A graph of substitution patterns

I have 828 play-by-plays from last season where the final margin was single-digits and for which the play-by-play has clean substitution data. “Clean” meaning that according to the play-by-play there were five players on the floor at all times. (This does not necessarily mean it’s accurate. There are probably some glitches as the curiously low figures immediately after halftime indicate.) Below is a graph of the average starters per team that were on the floor in those games, and the percentage of time that the entire starting five was on the floor. There’s been a lot of research on lineup construction in baseball, but to my knowledge there isn’t much going on in terms of how to best deploy one’s players in basketball. It’s a given that your best players will start the game, and it makes sense to distribute a player’s time on the floor across the entire 40 minutes of the game. However, it seems like the enterprising underdog could use a predictable substitution...