Lets say I've developed a model that is roughly 60% accurate at determining if the favorite will cover the spread or not in NBA games. The output from this model is in the form of probabilities like [0.3, 0.7] which would mean in this case that the favorite is likely going to cover. At first I thought about subtracting the implied probability given by the bookie from the probability given by my model to determine my edge and then base my bet sizes on the edge but then I realized there is an issue of accuracy. Lets say a bookie gave a favorite -110 or 52.38% and my model says 70% we would have basic edge of 17.62%. Normally with an edge I would do something like this to determine the max percentage of my bankroll to risk in this situation: edge/(1/implied_odds)=9%. The issue is that it isn't taking the model's accuracy into account.

How would you determine your bet sizing given the model's accuracy and the preceived edge to maximize gain while minimizing risk?

What does "60% accurate" mean? Does it mean that the model's projected outcome is correct 60% of the time? And is that 60% a constant, or is it instead a function of the output of the model, that is if the model says [0.3, 0.7] then it should be 60% accurate, but [0.2, 0.8] it would be 70% accurate (for example).

You can build an expected growth model in a spreadsheet but without knowing the probabilities of win/loss you cannot estimate your expected growth, and obviously your optimal bet sizing.

I thought about that a little while after I posted. So I went back and tested to see if the model's output probability correlates with the accuracy and it does but only kind of. For example, if I set the minimum edge to say 30%(which would mean the output is 0.8 or higher) the overall accuracy moves up a few percentage points but not at a 1:1. It is a lot easier to say that the model is 60% accurate and that the edge is the accuracy minus the implied odds(typically 6-8%).

edit

60% accuracy means that over 20k games it accurately picks the games where the favorite covered just over 60% of the time. I actually mean the f1 score is just over 0.6

?

With just a question mark it is hard to figure out which part you have a question about. Is it the favorite spread odds? -110/52.38% is pretty common. Is it the model output? In that example I said the output from the model is [0.3, 0.7] or roughly 70% for the favorite to cover.

I realized later on that the output probabilities is different than odds. In fact there is a very clear difference between odds and probability.

calculate the edge again

also I would say it's impossible to get 70% for any market

have you ever asked yourself why it might be covering 70% of the time?

There's a good chance you're wrong because that sort of edge is pretty close to unattainable, although I do have a 15% roi on a very niche sport with low limits across 3 seasons and over a thousand wagers so it might not be impossible if you're betting some super niche sport but on mainstream sports it shouldn't be possible - just google kelly criterion and act accordingly.

If you really do have a big edge obviously it should be trivial to grow your bankroll quickly. Worth considering that your model is likely to be flawed though if it's returning 70% winners at a spread but if it really is that's your ticket to being a millionaire, gl.

The model output probability doesn't correlate with the model's overall accuracy. The overall accuracy is just over 60% for picking games where the favorite covers the spread. In the case of [0.3, 0.7] the prediction is the index of the largest number so 1. The model would say in that example that the favorite was going to cover. I was incorrect in thinking that 0.7 in that case had any association with odds given by the book and if it did then I would have to also take the overall accuracy into account.

Here are its predictions for today/tomorrow. So far the first lost but we'll see it how it ends up:

================================================
Orlando Magic @ Boston Celtics on 2018-10-22
Spread bet on Boston Celtics
-11.0 -110
================================================
================================================
Chicago Bulls @ Dallas Mavericks on 2018-10-22
Spread bet on Chicago Bulls
5.5 -110
================================================
================================================
Phoenix Suns @ Golden State Warriors on 2018-10-22
Spread bet on Phoenix Suns
12.0 -110
================================================
================================================
New York Knicks @ Milwaukee Bucks on 2018-10-22
Spread bet on New York Knicks
11.5 -113
================================================
================================================
Indiana Pacers @ Minnesota Timberwolves on 2018-10-22
Spread bet on Minnesota Timberwolves
-2.0 -110
================================================
================================================
Charlotte Hornets @ Toronto Raptors on 2018-10-22
Spread bet on Toronto Raptors
-8.5 -105
================================================
================================================
Memphis Grizzlies @ Utah Jazz on 2018-10-22
Spread bet on Memphis Grizzlies
10.0 -110
================================================
================================================
Sacramento Kings @ Denver Nuggets on 2018-10-23
Spread bet on Sacramento Kings
11.0 -110
================================================
================================================
Los Angeles Clippers @ New Orleans Pelicans on 2018-10-23
Spread bet on New Orleans Pelicans
-6.0 -115
================================================


I ran this about an hour before the celtics game but the time I posted this was Q3 in the bulls game. ignore the first two if you want because I obviously didn't post them here before the games but I did run this before those two games.

You omitted the model's output. That would go a long way to answering your original question. For the moment all I see is a binary "play or no play", and so I assume you are only choosing a side if the model outputs a value exceeding some arbitrary threshold.

Either youre wrong or youre one of the best sports bettors in the world right now and should be a multi millionaire in a couple years. Hmm.

http://www.bettingtools.co.uk/kelly-calculator

Kelly Bet formula calculates the optimal betting strategy for fund growth, so if you're confident in your edge then theoretically a Kelly Betting Strategy should lead you to the most money.

Usually people using a Kelly betting strategy use a %Kelly approach, e.g. 33% Kelly if Kelly suggest $3000 you bet $1000. Kelly is more focused on fund growth than preservation and can be very aggressive, it's certainly wise to build in a margin for error on your estimation of your edge, plus considerations like cash flow utility (how easy it is to credit accounts, limit issues etc) and volatility control.

His question is how to quantify his edge. Everything after that is ABC.

No he is not

Exactly. Read it again.

He has clearly stated what his edge is..

Use the Kelly Criterion to determine how much to bet when you have an edge: https://www.pinnacle.com/en/betting-...T2LK6K2QWQ7QJ8

If you are referring to this:
...and this:
Then no, he has not.

Feel free to cite where he indicates he has quantified his edge.