BeatTheBookie goes Tensorflow

After gaining much experience over a complete season, it is time to set myself some new goals. Until now I just used or tested predictive models, which were invented or described by other people. Now I want to try something new. I would like to create my first own predictive model, which should of course provide a better performance as the current Poisson model. This is where Tensorflow comes into play.

Neural networks and especially Tensorflow as a neural network framework are the current hype topics in data science. Tensorflow already proved its abilities and is suitable to solve many problems. When searching the internet you find a large amount of use-cases. Simple picture recognition, cancer diagnosis, text translation, AI bots for computer games, nearly every problem seems to be resolvable with neural networks. So why not use it also for sports betting?

Which neural networks are suitable for sports betting?

When talking about neural networks, there is of course not only one kind of neural network, which is suited to solve any problem. There are many different types of neural networks, which have different areas of application. During my investigations I identified 2 types, which could be useable for sports betting.

Feed Forward Neural Network

FNNs are the easiest form of a neural network. The information of the input variables flows throw the network to the output nodes. The output depends on the different weights of the single neurons. The more the output differs from the expected result during the training phase, the more the weights will be adapted throw back-propagation. The classification problem of predicting a football match requires the output variables for the different results – home win, draw, away win. The network is able to recognize even complicate relations between the different input variables. So this kind of network should be usable to “just” put in variables, which correlate with the strength, the ability of a team or which describe the context of the game (e.g. motivation, table positions), and let the FNN predict the probabilities for the different outcomes.

LSTM Neural Network

LSTMs (long-short term memory) are a subcategory of RNNs (Recurrent neural networks). The special feature of these LSTMs is, that they got a long and short term memory like the human brain. They are able to build a historical context based on the data put into the network. Some factors are weighted more, some are weighted less.

A small example: If you provide the input sequence 6 -> 7 -> 8, you would like predict 9 as the next value. While for the input sequence 2 -> 4 -> 8, you would like predict 16. But for both sequences the last input was 8. The inputs before also have to influence the prediction result.

That’s why LSTMs are well suited for time series predictions or sequence classifications. And the performance trend of a football team meets a time series as I showed with the MACD analysis.

Start small – think big

Of course I am not able to simply build a complex LSTM neural network for MACD analysis of football teams. But this is my big objective, which I want to achieve step by step. I am interested, whether this is a suitable method to predict football matches.

On the way to this big objective, I have to learn and want to describe, how you are able to use Tensorflow for predictive sports betting. At the start I plan to build just a basic FNN based on the team attack & defense strength. By doing this, the accuracy of the neural network can be compared with the existing Poisson model and extended little by little.

If you have further questions, feel free to leave a comment or contact me @Mo_Nbg.

References:

[1] Kishan Maladkar, 6 Types of Artificial Neural Networks Currently Being Used in Machine Learning, 2018, https://analyticsindiamag.com/6-types-of-artificial-neural-networks-currently-being-used-in-todays-technology/

[2] Rivandra Kompella, Using LSTMs to forecaste time-series, 2018, https://towardsdatascience.com/using-lstms-to-forecast-time-series-4ab688386b1f