Algorithm to automatically guess sport bets

[raffish]

Hi everybody,

sorry if I post again about the same topic, but my previous post violated the forum's rules on spam and self-promotion and was thus removed.
I'll try to explain to you my business idea without any links nor self-promotion.
It will be a quite long post, but I think it will be worth for you too.
What I mean with this post is to present to you my idea and receive your feedback on it. What do you think of it?

My idea is to generate not passive, but very-little-active income having a method to automatically guess sport bets on websites without any sport knowledge nor reasing.

Concretely, I wanted to exploit the Machine Learning skills I learnt during my master thesis at university to create an algorithm guessing bets for me.
I concentrated on football bets with three possible outcomes (1 - draw - 2) on my favourite betting website.

What I did was to begin collecting daily all bets and related quotes and results (outcomes). This was the trickiest part as it required lots of time in the beginning, but I progressively automated the whole thing with macros which format and prepare data for me after raw copy-and-paste from the website, removing duplicate bets and so on.

What I actually do today is only to take a copy and paste of all bets on the site like 4-5 times per day while I'm at work. Then before going to bed I manually fill in the results of the day (5-20 min. work).
Then, once a week I update the decision tree which makes my predictions.

The interesting thing is that the prediction accuracy of my algorithm is linearly increasing in the number of matches filed in my database. Up to 7th October 2007, it has been trained on 4.974 football matches worldwide and proved able to guess 52,85% of their 1X2 results.

I’m very proud of this result because while being a betting aid for me and my website readers, it is itself a bet for me at the same time, as I bet on my skills in applying to this particular field the Machine Learing skills I have learnt at University.

What's more, I calculated that 60% accuracy is the threshold to begin having very interesting economic results. The goal I have is to achieve a prediction accuracy allowing me to earn DAILY something like 1% of the invested amount in bets.

For example, I want to achieve something like that: everyday I will spend a fixed 10.000$ on all bets on my favourite website, without knowing nothing on the matches I bet on, and be sure that in the long run I will always collect back 10.100 $ at the end of the day .

Also even only a 1.000 $ initial capital, making me earn 10$ per day in a brainless way could be good.

If 1% daily seems nothing to you, please consider if you wouldn't be happy to have your stock options behaving the same way ..

It took me a while to write all this .. please do not consider it spam, do not remove it and comment it instead

I'm really longing for your feedback on my idea !

[raffish]

25 visits and no replies?
Don't be so shy everybody
Tell what you think after reading

[da1prophet]

I've been a professional sports handicapper for the past 15 years and have the sharpest minds in the business on my speed dial. There's no such thing as a perfect algorithm to predict the outcome of sports wagers. The three way wagering used frequently in soccer (what the Euros call "football") is even more difficult to predict.

One thing you said that is dead wrong is that "60% accuracy is the threshold to begin having very interesting economic results". That's dead wrong for the simple reason that the theoretical breakeven points for sports bets is different at different moneyline prices. Now, if you were able to get all of your bets at even money or the standard -110 dealt for sports with pointspreads you'd be raking in the cash. At -150, however, your *breakeven* point is 60%.

The problem with your formula is that a good many plays that an algorithm would predict correctly would be priced higher than -150. Looking at next Saturday's English Premier League, for example, Chelsea is priced right at -150 and Manchester United at a whopping -500. At -500, theoretical break even is somewhere around 85%. So if Man U loses or draws you've lost nearly the equivalent of 5 units of your wagering bankroll. With the frequency of draws in soccer you're going to eat a few big losers like this over the course of the year. So if your algorithm suggested bets on Chelsea and Man U next weekend lets assume that Chelsea wins and Man U loses. For the purpose of tabulating wins/losses your algorithm has gone 1-1 but in terms of wagering units you've really gone closer to 1-5 (1 unit won on Chelsea, 5 units lost on Man U).

More later...

[raffish]

Dear da1prophet, THANKS THANKS THANKS for you post, because It is exctly the kind of professional reply I was in search of.

Maybe it is even too technical, as I don't even know the slang "-100" or "-150". It is common speach for who has been in professional betting for years, it is Arab for a Euro newbie making his home-brew stats on his own in Excel

I'm used to see bets phrased in this form:
Chicago Fire - Washington Dc United
1: 2.50
X: 3.10
2: 2.50
where "1" stands for betting that the team playing at hoem will win, "X" betting on draw and "2" betting on the team playing away win.
For Example, 1: 2.50 means that if I bet 1$ on Chicago winning and they win, I will be paid back 2.50$ .

I can show you how I calculated the threshold to begin making money.
1. I assume that every day I put a fixed "1 money unit" on every match on Betway website.
2. At the end of the day, my algorithm will have performed a X% accuracy overall in predicting the 1X2 results of all bets, and according to the wages of the bets it will have lead to a Y% earning on the invested money (wether positive or negative).
3. Thus I've plotted a graph with all such (X,Y) points, one for each day. You can see the graph here.
4. A simple linear regression on the point dispersion on the graph shows about 55% to be the threshold to begin making money in the long period, and 60% to allow making a 5% earning.

Pls comment on my work. I promise I will publish an article on my website to explain in details my work.

I'M REALLY LONGING FOR YOUR OPINION ON MY WORK, I GUESS YOU ARE REALLY THE KIND OF PROFESSIONAL WHO CAN PROVIDE VALUABLE FEEDBACK AND LIFT MY WORK'S LEVEL.

Thank you

[ixmatus]

It is an interesting idea. What neural net libs are you using?

I would imagine, that an evolutionary algorithm might be most suitable for this type of thing...

Use every two games as a new generation with the team (or teams, if you are tracking multiple ones) in question represented as a singular entity in the model.

Ex: Broncos vs Cowboys & Broncos vs Chargers; each occurrence of 'Broncos' is considered a singular entity.

Each Broncos entity will receive a breeding fitness magnitude based on (just off the top of my head):

• win or loss
  • win = 1
  • loss = 0
• win/lose gap
  • if desired team wins with large gap, higher fitness
  • if desired team wins with smaller gap, lesser fitness (but still above 1)
  • if desired team loses with large gap, very low fitness
  • if desired team loses with smaller gap, low fitness (but still below 0)
• time to win or loss
  • if shorter the win, the higher the fitness
  • if longer the win, lower the fitness (but still above 1)
  • if shorter the loss, very low fitness
  • if longer the loss, low fitness (but still below 0)

The algorithms for calculating who breeds with who could get pretty complicated, so, I am not going to attempt an example; but the general idea being that those multi-faceted fitness levels provide for multi-dimensional breeding selections.

This kind of system would obviously work best with large data sets. So, historical data with present time (or even real time data too, to see live breeding) data would provide a more mature system to breed your teams with than simply present time data.

You could breed multiple teams and use a simple script to analyze their breeding fitness amongst each generation to produce results of which team is fitter to win in a match.

Would be an interesting experiment...

[raffish]

The strange destiny of this post .. 2 weeks with 150 readers and no answers.. and then all at a sudden 2 EXTREMELY VALUABLE replies in one single day

I'm so happy with the smartness of people around this forum.

Back to the topic, I'm actually using C4.5 algorithm on individuals with 3 numeric attributes: 1 wage, X wage and 2 wage. The good thing about this is that I can bet on any football (soccer) match without actually knowing which teams are playing. It doesn't really matter. I only need the three numbers from the website in order to get a prediction on how to bet.

I also tried to add a 4th attribute (a label identifying the competition), but surprisingly enough this reduced the accuracy by 2-3%. I know that this can be due to the small number of matches recorded in my DB (actually around 5900, but increasing by around 500 every week), In fact 4-attribute accuracy is increasing over time faster than 3-attribute one, altough being lower at the beginning .

I renew my promise to publish a full report of my work on my website within this weekend.

I also kindly ask you why do you think that neural networks woudl be the most suitable kind of Machine Learning algorithm for the task?
I guessed that decision trees were what I was in search of : if "1" < 1.10 let's bet 1, else if "X">3.00 let's bet "2" and so on...

Isn't an if-then decisional tree the kind of final output we are aiming for ?

Pls remember that teh purpose of my

[ixmatus]

I'm seeing it as a fitness scale based on multiple criterion for a win/loss condition of a team and that team's win/loss interaction with other teams.

I am not sure how Futbol (sp?) works, but, my previous post was geared towards American Football and how the team match up hierarchy is constructed.

I'm only a hobbyist in using ANN's and EA's, so my concepts are somewhat immature; but, it seems to apply well in this scenario in my mind...

Maybe, when I get my computer working correctly, I will try and come up with an example python script.

[Jonathan0104]

Your algorithm only uses the data from the betting site? That is actually a very big problem!

You can't make any profit (in the long run) with an algorithm like that. Let me explain it using an example:

[1] Inter Mailand 1.20
[0] (Draw) 5.20
[2] Genua 13.00

Your algorithm will most likely recommend to bet on [1]. Of course you will win a lot of times. But the problem is the following: the betting site wants to make money. So in order to make money they will give you bad odds. In this example the "real" probability of the event [1] may be 1.30 instead of 1.20.

So if Inter wins you get 1.20 * [the amount of money you wagered]. But to break even you need to get 1.30 * [the amount of money you wagered]. If you bet like that a thousand times you will see that you lose money.

Your algorithm must be better than the "prediction" at the betting site in order to make money. You need to have more information than the betting site to make money.

Your algorithm has to "spot" the bets which have "good" odds. E.g. your algorithm says that the probability of the event [2] is actually 10 not 13 (as above). This means you have to bet on [2] because in the long run you make money. Statistically Genua will win every tenth game but you get paid 13 * [your money] -> profit.

This is a mathematical problem. I didn't use a lot of mathematical expressions because I don't know how much you know about mathematics/statistics.
I don't know if my text is comprehensible but I can't explain it any better. I suggest that you read some books about stochastics and math. There are some books who deal with betting and the mathematical aspects.

[raffish]

Good mornig Jonathan,

thank you for your reply, you have come to the point.
In the long run, I will quite surely be able to have a good accuracy for my predictions (70%), but guessing the easiest 70% of the bets (in other words: the easiest ones) is not enough to earn money in the long run.
It would be better to gain only the 5% more difficult bets to pay back the 90% failed easy bets and earn money too.
Don't be scared of talking mathemathics with me (I am a computer engineer with a master thesis on machine learning algorithms).
Any idea to help me strenghten my algorithm ?