"In Dallas, Doncic is reaching 2017-Russell-Westbrook levels of usage and his scoring efficiency is paying the price."
The only reason his TS is down from regular season is he is shooting 43% from FT line (0-5 game 4). His efg is actually up, .550 ->.563. Unless your saying his free throw is related to his high usage.
Nice stuff, but I have my doubts about the accuracy of Second Spectrum's data as well. I've seen plenty of writers (like LGW) use play-type data in analyzing what a player is good at without ever seeing anyone analyze how good Second Spectrum's data quality really is for that sort of analysis—especially when it comes to classifying a particular player's contribution as a "screen assist" or "P&R roll man" we're often dealing with low sample sizes that are susceptible to noise. Cameras and machine learning can do some impressive things, but they screw up too; for all I know, they might screw up more than the human scorers do. Heck, 583's metrics use tracking data to analyze players' defense—specifically, (if I remember correctly) it rewards or punishes the "nearest defender" based on shooting percentages vs. league average. Which, I have some issues with on its own¹, but I don't think anyone has publicly shared stats on how accurate Second Spectrum even is at fundamental stuff like correctly labeling the nearest defender, let alone slippery definitions like shot type and play type.
¹ Specifically, 538's approach to rating defense this way lets guys off the hook for getting blown by or losing track of their assignment completely. They at least try not to punish the teammate who rotates to attempt a difficult contest, by giving guys credit for defending more shots. (They acknowledged this shortcoming in the introduction to the "DRAYMOND" metric, but as far as I know there haven't been any follow-ups to analyze how much this skews the numbers.) Also, even Second Spectrum doesn't seem to give out any stats (or at least, 538 isn't using them) when a guy defends so well that the offensive player does not even attempt a shot; as far as I can tell, their metric probably actually punishes this a little because it rewards people for defending more shots. But I digress.
Anyway, it's great that there are people like you taking a skeptical eye to the statistics and identifying systematic bias in the way they're calculated. It's easy to talk up data-driven decisions and opinions, but those can be a lot more accurate when you know the caveats and limitations of the data you're using.
"In Dallas, Doncic is reaching 2017-Russell-Westbrook levels of usage and his scoring efficiency is paying the price."
The only reason his TS is down from regular season is he is shooting 43% from FT line (0-5 game 4). His efg is actually up, .550 ->.563. Unless your saying his free throw is related to his high usage.
Nice stuff, but I have my doubts about the accuracy of Second Spectrum's data as well. I've seen plenty of writers (like LGW) use play-type data in analyzing what a player is good at without ever seeing anyone analyze how good Second Spectrum's data quality really is for that sort of analysis—especially when it comes to classifying a particular player's contribution as a "screen assist" or "P&R roll man" we're often dealing with low sample sizes that are susceptible to noise. Cameras and machine learning can do some impressive things, but they screw up too; for all I know, they might screw up more than the human scorers do. Heck, 583's metrics use tracking data to analyze players' defense—specifically, (if I remember correctly) it rewards or punishes the "nearest defender" based on shooting percentages vs. league average. Which, I have some issues with on its own¹, but I don't think anyone has publicly shared stats on how accurate Second Spectrum even is at fundamental stuff like correctly labeling the nearest defender, let alone slippery definitions like shot type and play type.
¹ Specifically, 538's approach to rating defense this way lets guys off the hook for getting blown by or losing track of their assignment completely. They at least try not to punish the teammate who rotates to attempt a difficult contest, by giving guys credit for defending more shots. (They acknowledged this shortcoming in the introduction to the "DRAYMOND" metric, but as far as I know there haven't been any follow-ups to analyze how much this skews the numbers.) Also, even Second Spectrum doesn't seem to give out any stats (or at least, 538 isn't using them) when a guy defends so well that the offensive player does not even attempt a shot; as far as I can tell, their metric probably actually punishes this a little because it rewards people for defending more shots. But I digress.
Anyway, it's great that there are people like you taking a skeptical eye to the statistics and identifying systematic bias in the way they're calculated. It's easy to talk up data-driven decisions and opinions, but those can be a lot more accurate when you know the caveats and limitations of the data you're using.
This is great stuff!!! I just cross-posted at LGW.
And as someone who understands how tedious this exercise must have been, I absolutely applaud your efforts here. Pretty cool to see the side by side.