[joejoe19191]

Hi

I wrote a staistical model for mma predictions using machine learning tools. I cross validated this model on the last 1000 mma fights in the ufc using a very basic betting strategy - If the prediction is above 0.53 percent that a certain fighter will win, I will bet this fighter using the closing line odds at pinnacle. With this method the model is profitable and successful, predicting over 72% of fights correctly. However, this betting strategy is of course ignoring the odds themselves, which is problematic to say the least. If a fighter is a heavy favorite with the odds implying a better chance of winning than my model predicts, I still bet that fighter, despite the EV being apparently negative.

When I tried modifying the strategy so that I bet only when my prediction is above the probability implied by the odds, or even using the Kelly criterion, the performance declines to the point where it is no longer profitable. In addition I noticed that If I place the bet early on using pinnacle, I usually bet at prices worse than the closing lines (the model does not BTCL), Implying that the market does not agree with my model usually.

So... I'm unsure how to go forward. Is it reasonable using this model for real money betting using the basic strategy? Does the BTCL issue mean that the model will ultimately fail in the long run? 1000 fights making a profit seem to provide at least some statistical significance...

I will appriciate your thoughts on this matter.. I am a programmer with experience in implementing machine learning tools, but I lack statistics knowledge and betting know-how.

Thanks in advance,
Joe

[jizay]

When I tried modifying the strategy so that I bet only when my prediction is above the probability implied by the odds, or even using the Kelly criterion, the performance declines to the point where it is no longer profitable. In addition I noticed that If I place the bet early on using pinnacle, I usually bet at prices worse than the closing lines (the model does not BTCL), Implying that the market does not agree with my model usually.

It sounds like you learned that 1) the market is efficient, and 2) the lines incorporate new information close to fight time. What am I missing?

[joejoe19191]

As for the BTCL issue, I actually haven't made a statistics of how often I do beat it. So im not 100% sure that Idont beat it. But as for the market efficiency, IF I haven't done any mistakes than the market isn't very efficient, since I am apparently able to make a profit against pinnacle closing lines.

Also, I cant see how your post answers any of my questions..

[NunyaBidness]

As for the BTCL issue, I actually haven't made a statistics of how often I do beat it. So im not 100% sure that Idont beat it. But as for the market efficiency, IF I haven't done any mistakes than the market isn't very efficient, since I am apparently able to make a profit against pinnacle closing lines.

Also, I cant see how your post answers any of my questions..

The MMA Market isn't close to efficient whatsoever. There are specific subsets of lines that are even more inefficient.

I've done some limited modeling work with MMA and found it to be lacking compared to what I can do with 'feel' capping, which if anyone said that to me about any other sport I'd think they were a moron.

You can look at something like cagerank which is already modeling based on fightmetric data and see that it often comes up with ridiculous conclusions.

[Gradius]

As for the BTCL issue, I actually haven't made a statistics of how often I do beat it. So im not 100% sure that Idont beat it. But as for the market efficiency, IF I haven't done any mistakes than the market isn't very efficient, since I am apparently able to make a profit against pinnacle closing lines.

Also, I cant see how your post answers any of my questions..

1000 fights is a pretty large sample size. I'm not sure about MMA, but for the major sports, the rule of thumb is that if you can beat the closing line at about 60+% of the time, you should be able to profit in the long run. I would try going back and counting the plays where you beat the closing line, over your total number of trials.

Also, try using a Z-Score to see how many stand deviations you are. Anything above 2 might make it something to look into.

The MMA Market isn't close to efficient whatsoever. There are specific subsets of lines that are even more inefficient.

I've done some limited modeling work with MMA and found it to be lacking compared to what I can do with 'feel' capping, which if anyone said that to me about any other sport I'd think they were a moron.

You can look at something like cagerank which is already modeling based on fightmetric data and see that it often comes up with ridiculous conclusions.

Agreed, and this applies to other smaller sports as well (which MMA technically is).

I think in the future, there will be professionals that use math models that will better predict the "feel" part of the decision-making (similar to what Haralabos Voulgaris and Billy Walters have done), and it wouldn't surprise me if there was someone already doing that right now, with MMA (and profiting from it).

[TPowell]

would love to hear about the factors you are using in your MMA models. PM me if you wouldn't mind.

[Inkwell77]

I'm not sure what exactly the definition is of an efficient market, but for a while I did think the UFC market was not efficient.
The main reason: Blindly betting favs was profitable
But the past 6 months have reversed and now one would be down money blindly betting favs since the start of 2012.

I also know that following line moves in the UFC is not as profitable as it is in certain other sports.

So I don't think the UFC market is necessarily "efficient," but I do believe it is as efficient the other smaller sports/leagues.

There is a crazy subset in the UFC though involving the first 3 fights of each card. And I definitely believe that the first 3 fights on the typical UFC card are far from efficient.

[NunyaBidness]

There is a crazy subset in the UFC though involving the first 3 fights of each card. And I definitely believe that the first 3 fights on the typical UFC card are far from efficient.

That subset can be labeled differently and be more meaningful.