed since the middle of the last century. Support Vector Machines (SVM), Decision Trees, K-nearest Neighbors, and Deep Neural Networks (DNN) are the most popular algorithms. These algorithms are vastly used in science and industry. DNN recently got more attention due to its power to solve challenging tasks. It grows stronger every day parallel to the advancements in hardware technology. DNN is a great tool to recognize patterns in a given dataset. It is highly suitable for our task because what we expect from the algorithm is to find a correlation between the given ratios and the match results. The database consists of approximately 70000 match results and corresponding ratios by the result of years of web-scrapping. The deep neural network consists of 7 layers. Its architecture is shown below.
ReLu is used as an activation function for each layer and, for optimizer stochastic gradient descent with momentum is used. The model predicts with 55% accuracy of given matches. It is way more accurate than blindly guessing a result, which makes this predictor successful. The model can be extended by including more features such as performance statistics of football players, the current standing of teams, etc. However, it is not easy to collect all these data. To support improving the model and my future ideas, click here.