|
|
Real-Time Prediction of Football Matches Results
Drankou, Aliaksandr ; Bartík, Vladimír (oponent) ; Hynek, Jiří (vedoucí práce)
This thesis studies the problem of real-time football matches results prediction. It consists of several steps, including the acquisition of suitable dataset and training of the prediction model. The prediction model is represented by two types of neural networks: feedforward and LSTM recurrent neural network. Different combinations of input features are tested to achieve the best performing model. Both models achieved a classification accuracy of about 67.5%, where feedforward network accuracy starts from 54% at the beginning of the match and achieve 93.54% by the end of the match. In addition to widely-used metrics such as categorical accuracy and log-loss, each model is evaluated in the simulated betting environment.Experiments within betting evaluation have shown that LSTM can't compete with feedforward network, as in each betting run LSTM network ended up with a balance, dropped by more than 90%. However, the feedforward network achieved an ROI (return on investment) of 0.39% in a betting simulation run with one of the configurations. As a result, a neural network approach, especially the feedforward network, has proved to be quite successful in terms of predicting real-time football matches results. Moreover it allowed to build a profitable betting strategy upon it.
|
|
|
Real-Time Prediction of Football Matches Results
Drankou, Aliaksandr ; Bartík, Vladimír (oponent) ; Hynek, Jiří (vedoucí práce)
This thesis studies the problem of real-time football matches results prediction. It consists of several steps, including the acquisition of suitable dataset and training of the prediction model. The prediction model is represented by two types of neural networks: feedforward and LSTM recurrent neural network. Different combinations of input features are tested to achieve the best performing model. Both models achieved a classification accuracy of about 67.5%, where feedforward network accuracy starts from 54% at the beginning of the match and achieve 93.54% by the end of the match. In addition to widely-used metrics such as categorical accuracy and log-loss, each model is evaluated in the simulated betting environment.Experiments within betting evaluation have shown that LSTM can't compete with feedforward network, as in each betting run LSTM network ended up with a balance, dropped by more than 90%. However, the feedforward network achieved an ROI (return on investment) of 0.39% in a betting simulation run with one of the configurations. As a result, a neural network approach, especially the feedforward network, has proved to be quite successful in terms of predicting real-time football matches results. Moreover it allowed to build a profitable betting strategy upon it.
|
host ::
RSS.