BuzzerBeater Forums

You need machine learning to help you figure out what's really happening

Are you going to help me get started on that?

Instead of going overboard, maybe a relatively simple* approach could already improve this analysis.

For example using a logistic regression with win vs lose as a binary outcome variable. Then as predictors you can use a few "baseline" predictors like win% in past 5 games, past 10 games, whether it's home vs away, % salary difference between teams, opponent teams win % in past 5 and 10 games. "After" this you can add chosen tactic as a predictor and see what remaining predictive power exists for each tactic and what their influence on win likelihood is.

90% of the work is getting the dataset together correctly. Runnnig the regression is a few lines of code.

Context: What this analysis does is try to take away the other factors that also relate to winning likelihood and try to isolate what effect remains of purely the chosen tactic.



* DS/ML folks please keep the simple part in mind before commenting what all wrong about this

Yes! This is what needs to be done. It would show more accurately which tactics are dominate at the moment as some tactics may have inflated winning percentages from being used by the better team in a match up more often.

However it wouldn't tell per say what tactics are most viable in the game engine as tactics could perform poorly due to us managers not knowing how to form the right team to run them. This would probably need to be done through simulations of different team types to see what offenses just don't work under any conditions.

Just throwing something more into the conversation to either help the analysis or make it more complicated...

I might be an oddball and do things differently than everybody else, so my way of doing things should not be taken as an assumption that others do the same... but I sometimes lean towards some tactics based if I think my team is considerably better or worse than the other team. If other managers do the same, then the win % of some tactics may be a little impacted by the typical relative strengths of teams using them.

For example, if I think I've got away better team and want to run up the score and help my PD for tiebreakers in the standings, I am more likely to use R&G or PTB on offense. Similarly, if I think I'm likely to lose badly, I might put my best versatile score a little out of position to try to get an advantage and run Patient offense and hope for a hot scorer taking a ton of shots while slowing down the pace of the game and hope for a lower combined game score and lower PD. Therefore, my team is likely to have a higher win % when playing PTB and also R&G, but will have a lower win % with Patient since I'm usually the heavy underdog even before tip off.

I also suspect that BO will have a lower win % since that may be the "de facto offense" for a significant percentage of non-bot managers who are not putting a lot of time into managing their team.

There are probably a lot of other examples of data bias that may or may not be finding its way into the data analysis. Data bias is real, how to account for it is tricky.

Thank you to the development team and all the people continuing to improve the game, you are the awesome! I love this game!!!

Last edited by chihorn at 1/18/2022 2:51:43 PM

You shouldn't read too much from that data. Some tactics are used for specific purposes. Push the ball is used to get more points difference in easy games. That is why it has the best win% in div.1. Look inside and Motion are the basic tactics. And the rest of the tactics are used to avoid predictions or try to surprise the opponent or in desperate mode, which explains in part worse results.

The way to go for you is testing with hand-made teams.
All tactics are supposed to do it well when used by the right teams. If that is true and you find those teams, give us some clues so we can find those lineups also. This would be more interesting than just empower tactics so we can use all of them with the current standard lineups. If you do just this, predictions with be not longer usable and your effort will be counterproductive.

As they say in certain "whaling" communities...

Rate up is a lie!

Questions if inside isolation actually useful now, probably not... (119287415)

To say Inside Isolation got a boost, I would have thought I would have won this game more comfortably based on team ratings. I know I was away but all my ratings were better and yet I needed to hit 4 3's in the last minute to send it to overtime (pretty mental finish!). Similar fouls/free throws ands similar rebounding figures so just pure scoring I should have been better than my opponent surely?

Why is everyone posting one game they thought they should have won and claiming that it means there was most likely no significant buff to the offense. Maybe you don't have the right team composition for it? There's an element of randomness to each game so maybe this game was just on the more unlucky side?

I think it will require much more in depth analysis and really the best way is for the BBs to test different team compositions in simulations to make sure each offense is at least viable situationally.

What I'm saying is to bring more than a game or two as evidence. Outliers exist. Team ratings, DMI, salaries are only part of what matters in the game.