Human Prediction vs. Algorithm Prediction
Predictions are probability statements that are conditioned on all of the information one can obtain.
Do you think you can predict a UFC fight better than a model?
Well, I thought I could predict fights better than my machine learning models. I wanted to give the model and me a fair chance, so I decided to do 50 fights that range between 7 UFC fight weeks. I would make my predictions before the model to ensure that I was not biased towards my human prediction by getting more information than my model would have obtained. I want you to know that the model I use is about
58% percent accurate on all fights. It is about 65% correct on all men fighters above the flyweight division. It is about 43% on predicting any weight class below guys flyweight and all women divisions. I wanted to give you a little background on my model before giving out the sample results I produced over the last seven weeks.
Here is just a little background information on the typical discussion of a human vs. modeler.
I hear a lot from the MMA community say “models are too inconsistent.”
I decided to at least put myself to the test against my machine learning model. This predicts every single fight on the card that the model predicts, not just the fights I think are good bets. However, I think the results would be pretty similar, but I can always try another experiment!
So here are the results of the weeks!
As you can see, I did really well initially, but as time progressed, I started to do worse, and my model stayed consistent and got pretty close back to its usual accuracy, even with a small sample of 50 fights! If I go up against anybody, my model will eventually be at its consistent 58%, so if you can be better than 58% over a long period, then hats off to you! However, models will be the more consistent prediction over the long term than a human predicting the fights.
See, the big thing is human prediction has much variance, but yes, in small sample sizes like this, it is possible for a human to predict better than a model. However, if I had a bad day or a dull day where it affected me mentally, it could have been a factor in my predictions because I wasn’t thinking straight. However, if my model has an off night, it will still consistently predict the same way to eventually get to 58% accuracy. Again, a model will be more consistent over the long run than a human, and if someone tells you that is incorrect, they do not know probability and statistics very well!
– Brandon, a member of the beforethecagedoorcloses team
Let us know if you are interested in compiled UFC data.