Free winning NBA spread picks

[nomadcrypto]

The main reason for me posting these free picks is to establish a public history for my model and to establish a history of being able to successfully tackle hard ML problems such as spread betting. I may be interested in selling this model to one person/group once I’ve established a public record of the accuracy. The only reason I would do this is to build my initial sports betting bankroll and to help fund further research into other sports.  Right now I’ve only developed a winning spread betting model for NBA games but I will be working on under/over bets next and then moving on to MLB. I will also be working on a classification system for the picks as well as working on better calibrating the model’s output. Meaning that some picks are better than others in terms of accuracy and profitability.

To be as transparent as possible I will post all picks prior to the events and will post them at pastebin.com where you can easily see the date/time of the paste as well as see if the post has been edited. I will post picks here for 300x games or more and will update the first post in this thread with the current accuracy rate of all picks and will reserve the second for the current picks and link to the pastebin paste containing those picks. The third post will be reserved for links to previous picks and the results for that day.

For those familiar with machine learning I used a XGBoost Classifier for binary classification to predict if the home team will cover the spread or not. I used Stratified KFold(5x folds) to estimate the model’s performance and held out 2500x data points for validation data and back testing. The results(including accuracy, recall, precision, f1 scores and backtest results) are posted here https://pastebin.com/mz5VN07Z . Given the difficulty of the task these results are extremely promising. The backtest/validation data accuracy was 63.84% and the bankroll in back testing grew from $500 to $2.5 million over 2500x games(roughly 2 seasons/years). I capped the max bet at $5k in back testing and bet 10% of the bankroll on each bet. I do not recommend that you bet 10% of your bankroll on these picks.

If you should win some money and would like to spread some love my addresses for tips are:
Btc:
1LcEgr8JcKBFpynox6gKvonUsV1GSxnDUd
Eth:
0xcda88e2ccd5a9a45a678390637610dba3a48e6d7
Ltc:
Ld1vd12DvWfR36KRXHdm3cRApfBesyoXDL



Total picks: 48.0
Wins: 22.0
Loses: 26.0
Accuracy: 45.83%
Profit: -33.82
ROI: -40.99%





Edit:
Last night I rebuilt the dataset and retrained the model to use only pinnicale odds, built a rating system for the predictions and worked out the optimal(I may be wrong on this being "optimal") amount to bet on each rating type(bronze, silver, gold). In back testing the accuracy for each rating was:

total 2500 games in the back test/validation data
bronze rating is 55.87% accurate over 1235.0 sample
silver rating is 66.20% accurate over 722.0 sample
gold rating is 78.82% accurate over 543.0 sample

With the newly trained model there was a rather large improvement in the profitablilty of the model. The back testing bankroll increased by 16% over the previous model and the accuracy rate increased slightly from 63.84% to 64.08%. With the new bet sizing strategy the profit increased by 16.8% but also had lower volatility and that seems to be simply because we bet less on the bronze and more on the silver/gold. I used a "unit size" of 5% of the bankroll and bet the following:

bronze: 1 unit
silver:   2.5 units
gold:     5 units



From 2018-12-02 on all profit estimates will be based on those bet sizes. Each prediction will include the rating type as well as suggested units to bet on each game. Feel free to adjust the percentage of your bankroll that 1 unit repersents based on your acceptable risk level. Typically a unit size ranges from 1% to 5% of your bankroll. I am personally live testing with a 3% unit size.

Edit
The bet sizes may not be optimal and variance could account for some of the increased accuracy rate of the "bins". Be careful and do not bet more than you can afford to lose on these picks until we get more picks. Also, variance could account for the model having not shown a profit so far(as of the morning of 2018-12-03)

Edit 2018-12-07
After not getting the results anywhere near the back test results I started looking into it and I found that I made a fairly large mistake in how I selected the team players and got their stats for the live games. In the training data there are only players that played in each game which accounts for injuries. In the live data I had not built a system to account for players not able to play due to injuries or being out of a game for any reason. Being a fairly lazy programer I just used the players who played in the previous game for that team. This is what I suspect has caused the sharp difference between the past results and the live results. I've built a scraper for player injuries and am in the process of rebuilding the dataset/retraining the model. Hopefully I'll be able to have it ready for today's games.

[nomadcrypto]

picks for Friday 2018-12-07 https://pastebin.com/urHWBwZM
https://bitcointalk.org/index.php?topic=5077782.msg48332339#msg48332339
===================================================================
Toronto Raptors vs Brooklyn Nets on  2018-12-07
Rating: gold
Bet  5.0 unit(s) on  Toronto Raptors  at  -9.0
american odds -102 decimal odds 1.98 implied odds 50.50%
===================================================================
===================================================================
Oklahoma City Thunder vs Chicago Bulls on  2018-12-07
Rating: gold
Bet  5.0 unit(s) on  Oklahoma City Thunder  at  -8.5
american odds 103 decimal odds 2.03 implied odds 49.26%
===================================================================
===================================================================
Denver Nuggets vs Charlotte Hornets on  2018-12-07
Rating: silver
Bet  2.5 unit(s) on  Charlotte Hornets  at  -1.5
american odds -107 decimal odds 1.93 implied odds 51.69%
===================================================================
===================================================================
Sacramento Kings vs Cleveland Cavaliers on  2018-12-07 
Rating: bronze
Bet  1.0 unit(s) on  Sacramento Kings  at  -2.5
american odds -106 decimal odds 1.94 implied odds 51.46%
===================================================================
===================================================================
Philadelphia 76ers vs Detroit Pistons on  2018-12-07
Rating: bronze
Bet  1.0 unit(s) on  Detroit Pistons  at  -1.5
american odds -107 decimal odds 1.93 implied odds 51.69%
===================================================================
===================================================================
Golden State Warriors vs Milwaukee Bucks on  2018-12-07
Rating: silver
Bet  2.5 unit(s) on  Milwaukee Bucks  at  -1.5
american odds -110 decimal odds 1.91 implied odds 52.38%
===================================================================
===================================================================
Memphis Grizzlies vs New Orleans Pelicans on  2018-12-07
Rating: silver
Bet  2.5 unit(s) on  New Orleans Pelicans  at  -5.5
american odds -101 decimal odds 1.99 implied odds 50.25%
===================================================================
===================================================================
Indiana Pacers vs Orlando Magic on  2018-12-07
Rating: bronze
Bet  1.0 unit(s) on  Orlando Magic  at  -2.5
american odds -106 decimal odds 1.94 implied odds 51.46%
===================================================================
===================================================================
Miami Heat vs Phoenix Suns on  2018-12-07
Rating: gold
Bet  5.0 unit(s) on  Miami Heat  at  -8.5
american odds -104 decimal odds 1.96 implied odds 50.98%
===================================================================
===================================================================
Los Angeles Lakers vs San Antonio Spurs on  2018-12-07 
Rating: bronze
Bet  1.0 unit(s) on  San Antonio Spurs  at  1.0
american odds -105 decimal odds 1.95 implied odds 51.22%
===================================================================

[nomadcrypto]

Previous picks before model Update on 2018-12-07

Date/link	Total	Wins	Loses	Accuracy	Units	ROI
11-29-18	3	2	1	66%	+0.78	25.95%
11-30-18	8	6	2	75%	+3.42	42.73%
12-01-18	7	3	4	43%	-1.08	-15.46%
12-02-18	6	2	4	33%	-9.09	-70%
12-03-18	7	5	2	71%	4.04	35%
12-04-18	5	1	4	20%	-8.09	-69%
12-05-18	9	4	5	44%	-14.73	-65%
12-06-18	3	0	3	0%	-11	-100%

Picks after model updated for injured players/players marked as "OUT" for game

Date/link	Total	Wins	Loses	Accuracy	Units	ROI
12-07-18	10	0	0	0	0	0

[DarkStar_]

Does your model state by what margin the team will win? Most models work based on their own predicted result (example: GSW vs TOR, TOR is expected to win by 4.5 points so you would take GSW +9.5). Or, does it only predict whether they will cover the spread and not the margin?

[nomadcrypto]

Does your model state by what margin the team will win? Most models work based on their own predicted result (example: GSW vs TOR, TOR is expected to win by 4.5 points so you would take GSW +9.5). Or, does it only predict whether they will cover the spread and not the margin?

The model is a binary classifier which outputs either 0 or 1(or the probability that it will be 0 or 1). Part of the input is the spread, odds given for the home/away spreads, home/away moneyline, as well as various stats for each team and player.

I am going to work on points/point differential next.

[nomadcrypto]

We had a pretty decent night last night and ended up 2 and 1 win/loses. Total ROI was 25%. It was only 3 games so that doesn't really mean anything. I've updated the pick post with the most recent picks. Today we've got 8x picks

[mOgliE]

Man if your model really manages to go from 500$ to 2.5millions in 2 years you're in for a big fucking nobel prize xD

Anyway it changes from the "tipsters" who just gives out their opinion and label it as expertise.

[nomadcrypto]

Man if your model really manages to go from 500$ to 2.5millions in 2 years you're in for a big fucking nobel prize xD

I hope the future results are at least somewhat close to the back testing results. I would be ecstatic if it got close to a 58% win rate. Hell 55% would be better than most professional sports bettors.

Anyway it changes from the "tipsters" who just gives out their opinion and label it as expertise.

I do not claim to have any expertise when it comes to sports betting. I've never even watched a full basketball game, haha. I do, however, have a lot of experince with machine learning and with tackling hard problems with ml models.

[mOgliE]

Man if your model really manages to go from 500$ to 2.5millions in 2 years you're in for a big fucking nobel prize xD

I hope the future results are at least somewhat close to the back testing results. I would be ecstatic if it got close to a 58% win rate. Hell 55% would be better than most professional sports bettors.

Anyway it changes from the "tipsters" who just gives out their opinion and label it as expertise.

I do not claim to have any expertise when it comes to sports betting. I've never even watched a full basketball game, haha. I do, however, have a lot of experince with machine learning and with tackling hard problems with ml models.

That's why it's interesting.
We saw that tipsters and experts aren't efficient. Maybe machine learning can be. Would you be willing to discuss this a bit? Do you have telegram maybe?
I think it's a shame that such a great initiative is only a small thread on the forum. We might make something big out of it.

[nomadcrypto]

Last night was a really good night. Overall accuracy was 75% and roi was 42%. I've updated the pick post with 5x picks for today

[SyGambler]

interesting thread , I will be waiting till you reach 100 game then I will probably jump in
what are the bookies that you are using ? cause AFAIK only pinnacle has over 5K limits and it's the only bookie that won't limit you if you win so it would be great if you quote their odds

[nomadcrypto]

interesting thread , I will be waiting till you reach 100 game then I will probably jump in
what are the bookies that you are using ? cause AFAIK only pinnacle has over 5K limits and it's the only bookie that won't limit you if you win so it would be great if you quote their odds

My dataset contains odds from Pinnicale,5Dimes,BookMaker,BetOnline,Bovada,Heritage,Intertops,YouWager,JustBet and SportsBetting but to be honest I'm not really sure which books those odds are coming from because of the way I originally coded the model. It sounds fairly short sighted but in the first couple runs I simply picked the first spread/moneyline for each game. I'm in the process of fixing the training/testing/validation data to only use pinnicale. Sadly I can't bet at pinnicale because I'm based in the US but it makes total sense to train on their data since they're the sharpest book/don't limit winners. BetOnline is also pretty well known for being relaxed on limiting winning bettors and is basically the best option for US bettors.

edit
Also, betonline caps most NBA games to 10k

[nomadcrypto]

Here are the pinnicale specific predictions with the most recent spreads/odds

===================================================================
Toronto Raptors vs Cleveland Cavaliers on  2018-12-01
Bet on  Cleveland Cavaliers  at  13.5
american odds 100 decimal odds 2.00 implied odds 50.00%
===================================================================
===================================================================
Golden State Warriors vs Detroit Pistons on  2018-12-01
Bet on  Golden State Warriors  at  -4.5
american odds -107 decimal odds 1.93 implied odds 51.69%
===================================================================
===================================================================
Chicago Bulls vs Houston Rockets on  2018-12-01
Bet on  Chicago Bulls  at  12.0
american odds -110 decimal odds 1.91 implied odds 52.38%
===================================================================
===================================================================
Boston Celtics vs Minnesota Timberwolves on  2018-12-01
Bet on  Boston Celtics  at  2.5
american odds -109 decimal odds 1.92 implied odds 52.15%
===================================================================
===================================================================
Milwaukee Bucks vs New York Knicks on  2018-12-01
Bet on  New York Knicks  at  8.0
american odds 100 decimal odds 2.00 implied odds 50.00%
===================================================================
===================================================================
Indiana Pacers vs Sacramento Kings on  2018-12-01
Bet on  Indiana Pacers  at  -2.0
american odds -103 decimal odds 1.97 implied odds 50.74%
===================================================================
===================================================================
Brooklyn Nets vs Washington Wizards on  2018-12-01
Bet on  Brooklyn Nets  at  6.5
american odds -103 decimal odds 1.97 implied odds 50.74%
===================================================================

From tomorrow on all predictions will use Pinnicale spreads/odds

[nomadcrypto]

Last night wasn't all that great of a night. The total accuracy was just 43%(3 wins 4 loses) - 1x unit loss for the night. If we had bet using the new bet sizing/rating system the night would have ended up profitable(4.2% ROI) even with the low accuracy rating. Even with that losing night we're up 3.11 units total for all predictions. I made several updates to the model as well as implemented a rating system for the picks. There are 3x ratings: Bronze, Silver and Gold. As the names would suggest gold and silver have a higher accuracy rating in back test. I rebuilt the dataset to only use pinnacle odds and retrained the model on the new dataset. This increased the overall accuracy from 63.84% to 64.08% and increased the profitability over the 2500x games in the validation dataset by 16%. When I implemented bet sizing based on the rating the profitability increased by 16.8% and decreased volatility drastically. Please see the OP for more information and see the pick post for today's 6x picks

[BTCevo]

I just saw this thread and it is really attractive in my point of view. So about all these picks that you made, you divide it into 3 categoris, gold silver and bronze? And you are saying the accuracy is higher, are you saying that you have better odds and good prediction? Let me tell you this way,
Gold, lower odds, high percentage
Silver, normal odds, average percentage
Bronze, high odds, low percentage

Do you mean it this way? And you are saying about the machine, what is this about? You are not betting using your skills

[nomadcrypto]

I just saw this thread and it is really attractive in my point of view. So about all these picks that you made, you divide it into 3 categoris, gold silver and bronze? And you are saying the accuracy is higher, are you saying that you have better odds and good prediction? Let me tell you this way,
Gold, lower odds, high percentage
Silver, normal odds, average percentage
Bronze, high odds, low percentage

Do you mean it this way? And you are saying about the machine, what is this about? You are not betting using your skills

Out of 2500 games I held out for back testing the results for each class is:

bronze rating is 55.87% accurate over 1235.0 sample
silver rating is 66.20% accurate over 722.0 sample
gold rating is 78.82% accurate over 543.0 sample

If the future results are like the past results I would expect 7.8 out of 10 gold picks to be winners on average, 6.6 out of 10 silver,  and 5.5 bronze. The prediction classes are based on the accuracy rate of the model's output.

I am not using any handicaping skills to make these predictions. I am using a machine learning model. I would caution against blindly trusting this until it has a proven track record though. A blind folded chimp could pick 50% winners and with variance that could swing pretty wildly over a small sample size. Even with the extremely good back test results we could see some pretty big swings up or down over a small sample size. The past two nights have been losing nights for example  As another more experienced gambler stated he is waiting until it has predicted 100x matches and has a positive net result. I am personally betting each one of these games but I'm betting an amount I can afford to lose and I think the back test results are statistically significant.

edit
Here is a chart showing how wildly results can vary with a 55% accuracy rate

[nomadcrypto]

I wanted to give a quick update since the total predictions passed the minimum sample size for a confidence interval estimate. So far the model is performing within the 95% confidence interval but since the sample size is so small it is also within the 95% confidence interval for random guessing too. Below are estimates of what we could expect to see at our 31(current sample size), 50, 100, 200, and 300 sample sizes. The actual results may vary since the estimates do not account for the accuracy and proportion of each rating type just for the overall accuracy on the test and validation data.  

Bounds for 64% accuracy over sample size 31 upper 80.90% lower 47.10%
Bounds for 50% accuracy over sample size 31 upper 67.60% lower 32.40%

Bounds for 64% accuracy over sample size 50 upper 77.30% lower 50.70%
Bounds for 50% accuracy over sample size 50 upper 63.86% lower 36.14%

Bounds for 64% accuracy over sample size 100 upper 73.41% lower 54.59%
Bounds for 50% accuracy over sample size 100 upper 59.80% lower 40.20%

Bounds for 64% accuracy over sample size 200 upper 70.65% lower 57.35%
Bounds for 50% accuracy over sample size 200 upper 56.93% lower 43.07%

Bounds for 64% accuracy over sample size 300 upper 69.43% lower 58.57%
Bounds for 50% accuracy over sample size 300 upper 55.66% lower 44.34%

After 200x games I will be extremely confident that the future results will be similar to the validation/test data results.

This morning I did what is called "target shuffling" which is where I shuffled the target randomly on the training and test datasets and then trained the model on that data. The reason for doing this is to see how much better this model does than random. Out of the 100x iterations this model outperformed all of the random models. It out performed the best random model by 11%. Which given the difficulty of this task is a pretty significant result. To me this result as well as the consistency of the test and validation data results over decent sample sizes proves that this model close to, or is, state of the art for this task. I've not seen or even heard of a model that performs this well on this task.

[mosfellasbar]

less talking more picks please 

[mirakal]

I wanna try your picks and see if it's working.
I love NBA and I love betting so I think if this is gonna be consistent, it can help me win in the long run.
If possible, please do post on a daily basis because I love to bet different matches in a day, I am not contended with one bet only.

[nomadcrypto]

I wanna try your picks and see if it's working.
I love NBA and I love betting so I think if this is gonna be consistent, it can help me win in the long run.
If possible, please do post on a daily basis because I love to bet different matches in a day, I am not contended with one bet only.

Generally when I post picks they are every game that pinnicale has both spreads and moneylines posted for.

I found out that I introduced a leak of a sorts. This is mostly because I'm not a sports fan, just a data science nerd. The reason the live test results were so far off from the back test results is that I built the model/dataset using players that actually played in the game versus the team's roster. NBA teams have a roster size of 12-15 active players but generally 10 or less will play in any given game. For example, the Brooklyn Nets have a total roster of 17 players, 15 of which are active but only 10 players played in the game on Friday.  To successfully handicap a game you must look at each team's and player's home/away specific stats which is what my model does. The first several days I simply used the players for the previous game but that didn't account for injuries or player rotation. I've since built a scraper that gathers injury data but that still doesn't give me an accurate estimate of who will play in any given game. I'm in the process of building a model to predict the starting lineup for each team and then will use that lineup for a newly trained model using just the starting players.

TLDR: I'm not posting picks for a few days while I fix the leak.