 SBR Top-Rated Sportsbooks |
Best Sportsbooks List
|
|
| |||||
| #1 FanDuel | SBR rating 4.8/5 | Review | #6 BetRivers | SBR rating 4.1/5 | Review | #2 Caesars | SBR rating 4.7/5 | Review | #7 Fanatics | SBR rating 4.1/5 | Review |
| #3 DraftKings | SBR rating 4.7/5 | Review | #8 Betway | SBR rating 3.8/5 | Review | ||
| #4 BetMGM | SBR rating 4.6/5 | Review | #9 Borgata | SBR rating 3.5/5 | Review | ||
| #5 bet365 | SBR rating 4.6/5 | Review | #10 ClutchBet | SBR rating 2.9/5 | Review | ||
Entering Year 4 of My Computer Programming Model... College Football
2010 Net Percentage 64.2% Games Won 203 Under 63.5% Net Games Won 90 Games Bet 316 Over 65.8% 2011 Net Percentage 64.2% Games Won 213 Under 64.8% Net Games Won 94 Games Bet 332 Over 63.1% 2012 Net Percentage 61.9% Games Won 197 Under 63.2% Net Games Won 76 Games Bet 318 Over 61.8% Net 260 63.5%
Reply
These are my college football back tested results. Have I made a dime on college football? No, I have not. Year 1 I started with a blank spreadsheet and had to add every team name from scratch let alone the program. It took me a month even to get the product to the point it could print out any results. I busted out year 1 by week 5-6. Year 2 a similar result... Year 3 I revamped the entire system prior to beginning of the season, in order to start testing again. I added many features. I made it through the season but was down approximately 60-70% by the end and had to re-re-revamp the system prior to bowl season. Bowl Season was profitable, but overall I lost money Year 3 also. Year 4 I begin with my fourth iteration of my system. Every game has been treated fairly statistically, and I believe that I have identified a system that works. At what point are the data trends enough to have a reasonable view of success. I have tracked 1500+ games, tested and re-bet and retested to get to this point. I completely understand that I have not been successful forward looking but the tested results give me a fair amount of wiggle room to be successful. I caught a lot of grief about my college basketball system last year saying that it hadn't been fully tested. I went on to hit close to 60% of the games and earned about 300-400% over the course of the season. My question for the board is two fold. How many seasons is fair? How many games are fair? I have 5 Seasons of NFL games that yield similar results. But 5 Seasons of NFL is only about 750 Games. At what point am I allowed to feel comfortable that I have reasonable chance of success?
The tested resulted from previous seasons, were those forecasts made using formulas based on data ONLY from years before that? It sounds an awful lot like formulas based on in-sample data that's jiggered to produce a good-looking retrodictive result. But that doesn't at all mean it's predictive.
Absolutely understand your thoughts. I would say this that yes the formulas are tweaked to optimize results but the formulas are constructed with a lot of analysis and thought after watching tons and tons of games. So though some of the results are tweaked after the fact, that doesn't mean they won't be predictive. Dave
Guy loaded with machine guns comes to a broad side of the barn and randomly sprays it with a thousands of bullets.
Then he watches his results for a long time and very thoughtfully picks the largest clusters of the bullet wholes produced by his random shooting.
Then he takes a paint and draws bulls-eyes centered around the largest clusters.
Then he proudly declares himself the best sharpshooter in his country.
Is he?
I doubt it!
It seems to me, that his specific hypothesis ( "I am the best sharpshooter in a country and will hit the center of bulls-eye more often then anyone in the country")
should be confirmed by data (results of the firing over existing bulls-eye).
Collecting data first and then trying to invent specific hypothesis (in your case it would be "predictive formula") based on already collected data is not exactly
the smartest approach. In fact, it seems to me like an excellent way to get in a poor house.
Bottom line.
You can watch your past data all you want trying to discover hidden jams of undiscovered predictive patterns.
This is, by far, one of the favorite pass times of the human brain. Science even has a term for this - apophenia.
But once you done having this fun and think you got your golden nugget, this gut feeling of yours must be independently confirmed to the point of statistical significance by NEW, CLEAN set of data WITH NO TWEAKING ALLOWED. NEVER. EVER.
|
|
|
This post was nominated 1
time
. To view the nominated thread please click here.
People who nominated: MikeTizzy
|
HU...you beat me up pretty good with my posts about college basketball. My database was a little more extensive than my database regarding college football. There are obviously 3x the teams in college basketball and they more games than in college football which really puts it at 6x the amount of games tracked. With that being said, I had no proof that I was going to be successful in 2012-2103 except for all my past analysis. At some point, if the analysis is pure and the data is pure, you have to be trending toward something. I was successful. I bet no more than 3% on any given game and was up 300-400% over the course of the season in college basketball. Here I enter this college football season with the same mind set. I think you might misunderstand my tweaking. My basic premises remain the same. It is all about the weighting of the those parameters etc... You called me a data miner last year and I had to look up the definition. At some level, maybe I was but don't we all start that way until we figure out the relationships. So back to my original question? Do positive results over 3 seasons 1500 Games total mean I have something? For NFL, I have 5 seasons but I only have 750 games, but the results are solid. Does that mean I have something? Further more, College Basketball worked and yielded money this past season. So I now have 2 seasons of tracking and one season that it actually worked. At what point, can you believe you will be successful. As for no tweaking... I personally believe you are dead wrong. Your program should be constantly evolving and getting better... at least that is my plan. I think I have one down....with 2-3 well on their way. NBA will be ready 2014. MLB should be ready next season but we will see.
HU needs to get laid. S/he seems to get a woody at the idea of ripping someone who has something to say
I have no desire what so ever to argue with your personal beliefs.Originally Posted by usma1992
I just want you to know that you are not arguing this point with me.
In a nut shell, you are arguing with a scientific method.
I would bet on a side of scientific method (vs. your understanding of how things supposed to work in this world) twice a day and ten times on Sunday.
I'm not ripping anybody ever.
I'm just trying to help people who need help to start finally say something other than complete nonsense.
By definition, it will always be unpleasant and never cheering. Sorry, nothing I can do about that.
But it sure as shit can be very beneficial if considered carefully and thoughtfully.
Hu...I understand your thoughts...
and I appreciate them also. Essentially, you are saying that until the model is forward tested with good results without touching it than it is not worth anything. The problem I have is that your statement is that it is obvious and very lacking of much value. Many of the scientists do exactly what I am doing. Wall Street firms create models based on past performance all the time. Yet each mutual fund at the bottom say "future performance is not indicative of past performance".... of course it isn't. But, I always hate those same analysts that wait for a stock to go from $2-$6 and than put a strong buy recommendation on it. Where were they when it was $2? My point--- I have completely submerged myself into my computer programs for 3 Years. One program has actually finally delivered me not only money but a great performance... I personally don't believe that that means it works. I have 2 closely behind...I think. What I am trying to predict ... is when they will finally or have a realistic chance of being successful... Saying that they aren't successful until they are successful really isn't helpful or new. I am not arguing with you either. But from a scientific/statistical view point... Is 3 Years and 1500 games a realistic sample size? I guess I will find out shortly. Dave
Usma something should try creating your formula entirely based off say 1000 games of that sample and then test on the remaining 500. No offense but does sound like your fitting your data to your sample and expecting those type of results. This eventually could lead to negative results if you just continue modifying without a realistic reason why. A reason to adjust would be a rule change in a sport which effects scoring. I wish you nothing but the best and hope to see your picks to see how your model does.
Correcto mundo!
Sorry you feel that way.The problem I have is that your statement is that it is obvious and very lacking of much value.
Examples please.Many of the scientists do exactly what I am doing.
That's b/c in reality, they don't really give a shit how good those models are.Wall Street firms create models based on past performance all the time.
You see, they don't really make money from predicting future stock prices.
Their collective record in prediction future stock prices sucks balls.
What they do make money from is charging big fat fees for something that can be obtained for free.
They do it by confusing people to believing that they know something that justifies big fat fees.
They also very good in manipulating the system to the point that when their bullshit "models" result in monstrous loses
they get their money back from the same morons who pay them big fat fees for nothing.
Its ia great job if you can get it! Can you?
If not then you should not take a play from their playbook.
You should think for yourself.
Long and hard.
You don't have a 3-year, 1500-game sample, though. What you have is three years of back data and, from how it sounds, some formulas that were created to fit that data. Fully retrodictive. In effect, you have a sample size of zero when it comes to testing for predictive value.
Unfortunately, this is absolutely true. Have been there myself.
This scenario is not completely correct...
I did create a program using just 2 years and began the testing piece on year three. Unfortunately, I worked through all of year 3 with reasonable results but I had a major formula issue that I was grabbing the wrong teams defense so all of my fine tuning was using the wrong defense for the home team. It took so much effort to do it in the first place... when I fixed it... I did just decide to optimize the results across all 3 years. In my defense, It worked the last 4 weeks of 2012 after the optimization. In addition, I have attacked college basketball with the same discipline, desire, thoroughness etc... and even though it wasn't forward looking...it worked almost perfectly... I guess we will see come Week 5... Thanks for your input.
I am not sure this is accurate. Stock firms are in fact measured against indexes such as the S&P500, and their performance relative to that index represents how good their models are(were). Companies with a record of outperforming the index will very quickly be managing tons of major portfolios. Companies that talk b.s. about their models but then under-perform against the index won't get any clients, and won't make a dime.
That being said, I do agree with your earlier point... just as in finance, a sports data model is only valuable if it only considers previous data and then predicts future data. Tweaking/fitting a model to an entire historical data set, and then using that model on the same data set to show its predictive power within the set is not useful.
MMM heard of variance ?
See the Sharpe ratios of these companies ... its been a freeroll and will be
Yeah, sure. And I'll tell you why.
It is much easier to be unsure when you cherry pick bits and pieces out of the whole argument that are easy to be unsure about.
Little tricks like that (creating distorted version of the argument which is easier to attack instead of dealing with a whole argument)
are called "straw man" fallacy.
Lets not fall for this crap!
Here is the whole thing again including a few words you conveniently left out (i bolded them out just to be sure you see them this time)
Would you like to talk about it some more?That's b/c in reality, they don't really give a shit how good those models are.
You see, they don't really make money from predicting future stock prices.
Their collective record in predicting future stock prices sucks balls.
What they do make money from is charging big fat fees for something that can be obtained for free.
They do it by confusing people into believing that they know something that justifies big fat fees.
They also very good in manipulating the system to the point that when their bullshit "models" result in monstrous loses
they get their money back from the same morons who pay them big fat fees for nothing.
Last edited by hutennis; 08-30-13 at 06:21 AM.
|
|
^^ hutennis is a winning sports bettor. Listen to this guy, he knows.
Post your plays so we can track them here.
So basically if I had a model (and I do) that was tested on some past data just to help create it, but then FORWARD tested on a couple years of data using the exact same, unchanging parameters it was built upon, then the results are completely valid and the kind of scientific results hutennis loves? Hmmm interesting. If true, then I feel a lot better about my model
This is a basic of any form of modelling: you create on one set of data, you test on another. If you test on the same set you create from, you're no longer predicting, you're just tuning a model to get closest to the results that have already happened.
For example, if I create a model based on tennis results over the last decade, and then test on 2012+, ie only 1.5 years overlap out of 10, the results iirc were something like 59%, with 1500 bets.
1/3 Kelly, 1500 bets at 59%, starting bankroll $10. Median outcome is $146k, for a single year. Hrrrrrrm.
Good general rule: if the backtest is showing >54% over any genuinely sized sample, you've probably made an error. >55%, it's a miracle if not. >60%? The bookies would be bankrupt in a few days if their prices were that inaccurate.
Last edited by HeeeHAWWWW; 08-30-13 at 05:09 PM.
White swan vs black swan much? thats how many in wall-street go broke as well.
i made it at the very beginning of 2012, then used it on the remaining 2/3rds of MLB season, then used it on the entire 2013 MLB season. So I never had overlapping data with the creation or the testing
I do understand the concept everyone continues to speak to. However, I did that very thing with my NFL program. I tested based on 2009-2012 and than created 2008 from scratch and using the same parameters and weightings looked to see if it worked in 2008 and it did. I will be adding 2007 NFL season next week after I further fine tune and will be testing but I was unable to do that for college football. Let me restate, I began looking at 2010 with that intent. I had to reconstruct 2010 college football season from scratch and it took me 2 weeks... because I had to go game by game... Brutal. But, I had a coding error that I did not recognize until after 2010 was complete and I didn't want to re-spend the time testing it again. I heard this same argument last year prior to my 12-13 basketball season. I doesn't work until it works...until it is tested on a test set. Well, it worked... and worked extremely well. My belief is that if you have established the correct relationships and you continue to fine tune that is not only appropriate... it is mandatory as the game evolves and things change. I have what I believe very solid relationships between what is important and what is not... but again...we will see week 5 and beyond.
Couple of years of data means nothing. 2/3 of the season plus entire season means nothing.i made it at the very beginning of 2012, then used it on the remaining 2/3rds of MLB season, then used it on the entire 2013 MLB season. So I never had overlapping data with the creation or the testing
Those are pretty arbitrary parameters.
Statistical significance is the only measure. And that may or may not be obtain by calendar.
Moreover even forward testing is not immune at all to all kinds of BS.
Human nature is such that people will almost inevitably start compromising even forward testing in order
to confirm not what it is but what they want it to be.
The best thing would be betting according to your rules in real time. All be it betting only $1.
This way you have real time odds (not a wishful thinking odds) and you have hard, independently kept records.
You do that, reach statistical significance ( however long it might take) and then you can feel a lot better about your model
I did bet my plays every day, I did pull the odds daily manually for every play. I never changed how the model works, and all my bets were made in real time. How long do you think it would take to reach statistical significance if I continue to place bets in real time and continue to make them according to an unchanging model?
Very simple yes or no question. When testing week 10 of 2012, was any part of your formula created from games played after week 9 of 2012? If yes, your testing is invalid.
Think about it...it's week 1 of 2013. You can't use data from week 1 of 2013 to "tweak" your formula, as it hasn't happened yet. The purpose of your testing is to simulate actual conditions. Incorporating information that didn't yet exist when the (simulated) bets were made gives you an advantage that you can't possibly have in real life.
It is not about "how long"
It is about numbers.
For example
After 500 plays with implied probability of 50% you have correctly predicted 270 wins (20 more than you are untitled to by simply the chance). Excellent 54% win rate!!!!
Is it statistically significant? I would not say so. You chance to get a result like this by simply getting lucky is ~4%.
There is nothing significant about that. People get lucky to the tune of 4% all the time. It is trivial.
Now, if you were able to sustain the same record after 1000 plays (540 wins) your chance to get there by luck alone drops to 0.6%
which is a lot more interesting.
And if you would to keep your winning pace after 1500 plays (810 wins) your luck trash hold would drop to 0.1%
Now you can certainly start feeling pretty good about the whole thing.
So how long would it take? You tell me.
It depends on what you deal with , I guess
But yet again, you can never be certain.
After all, 1 in 300 000 000 event happens in United States every day.
Last edited by hutennis; 08-30-13 at 10:17 PM.
Well my sample size on totals is 2114 games and I hit 53.17%, of that 2114 game sample, there were 789 that I deemed "strong bets", they hit 53.74% (really close to the 54% you gave the example of 270/500). So "strong bets" hit 53.74% over 789 game sample, and the rest hit 52.83% over the remaining 1325 game sample. All games in my sample were tested AFTER my model was finalized and numbers and lines gathered daily as the season progressed, so none were used in the creation of these results. So based on what you've said, the chance that luck helped me get almost 54% over my sample is under or around 1%?
No, it is much less.
If everything above is correct, including the assumption that true win probability in your totals is 50%
your chance of simply being lucky is 0.2%.
Congratulation! You are one of the best handicappers in the world!!!
It makes it even more impressive that you got to be that good without any clue on what statistical significance means.
Looks like math is truly overrated!
Last edited by hutennis; 08-30-13 at 10:59 PM.
Lol I have a decent idea of what it means, I've learned from reading around the internet and forums, so I've educated myself on that. I just wanted to know that you agree with what I believe is a statistically significant model. But it's good that we're on the same page anyway at least when it comes to luckiness of models and what data is significant.
You never told me what you believe is a statistically significant model. You only asked questions.
So I have no idea whether we are on a same page or not.
But I gotta tell you, if it's only took you reading around the internet and forums to get to the point that only 0.2% of the handicappers in a world are better than you are - you are certainly one special dude.
The two words dont go together
This is true, but the sample size issue can be greatly alleviated (at least in a high liquidity market) by incorporating market information.
gamblingisfun- here is my advice. Take your entire subset of games that you actually bet on (which were indicated by your model) and track your closing line value. If you are not ahead of the closer at least 60% of the time, your results are almost certainly a fluke, even with a sample size > 2000. A simple binomial test for significance can give you a p value immediately. And the nice thing about that test is that even with a small sample size of ~100, you can get a very good idea of whether you are on the right side consistently. I'd be interested in hearing the results either way.
Reply With Quote
MikeTizzy gave hutennis 2 SBR Point(s) for this post.
