[usma1992]

reasoning. Would help.

Both systems work off the same principles except one system allows you to bet more games. I have strict money management criteria. 3% per game...no more than that.


System 1: 57% success rate over 2300 games. The benefit of this system is that on a daily basis I have more money on the line. So financially on a slow day of only 17 games... if one system allows 14 bets and one allows 6 bets... I have 43% of my role on the line vs. 18%. Financially, the upside is greater. In addition, you are spreading the risk over more games, so I believe your downside is also capped.

System 2: 59% over 1300 games. The benefit is obviously the 2%. And I have run the numbers, 2% adds up to a significant amount of money. But that is offset by not having much of your bank role on the line daily. In addition, It is easier to have an off day with 6 bets than with 14.


57% is very strong and I know that. But 59% is really strong and obviously the extra 2% makes a difference. Just could use some opinions.

Ty in advance...
Dave

[hutennis]

To start with...

When I don't see Avg Implied probability posted next to Winning %% I become really suspicious on many levels.
I mean, common! Don't you understand an overwhelming importance of it?

[usma1992]

Educate me...My degree is in electrical engineering... but I have been far removed from statistics for quite some time. Clue me in, what am I missing... Ty for the reply... these are the picks from the 57% model... in the 59% model it keeps all for evening games an Temple,Wisconsin game and instead of 10 bets only has 6. Please clue me in on the average implied probability. I have all the data, how can I calculate it.

12:00	Syracuse	USF	126.5	O
12:00	Iowa	Michigan	142.5	N
12:00	St. Peter's	Loyola (MD)	116.5	O
14:00	Siena	Rider	124	N
16:00	Texas-San Antonio	Denver	123	U
16:30	Temple	Kansas	142.5	U
16:30	Wisconsin	Nebraska	117.5	O
17:00	Manhattan	Iona	143.5	U
17:30	Florida	Yale	128.5	N
17:30	Wichita State	Bradley	127	N
19:00	Northwestern	Minnesota	131.5	N
20:00	Tulsa	Southern Methodist	128	U
20:00	Colorado	Arizona State	133	O
20:00	North Carolina	Virginia	133.5	O
22:00	Oregon	Oregon State	147.5	O

[u21c3f6]

Are all of your wagers @ -110?

Joe.

[usma1992]

Yes...except for the baseball systems...

[hutennis]

Yes...except for the baseball systems...

What is another side?

-110/-110 or -110/+100?

And when you say all of them you mean every single wager is placed at the same odds?

[usma1992]

Sorry... these are all college basketball totals... all are at -110... thought that was understood. So my breakeven is 52.38%. But back to my original question. 57% over 2300 games or 59% over 1300 games... all at -110/-110....

[hutennis]

Ok, never mind.

Here's what we got.

You are saying that with avg probability of winning the bet = 50% you won
57% or 1131 out of 2300 games and
59% or 767 out of 1300 games.

The probability of getting results like this by luck is nice, big, fat 0.
So, basically you are sport betting genius, the best ever.

Yet, you don't know what an implied probability is, you don't understand importance of posting odds,you've been removed far from statistics
and you have trouble figuring out on your own witch system will yield you more dough in a long run.

And yet, with all this ignorance on a elementary subjects, you did manage to develop
indisputably the best sport betting models of all time.
Models that could be sold/bought for countless millions tomorrow.

I don't know. Does not look good.

[Miz]

Do your training and test sets overlap? Meaning, are you testing the system on the same dataset that was used to develop it?

[usma1992]

I understand you comments hutennis and I appreciate your comments...still unclear about the implied probability piece, but yes I am assuming a 50% chance(honestly on a college basketball total) not sure what else you would assume, thought that was understood. Understand the importance of posting odds...(80-90% of the bets I place are -110/-110), so I apologize if that wasn't posted. For me, my baseball program is the only time that comes into play and since I only play totals, most of the time even in that sport it's -110/-110. All my data is based on totals... very rarely if ever bet the spread.

As far as what will yield me the most dough in the long run. Please don't show your ignorance in this subject. I believe that yielding the most dough has to be compared to the amount of risk taken. If someone bets all their money on one game and doubles up... I find that quite different than someone that doesn't allow a more than a 3% bet on any individual event and yields the same results.

My quest has always been to create a mutual fund of sportsbetting. Only allowing 3% on any given bet to minimize risk. As far as developing the best sports models of all time. Yes, that is my goal... not to quote "Miracle", that is why I want to pursue it.

But yes, I am well on my way to producing that model. From a 4 year old laptop and can download all the games, update all the statistics and place my bets within a 10 min timeframe. I have fully tracked app. 8000 college basketball games of which I have now have the results that I published above. This year has been great so far and yielding well over 60% the previous weeks. But I continue to develop find tune etc...

As far as looking good. Yes I really am... I was just asking a simple question about risk/reward and who has a decent experience with it relative to sports betting.

[usma1992]

Miz I developed the system and tried to learn the game of college basketball over the last couple of years to try and determine what was important. Than I developed the system and than after I developed it, I began trying to figure out the best way to address the data and optimize the results.

Now I am using the previous results and trying to predict the future results. For the first time in 3 years, I am winning consistently. Colllege Bowl Season predicted 58%+ on 24 of the 34 bowl games. Tonight the system predicts and under bet on college football(Kent State) and says don't touch the Bama game....

[u21c3f6]

If you are going to wager only 3% on either system, then you will make more money with the 57% 2300 wagers. However, from a Kelly point of view, you can wager a higher % on the 59% 1300 wagers which would result in more money assuming that your win rate continues for future wagers.

If I understand you correctly, your 59% is a subset of your 57% wagers. I would wager on all 2300 events and increase my wager for the 59% subset (not more than Kelly, which at 3% you are very far from the max wager).

Joe.

[Miz]

Is your approach a model or a system?

System is a naughty word around here. So let me ask it this way...

Do you quantify the difference between your projection of the margin of victory (or total) and the book's line?

[usma1992]

That u21c3f6...I thought the same thing except, I am very hesitant to wager more than 3% because I believe that if I have created a sustainable system that it is just a matter of time to make the money I need. Yes the 59% is a subset of the 57% of the wagers. If I am using 3% on the subset of the 57%(not including the 59%) what should I use on the 59%(Subset)... How do I calculate the Kelly value?

For instance, yesterday... the 57% program yielded results of 54% and the 59% program yielded results of 46%. So it happened in that specific scenario the 59% program tossed out the bets that actually won.

[usma1992]

Miz... I have created 5 models, NBA, NFL, college basketball, college football, and NFL. For the moment, I shelved the NFL program... too hard to predict. These are all computer models using statistics and some algorithms. NBA is only in its second season. College Football is finally showing signs of life. College Basketball is my favorite and that data that I what I provided in my previous post.

I have tried to quantify the difference several times and I will probably eventually get back to it. I look at individual games and determine at what total value they would change to a neutral bet and than eventually to the opposite bet.

A good example is the UVA game a couple of weeks ago. It opened at 104 total...by the time I got my bet in it was 107. I checked it was still an over bet so I bet it. The line closed at 111. I tested my program and 111 it was an over bet at 112 it went to neutral. That particular scenario really made me feel great... because the line went to the exact point it changed on my program 111-112.

[Miz]

What total did you project for the UVA game? How did you arrive at that number?

[usma1992]

I have complex algorithms...that look at specific parameters from 3 different directions... so it isn't really as easy as what number did I arrive at... It more about what numbers fall with a neutral area depending on the characteristics of how the game unfolds. Not trying to skirt the question, but I place a plus or minus 5 points on the total that I won't bet it. so since it changed from Over to neutral at the 111 point... my program projected 116 or 117 so my system allowed me to bet the game. It turns out the scored 113 points total. I try to use a 10 point neutral window that I don't bet. What I find astonishing is that the system projects the amount of games that it does. 10 points seem like a pretty big neutral window. But it works...so far...

[jgilmartin]

Do your training and test sets overlap? Meaning, are you testing the system on the same dataset that was used to develop it?

The most important question.

OP, this will determine Kelly staking:
http://www.sportsbookreview.com/betting-tools/kelly-calculator/

[statdude]

what if you just overlap the systems?
are these back tested results or tracked live over time?

[u21c3f6]

Using the Kelly calculator, full Kelly for the 57% is: 9.7% and for the 59% is: 13.9%. I personally use something that is closer to half Kelly as I do not like the swings of full Kelly and it helps to prevent me from overbetting due to miscalculations of edge.

If you use half-Kelly, then your wagers would be 57%: 4.85% and 59%: 6.95%. If those are too high, then you might think about 57%: 3% (your comfort zone) and 59%: 4.3% which maintains the same Kelly % difference. This of course is based on your ability to continue your winning %'s. But as I stated earlier, by wagering less than full-Kelly even if you come up a little short in your winning %, you will still not be overbetting. Only if your winning % drops significantly will you be in danger of ruin.

As a side note, since I don't bet full-Kelly, I don't worry about doing simultaneous event calculations since the change is relatively small and I am already betting only half of the max.

Good luck.

Joe.

[HeeeHAWWWW]

Do your training and test sets overlap? Meaning, are you testing the system on the same dataset that was used to develop it?

That was my immediate thought.

[Miz]

Ok, no need to divulge details about your model. I certainly don't say much about mine either. Best of luck on what you decide. You can certainly calculate the growth and kelly stakes on your own.

My opinion on the immediate choice between the two would be to play fewer plays at the higher win rate. I would stake assuming a 55% win rate. Then you won't go bust when the forward betting doesn't go exactly like your backtesting did.

Good luck on your plays

[hutennis]

I understand you comments hutennis and I appreciate your comments...still unclear about the implied probability piece, but yes I am assuming a 50% chance(honestly on a college basketball total) not sure what else you would assume, thought that was understood. Understand the importance of posting odds...(80-90% of the bets I place are -110/-110), so I apologize if that wasn't posted. For me, my baseball program is the only time that comes into play and since I only play totals, most of the time even in that sport it's -110/-110. All my data is based on totals... very rarely if ever bet the spread.

As far as what will yield me the most dough in the long run. Please don't show your ignorance in this subject. I believe that yielding the most dough has to be compared to the amount of risk taken. If someone bets all their money on one game and doubles up... I find that quite different than someone that doesn't allow a more than a 3% bet on any individual event and yields the same results.

My quest has always been to create a mutual fund of sportsbetting. Only allowing 3% on any given bet to minimize risk. As far as developing the best sports models of all time. Yes, that is my goal... not to quote "Miracle", that is why I want to pursue it.

But yes, I am well on my way to producing that model. From a 4 year old laptop and can download all the games, update all the statistics and place my bets within a 10 min timeframe. I have fully tracked app. 8000 college basketball games of which I have now have the results that I published above. This year has been great so far and yielding well over 60% the previous weeks. But I continue to develop find tune etc...

As far as looking good. Yes I really am... I was just asking a simple question about risk/reward and who has a decent experience with it relative to sports betting.

Look, I don't know, maybe it is just me, but...

The best sport betting model of all time from 4 year old laptop operated by a guy who "has been removed from statistics"
sounds not just too good to be true. It sounds intelligence insulting.

So far you have received a gigantic benefit of the doubt here. I don't even know why guys around here are even talking to you to the extend they did without you proving first that you are not a delusional dreamer or what they call an attention whore.
And proving is very easy. You are not asked for your records or details of your model.

Just start posting your picks ahead of time so we can get into some statisticaly significant neighborhood together.
That would do wonders for your credibility.
Otherwise it is just a big steaming pile of BS and no rational person should spend a minute of his time on it.

[usma1992]

I understand... your comments... but please don't take this the wrong way. There is no benefit for me to lie... in sportsbetting the truth is always in the pudding. You either win the games or you don't... there is very little middle ground.

If it is working, it is working...for the first time in 3 years it is working and yes I have spent 3 solid years on my program. I actually have gotten investors year 1- year 3.

Last night on the 59% model I went 2 out of 6 and on the 57% model I went 3 out of 10. I have no problem saying that... and honestly have no problem posting my picks. Most of the time however, I end up running out of time. And posting my picks become very low priority. Most of the time, I am watching the games trying to come up with an extra .5%.

Again, no reason to lie... At this point I have tracked about 8000 college basketball games and my program shows that of 2300 games I hit 57%... I would love to get that higher and work everyday to do that. Unfortunately, I have had to take a full time job and an 8 year old son...so I don't always have the time... but I track every day.

NBA not ready, MLB questionable, NFL... shelved it because I can't get it to work... College Football, I think I'm there... college basketball also...

Anyway, to those that provided feedback, I appreciate it. For the last 3 weeks, my programs have worked consistently for the first time ever. Hopefully... the march to march madness will show the same....

And I am going to think about the Kelly factor to see if that makes sense...

[usma1992]

To the question of training vs. testing... the data. Basically, I have recorded real time data for half of 2010-2011 all of the last year and all of this year up to this point. I downloaded statistics everyday or once every 3 days during those time periods to ensure, I was looking at the exact data that Vegas looked at when they made the line. All my data and spreads are pure.

Than when I identfiied what I felt was important and weighted the charateristics according, I created my approach to how to view those characteristics and over time I fine tuned it to the point, I am at now.

I understand that past data is not indicative of future performance... but the program for college basketball and football are performing ahead of the games now. Each year, I anticipate fine tuning further... but right now it is working and the tweaks, I make are minor as I collect more data.

[usma1992]

Hutennis,

The one thing I thank you for is understanding the value... It is hard to explain to people that hitting 57% is worth millions. Most people think 57% should be easy and that 80% is where you need to be. Even my friends have a tough time grasping how valuable my model is. If you are not inside the sportsbetting industry... you really have no clue.

Ty
Dave

[usma1992]

18:00	Illinois-Chicago	Cleveland State	125.5	N
18:30	Notre Dame	Cincinnati	134.5	O
19:00	Georgia State	Hofstra	127	O
19:00	Indiana	Penn State	138.5	N
22:00	Southern Utah	Sacramento State	134	O

[matthew919]

To the question of training vs. testing... the data. Basically, I have recorded real time data for half of 2010-2011 all of the last year and all of this year up to this point. I downloaded statistics everyday or once every 3 days during those time periods to ensure, I was looking at the exact data that Vegas looked at when they made the line. All my data and spreads are pure.

Than when I identfiied what I felt was important and weighted the charateristics according, I created my approach to how to view those characteristics and over time I fine tuned it to the point, I am at now.

I understand that past data is not indicative of future performance... but the program for college basketball and football are performing ahead of the games now. Each year, I anticipate fine tuning further... but right now it is working and the tweaks, I make are minor as I collect more data.

I think you misunderstood the question of train/test set overlap. Try this: take your full data set, and randomly separate in into 2/3 (train) and 1/3 (test) subsets. Then train your model and estimate whatever parameters you are estimating using the 2/3 set. When you have your parameters, apply it to the test set, and record the results. Ideally, you should start by using a cross validation approach on the training set to estimate your parameters, and THEN apply it to the 1/3 sized test set.

What that means is to take the training set and randomly break IT down into 2/3 train 1/3 test, estimate parameters on the 2/3 sized set, apply to the 1/3 set. Repeating this thousands of times will allow you to optimize the parameters BEFORE you apply it to the original test set, and (hopefully) achieve better performance. Hope this helps.

Matt

[HeeeHAWWWW]

Here's an example from tennis: I have a relatively simple model whose purpose is calibrating other elements. If I run a 2012 backtest based on matches/stats from 2004-2011, it produces 52.5% for the 2012 season, over about 1000-1100 wagers. If it is set to calculate the 2012 season based on 2004-2012, it hits 57.3%. This sort of discrepancy is absolutely iron-clad predictable, because the latter isn't a genuine backtest, it's using information from the "future".

[Miz]

Matt and HeeeHawwww, correct and correct.

[usma1992]

My approach was to first identify what I felt was important. I set up my training set based on the data I had from 2010-2011 and than tested it over the course of the 2011-2012 season. After a month of the 2011-2012, I changed the weightings of the parameters so that the weightings still worked over the course of 2010-2011 and with the current data. And I would test the next several weeks and than make adjustments. I found that two seasons at a minimum are mandatory than I made minor adjustments over the course of this season. I found that after December 11... my planned start of the betting date...The program has worked well, actually very well. But I have continued to make adjustments over the course of this season to continue to fine tune and test further. Published are the results on the far left and the amount of games second from the left... all the other data is just how I track the individual games. January 2011 is off a little because I was toying with it... but u get the point.

2011 January	45.33%	150	0.0%	0	0	41.9%	18	43	45.7%	16	35	47.2%	34	72	-1	3.5	7
2011 February	56.58%	281	59.0%	59	100	0.0%	0	0	0.0%	0	0	55.2%	100	181	10	1.1	2
2011 December	55.08%	187	51.9%	42	81	0.0%	0	0	0.0%	0	0	57.5%	61	106	-7	2.8	4
2012 January	61.22%	245	66.7%	64	96	0.0%	0	0	0.0%	0	0	57.7%	86	149	-2	2.2	4
2012 February	63.33%	240	59.8%	52	87	0.0%	0	0	0.0%	0	0	65.4%	100	153	23	1.6	5
2012 March	54.88%	82	45.5%	10	22	0.0%	0	0	0.0%	0	0	58.3%	35	60	-6	2.3	5
2012 December	60.65%	155	68.4%	52	76	0.0%	0	0	0.0%	0	0	53.2%	42	79	5	4.7	1

[matthew919]

1. That's an awful lot of wagers per month.
2. I'm counting 3-7 last night, and so far 0-2 tonight, which makes those numbers very hard to believe. Although if they are legit, then good job.
3. Are these games you actually bet on using your model AFTER re-training on the pre-2012 data? Or when you updated your model in 2012, did you train on these data? It sounds like you trained on them, but your wording is extremely confusing. If so, then these results are relevant to the training data only, and they are not valid as benchmarking statistics.

Also, for future reference, work on wording things more clearly. For instance, when you say "I changed the weightings of the parameters so that the weightings still worked over the course of 2010-2011 and with the current data," I have no idea what you mean by "current data." Furthermore, when you say you were "toying with" your model, I need a better definition. In statistics, there's only "training" and "testing", not "toying with." You need to be explicit about exactly what you trained on, and exactly what you tested on, and know that there are hard boundaries between those two data sets that cannot be blurred without invalidating your results.

[Miz]

Best of luck Dave. I think quite a few people have given you suggestions and ideas, all of them constructive. I hope you choose to use them.

[usma1992]

On the 57% program you are correct... I am 3-7 last night and 0-2 tonight. I get that the numbers are hard to believe but they are true... but those numbers reflect the 59% program... except for the Jan 2010-2011 because I was checking something before I posted. I think... I am going to go with the 59% program after the last couple of days. I am not trying to knee jerk... but I have gone 3-7 and 0-2 on 57% model and I would have been 2-6 and 0-0 since no bets were created tonight...

When I test... rather than tweak(my words)... I have a centralized parameter page that allows me to change the weightings of several parameters all at the same time and see how they are alter the results of the previous years, in addition to how they reflect not only the current year's December performance, but also the current year's January performance.

If I change the weightings to reflect positive performance this season and it doesn't work the previous seasons it obviously has no validity over the long term. It has to work consistently over several seasons, several months within each season, with different team compositions...

To your point... I am not an English major by any means...

[usma1992]

Miz... thank you ... I need it...