Replies: 47 comments 37 replies
-
|
Some feature Ideas: |
Beta Was this translation helpful? Give feedback.
-
|
I also want to add season games played and season games won as a team input features but the api does not have that data. maybe it would be easy though if i sort games by date first, to be honest i kinda think they are already sorted by date. but then the problem still becomes how to get new values for future games, i don't want to rely on added up values. |
Beta Was this translation helpful? Give feedback.
-
|
I also think it would be really great to eventually get historical spread data setup and start predicting error in spread. |
Beta Was this translation helpful? Give feedback.
-
|
this is next up! |
Beta Was this translation helpful? Give feedback.
-
|
Got this done today, really improved performance, getting up to 70% accuracy now! I will try and push updates tomorrow. Also I think moving players per team from 9 to 7 helped. |
Beta Was this translation helpful? Give feedback.
-
|
Also interesting data it appears as though point spread's have gotten less accurate over time. with a lowest accuracy of only 62% during 2021 |
Beta Was this translation helpful? Give feedback.
-
|
Okay pushed updates for players per team, and added point spread as a input feature. Also the training and testing datasets are now split up in csv folder. Getting much better results now with point spread as an input feature. I think we are actually producing positive expected value now or at-least very close. I am working on writing some better metrics to track percent correct with spread applied. |
Beta Was this translation helpful? Give feedback.
-
|
Today i got new data working both with historical and future games from nba_api python package. The main new data input features are games played, games won, and games lost on a daily level(per season). I will update repo either tonight or tomorrow depending on things. I want to add point spread as input feature getting us one step closer to spread error predictions. This is really great though we have a new data source! |
Beta Was this translation helpful? Give feedback.
-
|
this was bad idea. nba_api python package is pretty bad. I could'nt get it to return responses in under 30 seconds and timing out every other try, tried on server too its not my ip. Apparently they have alot of trouble with nba.com blocking request. many issues on there repo about it. I tried all kinds of ways to get around but i gave up and decided to use a different api instead. thats when i found this one: |
Beta Was this translation helpful? Give feedback.
-
|
Okay gona be tomorow before update. I think im going to lower playersPerTeam from 9 to 7. I wanted to write down this feature idea before i forget: |
Beta Was this translation helpful? Give feedback.
-
|
Done, I really like this feature too. i think it adds alot of value, lets people see all the data of the games predicted, and can do alot with csv files like graphs and equations.... |
Beta Was this translation helpful? Give feedback.
-
|
It would be great to have a feature that would mark players as injured. Making the players name in edit stats clickable to mark player as injured and get next best player. I think this would be difficult to implement. I guess i could make a url edit/remove// then basically add that player to a blacklist and re run best players. This seems difficult. if i do make this feature it probably wont be soon. |
Beta Was this translation helpful? Give feedback.
-
|
Noticed a few players on miss matched teams. or atleast will be when trades go through. 2023 NBA trade deadline buzz: Feb. 9. Kyrie Irving is officially leaving the Brooklyn Nets to join the Dallas Mavericks. |
Beta Was this translation helpful? Give feedback.
-
|
remember to update rosters tomorrow or atleast before games start again |
Beta Was this translation helpful? Give feedback.
-
|
Research xgboost maybe could be useful |
Beta Was this translation helpful? Give feedback.
-
|
The api for todays games is not updating properly, the api endpoints probably changed. Hopefully it will be fixed soon. If not i will submit issue on the repo. They have been good about getting things updated in the past. Here is example api url without updated scores. |
Beta Was this translation helpful? Give feedback.
-
|
Looks like balldontlie api is still broken. I submitted issue but no one has responded, maybe someone can comment on my issue and be like same, and then they see it. Anyways Im gona try to build out the export games function tomorrow, and add more metrics like expected value, and spread accuracy. |
Beta Was this translation helpful? Give feedback.
-
|
Think i found the culprit of no updating scores here: committed yesterdays right when it stoped working.... |
Beta Was this translation helpful? Give feedback.
-
|
Messing around training the model. Im finding that if you overfit the model really bad (over train, way too many epochs). Then it produces better expected value. Ive got up to 55%. This isnt really a good solution though as the model isnt working as it should and produces extreme results too often like teams winning by 100+ points. I want to try and and get win/loss streak features soon. |
Beta Was this translation helpful? Give feedback.
-
|
Talked to the owner of the games won/loss api we are using. Cool guy, think he is gona add streaks as-well, which would be amazing. still would have to find source for historical data streaks, or add them up using game dates. I know before I didnt want to add them up, but thats for future games. If an api keeps the current count, I have no problem adding up historical data because it will only need to be done once, not every-time you need the current count. This would be cool. I also want to work on accounting for mid season trades, and injured players. I think this can be done by using the players actual playtime to determine if they played....then if they played use there season averages as usual like average min. |
Beta Was this translation helpful? Give feedback.
-
|
Added win/loss streak to web app. Need to work on creating historical streaks with dataset so we can start using streaks as an input feature. Also looking into injuries today. |
Beta Was this translation helpful? Give feedback.
-
|
Any one have ideas on how the win/loss/streaks should be displayed? if i put dashes or slashes it looks like a date, and is kinda confusing if i put : in-between then it looks like a time. like 10/10/2 or 10:10:2 or 10-10-2 they look like date times. Now there is just space in-between but i dont think its very clear.... any ideas?
|
Beta Was this translation helpful? Give feedback.
-
|
maybe two spaces! |
Beta Was this translation helpful? Give feedback.
-
|
Okay next up is more stats on graphs. also making a search player function to add and remove players from games. also going to at-least make little alert that detects if players are injured, gona be difficult to automatically match the right player and remove them so at first this will probably have to be done manually. also i wanted to experiment see what expected value would look like if we only count predictions with margins greater then like 1 point. |
Beta Was this translation helpful? Give feedback.
-
|
also the get data form cant get future or past spreads, only if the game is today. maybe i can improve this in future but to be honest i dont like that form and its way better just clicking on todays games, so dont want to spend alot of time on that. |
Beta Was this translation helpful? Give feedback.
-
|
Just realized that both team rosters and season averages can be updated at the same time. going to implement this soon make it easy to update saved stats and rosters. |
Beta Was this translation helpful? Give feedback.
-
|
Okay i added a bunch of stuff today. I havent pushed any the bigger changes yet but will try to update tonight (probably tomorrow). The updates already live include keeping scroll position for games, ive been trying to fix that for a while but it always caused other problems, this fix isnt perfect it kinda flickers wrong scroll position on first load, then updates. so i added a nice opacity transition on page load. kinda covers it i think. also kinda fixed the new game form, but honestly im considering just doing away with it. Unless i can save it by making it better, i was thinking it would be alot smoother to select date first, then be shown a list of games on that date, i could probably reuse the todays games function, this is good idea! Also the training script is updated to track 4 levels of expected values, 0 point up to 3 point margins, meaning it only counts games that have a large enough deviation from spread to prediction... only counting strong predictions against the spread..... This really creates some spectacular results, achieving upwards of 60% over the spread, thats plenty to cover fees and more!! So then i added the new metrics to the game django model and added some overall fields like percentages to the profiles model. Each game now has fields specifying if the game won over the spread, the exact margin amount, and then the 3 category margins of 1, 2, and 3 point margins. all the data is there and being updating properly now, im just not exactly sure how im gona integrate into the graphs and donut. I might delete the donut and just add like 5 bars to the other chart, not sure. when i figure out the charts, ill push the update with new metrics. |
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
Got this built today. Works really well and got rid of the old form! I will say there is no historical spread data still which really sucks, its would be easy to implement but the spread api only lets u get historical odds if you pay them 10$ a month.... you can still go into admin and input spread yourself. really happy how the date picker turned out tho! the metrics should be fixed too, tracking margins and wins against spread. might add those to the export function too. I want to work on the player lookup function i think it will go under the player data on the edit page. |
Beta Was this translation helpful? Give feedback.
-
|
API spread data is fixed! i also built a predict all button, this button docent remove any injured players, and exclusively use's model slot 0. Really easy to go back and predict a bunch of games now, and test models. I think im gona work on writing a document, call it theStrategy, explain how the project works, and ill try to specifically articulate things there. |
Beta Was this translation helpful? Give feedback.
-
|
Really happy with how the new model is performing, getting much better results with smaller layers. I think the swish activation was causing some overfitting as-well. Really easy to test models now with the predict all button, i made a get all scores function, but im not gona upload that yet, mainly because it makes too many api calls too quickly. Worked on making an actual homepage pretty happy with where it's at, displays todays games on an infinite carousel. Probably will see errors with that in the coming days. Check out the new homepage! https://nbadata.cloud/home/ |
Beta Was this translation helpful? Give feedback.
-
|
Home page looks really good now! still want to add more text, not sure how that will look. I made a search players page, shows results in a table. Each player has a detail view as-well. Maybe i could link the detail view on edit game view for each player that would be cool. For now the detail player view has 2 buttons, update season average stats, and update team. Very use-full in future when players are traded. |
Beta Was this translation helpful? Give feedback.
-
|
can now click on players name's in the game edit view, and redirects to player detail view. |
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
added team detail view, shows current roster. can now click on team logo image either on game edit view or player detail view, it will then redirect to team detail view. also player names are clickable in the team roster table. |
Beta Was this translation helpful? Give feedback.
-
|
https://nbadata.cloud/predict/team/MIL |
Beta Was this translation helpful? Give feedback.
-
|
Hopefully api will be back online soon. I build out a team list view, with all the teams and can click on them to go to team detail view. Also improved team detail view a bit, looks better now. |
Beta Was this translation helpful? Give feedback.
-
|
got alot done today, added 2021 and 2022 seasons to training dataset. also made a system so that games can be marked as bets, and then there is a separate page with bets and stats on games marked bets. Also i think the default model is kinda overfitting, been having good luck with really small models, 32 layer1 16 layer2. will update default model and push updates tomorrow. Also I redid the team data so we calculate and add up win/loss/streak, no longer using saved dataset, just adding up values... |
Beta Was this translation helpful? Give feedback.
-
|
Got alot done today, bets page is looking fantastic. Uses time series data to display profit, can sort by team as-well. I want to add a time select for 30,90,180 days etc... Also maybe add bet amount input. |
Beta Was this translation helpful? Give feedback.
-
|
In the future, work on improving the Model Training Center by adding the following features: Enhance data drop options to allow dropping players per team (from 7 to 0) and specific player stats (e.g., points or defensive rebounds). This will help focus on the most relevant data and reduce noise. Implement test game and train game seasons sliders to provide better control over the number of games in test sets and the amount of historical data used for training. This will help tailor the model to capture recent trends effectively. Additionally, include time series data graphs for model evaluation to visually assess the model's performance over time. This will make it easier to identify patterns and areas for improvement. |
Beta Was this translation helpful? Give feedback.
-
|
Created game series training chart, displayed batches of 5 games. |
Beta Was this translation helpful? Give feedback.
-
|
Added data feature options |
Beta Was this translation helpful? Give feedback.
-
|
We now gather the previous game for each team. Each previous game is input using the top 5 players by playtime's game stats along with the game score. 3x the input features and hits the api's really hard, so i removed predict all button for now. each historical game is now also formed with the previous game. |
Beta Was this translation helpful? Give feedback.
-
|
|
Beta Was this translation helpful? Give feedback.
-
|
👍🏼
…On Sat, May 6, 2023 at 11:33 AM Neal ***@***.***> wrote:
I am going to delete these messages now, please create a new topic if you
have any further questions, as this is the new feature section.
—
Reply to this email directly, view it on GitHub
<#2 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AY7NRWWCMNPZS76BGUVDN5DXE2KPHANCNFSM6AAAAAAUUAVFHE>
.
You are receiving this because you commented.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
just became a subscriber, is there a more detailed guide explaining what each feature is for training models? |
Beta Was this translation helpful? Give feedback.
-
|
Hey thanks for subscribing! Some of the settings have little popup text boxes that describe what the settings do. All of the features supported for our models directly translate to tensorflow. In general this project is kinda like a wrapper for tensorflow. When you select one of the settings in the web app, all its doing is translating that to tensorflow parameters and model architecture... I will try to write some better documentation, or might make a page specifically for info on training models. |
Beta Was this translation helpful? Give feedback.
-
|
Have an idea for a RLHF feature. So im thinking it would be great to continuously train models and update on more recent games. a button probably at the bottom of each game, when clicked will train the model on the individual game. Which model? i may have to build another model selector, or actually if its a previous game we should just make this only available for finished games, and also just make it select the model that was used to predict the game initially. Then we load up the saved dataset make sure we have the final scores, and train the model on the individual game... could be cool feature and help bridge the gap between historical data and future data... also ive been looking at a few different more professional looking data source's, i really think ensuring data integrity and authenticity is paramount. There are only a few first hand data recorders like espn and nba.com, but i want to work on a model that is based mainly off of game history. Instead of inputting the last previous game for each team, i think inputting a 5-10 game history, kinda like a sliding time window, or time series might improve accuracy. |
Beta Was this translation helpful? Give feedback.
-
|
This is awesome. I commend you. How well have the predictions been? And what are your thoughts on the various models used for prediction like whats used for https://github.com/kyleskom/NBA-Machine-Learning-Sports-Betting/ . I have found the XGBoost models to be consistent. |
Beta Was this translation helpful? Give feedback.
-
|
Awesome thanks for the insight, ill give it a read this weekend. Do you think you can also include the odds from bet365? |
Beta Was this translation helpful? Give feedback.
-
|
Ill look into it! |
Beta Was this translation helpful? Give feedback.
-
|
btw did look into bet365, the api we are using for odds data does not support bet365. I could contact the api owner and see if they can add support, not sure if its worth it tho. Most book makers tend to have very similar lines, as soon as a different casinos start posting different lines it becomes possible to arbitrage the difference and take out multiple bets with guarantee profits. From what ive seen odds tend to be identical except for the few edge cases. Let me know if you really want 365 support still, as of now im not planning on implementing this. Still possible though, I have thought about adding an arbitrage engine with much faster and wider vegas odds support but thats gona have wait for now, plus arbitrage gets you banned super fast, i think positive ev model betting has much more longevity until the inevitable back off. |
Beta Was this translation helpful? Give feedback.
-
|
Not sure for what it's worth, you are correct the odds are similar but sometimes the lines of higher sometimes especially on the over's. I only asked because bet365 if what i primarily mostly have access to in my country unless we use our local casino. But i think you also mentioned we could input our custom lines? |
Beta Was this translation helpful? Give feedback.
-
|
I agree, Custom lines are not currently support on the web ui. It is possible to manually go in find the game csv and edit to anything... on the training dashboard there is a checkbox for spread data, when checked the model is trained on the home team point spread from fanduel. The vegas line is used as input. If the spread isnt checked in the data features when you go to a game will still show all the data but at the last moment it will drop anything the model was not set to input..... I think best option would be to just use default setup and maybe a bit of line sniping waiting for the price to match what you want it to be.... I mean in a perfect world every book keeper would be supported, and probably would be easiest to set for the model when training. Then just load back up the saved setting and use the book keeper set when training. Thats probably how multiple vegas lines should work in future.... A couple notes if you input 10 spread lines that are the same or similar, you get sparse data and multicollinearity which basically means the model gets confused which variables are which... So it definitely should be a model setting. I would say the biggest road block to building that is just the fact we dont have historical vegas odds for anything other then what we are using... for live games it fetches a bunch of lines from different books, but that data only goes back like a year. where as our training data goes back 10+ years, originally i did extract transform and load the original spread dataset my self, so it would pretty much same process again, basically getting all the historical data correlate across data sets, then find a live data source and connect it all.... Like I definitely agree this would be a nice feature and i plan on expanding the model settings more, im not sure how soon this will get done. I think this does offer like better user experience if your using the specific vegas line being added, however there are many books, and i kinda think its like increasing the complexity without increasing the singular accuracy. I think in general it would be better to really get the system performing great on a single vegas line before adding a bunch more.... still in the stage where trying to improve model accuracy and figure different tricks and nifty techniques, as opposed to widening support for the system.... great ideas though i really appreciate the feedback! |
Beta Was this translation helpful? Give feedback.
-
I have more stats to track now that we have spread, so i might make separate page just with graphs.
What features would you like to see?
Beta Was this translation helpful? Give feedback.
All reactions