Deep learning and sports forecasting

I’m a software developer and I’m just getting started on ML with this course.

As advised by @rachel, I decided to have a project in mind from the beginning. my goal is to make models that can predict the outcomes of sports matches.

I know there has been some research done here, mainly in the form of ranking systems, like Elo, LRMC , or a combination of many of those.
However, none of these approaches provides a way to consistently make a profit on the betting market.

So, I’m very interested in knowing if someone sees potential on deep learning algorithms making unprecedented performance on this kind of problems.

It’s definitely doable but I think the issue you’re going to have is what data you need to be looking at and how you’re going to consistently collect it.

I’d be surprised if the bookies setting odds aren’t using some form of machine learning to do so so you’re looking at trying to create a model that’s more accurate than theirs. I think you’d need deep pockets as well as the variance inherent in sports can catch you out.

But it’s an interesting problem domain. I’m not sure where live sports data resides but i’m betting it’s out there.

With the right data and a deep neural net you can do anything… :wink:

Thanks, that sounds exciting.

What I want to do is use an algorithm to discover patterns in the games. For example, it could discover that team A always loses against teams that have scored more than 40 times per season. So although team A has a very good ranking, we can see that its defense is not enough against very strong offensives.

So, what do you mean the right data? Are you referring to quality (like data with detailed info on each game and play, to be able to find more interesting patterns) or quantity? If its quantity, we have to take into account that data from many years ago, will not necessarily be relevant.

I think any basic patterns will be well known to the oddsmakers.

You’ll most likely need to come up with additional data they’re not looking at to get a real edge. That could be everything from individual player performance per game, injury information, weather info, news, weather, etc.

My guess is that without additional data you’ll be hard pressed to come up with anything that beats the odds.

Which isn’t to say you can’t find these patterns you’re talking about and I think you should explore the project. Start simple and then expand the data you’re using. Just don’t expect to make money with it right away.