I have predictions from my machine learning model. So far this season it has a 87% strike rate on all games and markets.
Looing to know what everyone thinks about them!
These are my model's 4 predictions for SuperBowl LV.. I know sports betting in the US is legal for the first time for many. Although I do not bet and this is a purely academic project I hope this provides you with extra data. Coming from a Quant background in finance, creating models for securities, I have to say American football lends its nicely to analysis unlike other sports. The play by play nature make patterns very obvious edges quite easy to take advantage of.
This board is surprisingly dead for what is usually a very active time in. very interesting data thank you for the post. were u using this model all season ? also have u applied ur quant to other things such as the stock market or other sports ? very interesting subject please post more !
I've been on the chatrooms too long to not know if that's sarcasm, so i'll just assume you are serious.
If you are thank you for such a refreshing response. I worked for serveral years in investment banking basically as a quant working with securities, and programing models for high freq trading. I consider myself to be more inteserested in the mathmatics behind this. I don't really bet but if I did it would be on football. The play by play nature makes it very easy to analyse and and edge is gained fairly easily.
My machine learning model Automatically scalps any sports news sites for weather, injuries etc. and decides whether it will effect an of the other variables. It is a supervised model meaning I had to programme it first however now it corrects itself without me, and gets better with each game.
This was an amazing season 87% strike rate on all markets (taking into account variance). I am very keen to see where I can take this perhaps build a group for predicitve sports analytics.
Thanks again for taking an interest
I have to say i've only ever used GPT-3 for voice recognition and similar tasks and wouldn't know how it would work for linear regression for predicting outcomes data in sports. This space changes very quickly in finance not so much in sports analytics so it's fairly easy to make a good model with primitive software. I find the limiting variable is easily Data, that being said next-gen stats aws and others have made football and basketball to a certain extent the best sports in the world to analyse.
Yeah I didn't want to go into details initally as most people just want the sexy headline. Being mathematically minded, I also would never take that strike rate at face value (you'll be surprised how many do, especially in finance). I have a SQL dataset with the outcomes and bets that would have been made from them based on the line Bet365 are offering. ALL these bets are against the spread and NONE are moneyline (that would be too easy). I personally don't bet so I didn't wager any money per se (I will be having a cheeky one on the superbowl, just to stand by my predictions). There were approximately 3 wagers per game and most games got a bet for all of this season. 624 bets (154 on the spread, 240 on passing yards, 134 under/overs and the rest miscellaneous). These were always a single bet (NEVER parlays etc.)
Do I expect a similar result again? no, it's unheard of. However with the right money management and no emotions It could prove very profitable.
I am making my next YouTube video on exactly how I train my models, how I get data and the parameters used. I hope this answer your questions, thank you for taking an interest. I would love to know if you are training any models currently, and how successful are they. Whihc parameters do you find to be most successful?, Cheers
Chris - Thanks for the response.
I have been doing this a long time. Back to the 90s where it was a struggle to maintain enough computing power in databases to chug through calculations - and where it was very difficult to even get stats to populate a database. In the 2000s, I employed a team to help me with it - just because it got bigger and more comprehensive and I was spending too much time doing it and the computer tech passed me by. I feed my guys queries and they feed me results.
The source database I put the most weight in are adjusted stats for strength of schedule, with smoothing for recency. Pace, yards per play, and in game variance probabilities are several I use often. I have found that simply identifying games that have a high predictive likelihood of falling away from the spread number (high standard deviation) are games I concentrate on during the regular season. I try not to be involved in games that are within 1 score of the spread - either way.
It is a constant battle. As you say - 87% over 624 bets is unheard of - although you haven't mentioned the cost of those bets. I would assume the cost of those bets brings your winning $ percentage way down - but that is just an assumption. If your ROI is 87% - I will hire you and you can retire from anything else you do - and that is not a joke.
Anyways - a little explanation why a post like this goes largely unnoticed here..... These are crack addicts here. They want the winner of the day - they need a fix. Most dont care about working on getting better or how you come to a pick - they just want the pick. A crack addict doesnt care how crack is made.
There are other reasons a post like this goes unnoticed.
1) 87% is nearly impossible over a significant sample size and makes you sound like a tout. So does you "consider subscribing"
2) When you say the margin of victory is 11.22445454, that means you have no understanding of significant digits and calls into question your mathematical skills.
I am not making any judgements on you or your model or your claims. I hope your model is that good and you get rich off it.
I work with European odds (much simpler to put into a computer) always over 1.83, with an avarage bet of 1.89. I'll assume you are familiar but for anyone reading that means for every 1 dollar put down you get 89 cent.
Quite cheesy but people like you are the reason I wanted to branch out. I know enough financial analysts but it takes a different breed to try to predict sports. Bill Benter has been a hero of mine since reading his book about 12 years ago.
I agree, the most profitable thing I did with my model is not to try and force a prediction for every game. If the model is not sure it will not give a prediction and look for another market in the same game.
Yes, I will not sustain the prediction accuracy (just by the law of avarages) but I know I can consistantly 60%. Again quite cheesy but money doesn't motivate me, it's more pushing the boundaries of AI and getting more people interested so this field can expand.
I agree with your first point. However I am not selling a course or picks and never will, but I realise how it looks when taken at face-value. The subscribing comment does sound a little foolish, but I really just wanted people interested in predicitve analytics to reach out and engage and maybe form a community, didn't really know how else to word it.
Regarding the second point I clearly remember my high school math professor talking about the dangers of continuous rounding, as well as in college. Ever since I try to leave as many decimals in the programme as you can get results very different with each operation even with 4 sig figs. My general rule is to leave them as it doesn't effect the outcome in any real way. Any predictions will be far from current bookmaker line to be counted as a prediction so working with such small numbers is insignifcant relative to the whole number parameters. That was a screenshot of the programme interface so rounding wasn't possible.
I will train a duplicate model using signifcant figures throughout just for you (and to prove a point) and see any differenceis found in accuracy.
From my conversations, it does seem that getting rich in sports betting is more about controlling emotions and bankroll managment, however time will tell
Please don't misconstrue any of the below as anything other than friendly observations, leading to or continuing friendly banter in this thread, thank you for starting it.
With a tremendous amt of respect, seriously, and zero sarcasm(for now), the problem with this forum is that with great frequency there is always someone with a system - and although the story behind the system varies greatly in interesting buzzwords they all have something in common and that is a past record of some really high winning %.
They then proceed to take that high % and make some predictions here and if they do decently for a small handful of posted plays/predictions they then look to gain subscribers, go tout or sell something, so unfortunately you will be incorrectly lumped into that category, just due to past experiences.
Buy as you can see many, such as myself will give the benefit of the doubt.
Interestingly my calculations from my models are KC by 9 and a total of 51 = KC 30-21 - VERY close to yours.
My issue overall (not including my many mental issues) is with choosing NFL as the sport in which to make $.
There is tremendous luck factor, lucky bounces, weather, bad/missed calls by the officials and games and pointspreads can be determined because of a single defensive missed tackle or slip. ANd this year we have covid. This all can happen in any sport but is greatly magnified by an extremely limited 16 game NFL sample size. Two bad calls, turnovers etc during a 16 game season and you went from a team win% of .563 to .438
You're a quant in finance, imagine if the equities, single or portfolio you are modeling are calculating beta exposure based on 2 years of history instead of 10. That's just not enough history.
This is why my eyebrow was raised upon reading your, "American football lends its nicely to analysis unlike other sports. The play by play nature make patterns very obvious edges quite easy to take advantage of." Would love to hear more on this from you as this statement makes me look like this:
A good thing about NFL IMHO is the liquidity of the market. A GREAT thing about betting NFL is that many times the market is wrong and a great edge can be had, but these usually occur in the first 4 weeks of the regular season when a team's "last year reputation" is driving the market on this year. Before the market can catch up there are a lot of good +EV line value games to be had.
Again, just my opinion. I wish you luck with your model.
Really great post. As you can probably tell I'm not a big NFL fan. I don't follow drafts or stories surrounding players (sort of the reason why i'm posing here as I know I need a new angle). I like the math behind games. Football in general is a very simple game. The permutations are small compared to soccer or rugby. The offence either completes or doesn't no second chances per drive. My feeling is that people tend to individualise games making a god out of each and every QB. Really it boils down to completion % and yards gained.
If you compare that with Rugby (A sport which I know quite a lot about) it's just so much harder to predict. Fatigue sets in and anything can happen whereas in football all players are ready to go, full concentration. It means there are very few plays that stray too far from normal. Like how in the NFL there are nearly always an average of 10 points scored in the 1st quarter. When compared to soccer the first 15 minutes average each year bounces around with no discernible patttern.
But yes I agree Data is king, and lack of meaningful data is a real flaw in this system, and why people steer clear of it. I however look at it like, If I have the same data as my competitors and analyse better, the model will always be more accurate.
I cannot say this is analysed in an extensive way, as there would be too many variable with little data available from player to player. The main parameters for players are their rankings at their position. So average running yards, passing yards, sacks, funmbles (essiantially my own modified power rankings) however because this is so superficial it is clearly a weakness. Personally I feel if someone/a seperate model was able to know the players very well it would definitely be an edge, I simply do not know enough about the players and don't want to get too invested as emotions would clearly cloud my judgement.
At the moment I simply don't have the knowledge or computing power for such an analysis (or mathemthical ability tbh) to analyse each player. But if there was a way of condensing or averaging each players contribution I am certainly willing to try.
What is the most important factor when analysing how impactful a player will be as an individual?
87% on the season? Care to post wk1-20 please? Here’s the thing: a computer may get it right a decent amount? But this isn’t a computer game. AI fails to take into account the human element. I agree Chiefs by a large margin? I say 8-9pts. However! TB has the better Defense and the PROVEN QB in this game. Odds are highly stacked against him. Can he overcome them? If anyone can? He’s the ONLY one who can.