Predict NBA games, make money — machine learning project

Predict NBA games, make money — machine learning project

Past performance doesn’t indicate future performance… unless…we can get 136% returns….? :wink:

The bettor and the bookie don’t have much in common — one could describe their relationship as a rivalry, a duel, a war. But in their sleep, they drool over the same fantasy: a model of historical performance so perfect it predicts the outcome of future games with pinpoint precision . With deep learning, this might be possible — or at least easier than previous data science techniques.

I read a lot of good and bad journal articles to see if this was possible and here are the good ones:

The underlying assumption is that NBA markets are inefficient (the price or betting line does not reflect all available information for the game) — and maybe more inefficient than most markets because of the bias of hardcore fans to just bet on their favourite team. If you can bet against the inefficiencies of the market, you can make money. One of the ways we could identify inefficiencies is by analysing the cold hard numbers.

Although many models that attempt this challenge are accurate, most are nowhere near close to making a profit. The reason for this is simple: the bookies are extremely accurate as well. Even if you match the bookie accuracy, you will lose because of the 5% betting fee/house edge.

The graph on the left is bet365 money line vs actual win percentage. A successful model will have to predict those minute fluctuations of the bookie from the perfect regression.

My model, built in Python with Tensorflow, analyses the past 11 NBA seasons and in many ways, is similar to every other deep learning model that attempted this problem with one crucial difference — it uses a custom loss function to decorrelate from the bookies. We are trying to pick the games the bookie has misgudged the actual percentage.