ive been building nba prop models for years now and the biggest mistake i see people make isnt bad data. its treating all data the same. so heres four things you can actually go implement right now that moved my numbers significantly
first. stop using season averages. theyre polluted. what you want is a weighted recency window that decays older games. think of it like this. a game from 6 weeks ago should not count the same as last night. you can use an exponential decay on the last 15 to 25 games and your signal gets way cleaner immediately. its not hard to code. just weight each game by how recent it is and let the older stuff fade
second. opponent defense needs to be position specific not team wide. team defensive rating is an aggregate stat and its useless for props. a team can be bottom 10 against guards but top 5 against bigs and if youre taking a point guard under against them because their team drtg looks good you are using the wrong number. isolate by position. it takes more setup but its probably the single biggest edge i added
third. track minutes trends separately from usage trends. most people merge these but they tell you different things. minutes tell you opportunity. usage tells you what the player does with that opportunity. a guy can see a minutes bump from a rotation change but if his usage rate stays flat he is just standing in the corner longer. he is not getting more shots. these are two different signals and you need both
fourth. add a variance filter. if a player had two crazy outlier games in the last 10 your model probably thinks hes trending up. but if you strip those two outliers his baseline hasnt moved at all. run a quick check on whether the recent trend is being carried by outliers or if its structural. if its outliers only. skip the pick. you dont need it
none of this is rocket science. its just that most people stop building their model once it outputs a number and never add the filters that actually separate signal from noise. these four things alone will clean up your board a lot. the rest of what i run is deeper but this is the stuff anyone can build in a weekend