Love following everything you post here and on the site. Got a question for you on the Opponent Strength Adjustment component, if you get a second.
I'd imagine that for major conference teams (and probably low majors in the reverse direction - but who cares about them), opponent strength and date of matchup might be somewhat correlated - with most playing the weakest parts of their schedule during the non-con season. If teams do not really change over the course of the season, that's a moot point. Or if there's enough strength variation within-conference, that's also probably moot. But I think there's an argument that both of those assumptions would be incorrect.
And two related concerns: 1) The lowest Big 12 team is 141 in your rankings, how reliable would the Strength Adjustment be outside of that range? 2) All Big East teams are above 73 in your rankings (except Georgetown at 182 and DePaul at 307); are games against those two teams have disproportionately high influence on the rating?
I am not questioning your model, it just seems hard to estimate! (Maybe not if I had a statistics PhD, haha.) Is there is a simple-ish enough explanation for why these are not concerns?
To your first question, I account for the recency of each game in the model so that games played more recently get more weight than games played way earlier in the season. I chose this weight based on what led to the best future game predictions in my previous season backtesting.
Something else that I do in future game predictions is that, say, a 16 seed is playing UConn in the first round of the tournament, but the best team they have faced all season is the team ranked 80th, I don't allow for the strength adjustment to go past the best team that they have faced all season, so the opponent strength adjustment for that game would cap UConn's strength at 80th.
To your DePaul / GTown question, to a certain degree, yes those games would carry more weight, but that's something that I'm going to fine-tune this offseason to see if that actually leads to better predictions or not.
Love following everything you post here and on the site. Got a question for you on the Opponent Strength Adjustment component, if you get a second.
I'd imagine that for major conference teams (and probably low majors in the reverse direction - but who cares about them), opponent strength and date of matchup might be somewhat correlated - with most playing the weakest parts of their schedule during the non-con season. If teams do not really change over the course of the season, that's a moot point. Or if there's enough strength variation within-conference, that's also probably moot. But I think there's an argument that both of those assumptions would be incorrect.
And two related concerns: 1) The lowest Big 12 team is 141 in your rankings, how reliable would the Strength Adjustment be outside of that range? 2) All Big East teams are above 73 in your rankings (except Georgetown at 182 and DePaul at 307); are games against those two teams have disproportionately high influence on the rating?
I am not questioning your model, it just seems hard to estimate! (Maybe not if I had a statistics PhD, haha.) Is there is a simple-ish enough explanation for why these are not concerns?
To your first question, I account for the recency of each game in the model so that games played more recently get more weight than games played way earlier in the season. I chose this weight based on what led to the best future game predictions in my previous season backtesting.
Something else that I do in future game predictions is that, say, a 16 seed is playing UConn in the first round of the tournament, but the best team they have faced all season is the team ranked 80th, I don't allow for the strength adjustment to go past the best team that they have faced all season, so the opponent strength adjustment for that game would cap UConn's strength at 80th.
To your DePaul / GTown question, to a certain degree, yes those games would carry more weight, but that's something that I'm going to fine-tune this offseason to see if that actually leads to better predictions or not.
Ry Daw's above comment is quite probative. Would very much appreciate Evan's response when time permits.
Replied above.