BPFNN Based Football Player Value Prediction by Introducing Historic Information

Abstract

With extensively successful applications of artificial neural networks in many scenarios of classification and prediction, research related to the world most popular sport - football - has undoubtedly attracted researchers to carry out their works with aid of such powerful tools. Despite that most of these works are oriented to predict match results and game strategies, only very a few are designed for player value prediction, which is of great importance with the rapid growth of the football players’ transfer market. In this paper, we propose a BP feedforward neural network framework which integrates both the empirically determined variables and some historical information. Comparative ablation experiments on one dataset FINAL verified the necessity of some factors especially the highest_values which inspired the inclusion of available historical information into the framework. For further verifying the efficiency of integrating historical information, we embedded the value columns of FIFA17 to FIFA21 into the data of FIFA22. Comparative experiments showed an 11% improvement on the precision with the help of historical values, verifying the efficiency of this strategy. Experiments were also conducted for determining the choice of the number of layers, the ratio of the training set and test set, the learning rate, and the process of the missed information.