MBB Nerd Polls 2024-25

Calculating "average lead" means that starting hot is more important than ending hot. Hmm.
Duke wears teams down with its high quality depth during a game - a reason we normally have strong second halves. Torvik's method won't capture that. Interesting.

Edit: Nevermind. I see CDU's additional explanation. Makes more sense if this is just a secondary factor.
 
Last edited:
To be clear, neither Torvik nor Pomeroy explicitly care about final margin. They both care about adjusted efficiency margins. Also to be clear, Torvik doesn’t just use average margin. His metric just includes average margin additionally as a way to adjust for running up the score late or a team storming back once the game was already “over.” It is a “game control” measure as an auxiliary input rather than the primary measure.

As for why Torvik viewed the game as equal to the ND game, it is a combination of the “game control” measure (we had much better game control in the ND game, whereas we didn’t in the BC game) and the fact that BC is much worse than ND.

That said, I wouldn’t worry too much over it. A 93 game score is still very good (roughly about 15th nationally). To rate about 15th nationally in ehat clearly wasn’t our A game is fine.
It is true that both systems are possession based, not game based. But, for example, in KP the raw off efficiency is points per possession, which in turn is the teams final score/# of possessions in game. The adjustments work from that raw #, I believe.
Duke wears teams down with its high quality depth during a game - a reason we normally have strong second halves. Torvik's method won't capture that. Interesting.

Edit: Nevermind. I see CDU's additional explanation. Makes more sense if this is just a secondary factor.
It’s just a part of Torvik’s formula but it is a significant part. It is why our offense is rated 8th in Torvik and 4th in Kenpom. Btw I think Torvik’s adjustment is valid and the (very slightly) superior measure.
 
Last edited:
It is true that both systems are possession based, not game based. But, for example, in KP the raw off efficiency is points per possession, which in turn is the teams final score/# of possessions in game. The adjustments work from that raw #, I believe.

It’s just a part of Torvik’s formula but it is a significant part. It is why our offense is rated 8th in Torvik and 4th in Kenpom. Btw I think Torvik’s adjustment is valid and the (very slightly) superior measure.

The difference between #4 and #9 in Torvik’s offensive rating is 1.2 points (125.5 is #4, 124.3 is #9). It is a similar story for KenPom, where #2 is 125.1 and #6 is 124.2. So I wouldn’t say that Torvik’s “game control” metric is a significant driver, but rather that there are just a group of teams with VERY similar ratings, and a small difference can shift 3-4 places.

Our ORtg at KP is 124.8, while at BT it is 124.3. So there is a negligible difference.
 
Makes sense to me. I’d rather bank my lead early. Then you don’t have to pull away late!
And I would rather take the lead early and steadily increase it - steadily sucking the life out of the opponent. That eliminates late surges that make the game close.
 
It is true that both systems are possession based, not game based. But, for example, in KP the raw off efficiency is points per possession, which in turn is the teams final score/# of possessions in game. The adjustments work from that raw #, I believe.

It’s just a part of Torvik’s formula but it is a significant part. It is why our offense is rated 8th in Torvik and 4th in Kenpom. Btw I think Torvik’s adjustment is valid and the (very slightly) superior measure.
Yeah, isn't adjusted efficiency margin essentially scoring margin adjusted for competition, game location, and tempo?

KenPom appears to use the actual final score to calculate adjusted efficiency margin. Torvik uses "both the actual score and GameScript-derived score". As has been discussed, the GameScript-derived score is based on the average margin over the course of the game. Torvik's FAQ provide more detail, but he doesn't mention how he weighs the actual score and the GameScript-derived score. As far as I can tell, he appears to give ~50% weight to each.

To see how this plays out, consider the ND and BC games.

Duke beat ND by a final score of 86-78 in 65 possessions. Duke's average margin over the course of the game was 11.1, so Torvik's GameScript-derived score was 93.1-70.9.

The table below shows the Adjusted Margins based on a) the actual final score, b) the GameScript-derived score, c) an average of the actual and GameScript-derived scores, and d) the value Torvik reports on his site.
Duke/NDActual
score
GameScript-
derived
score
Average
(actual &
GameScript)
T-Rank
Site
AdjOE130.4141.1135.7135.6
AdjDE111.8101.6106.7108.5
AdjEM18.539.529.027.1
GameScore85989493

For the BC game, the actual score was 88-63. Torvik's GameScript-derived score was 81.2-69.8.

Duke/BCActual
score
GameScript-
derived
score
Average
(actual &
GameScript
T-Rank
site
AdjOE130.9120.1125.1124.5
AdjDE93.2102.697.699.9
AdjEM37.717.527.425.8
GameScore98869493

For these two games, the adjusted efficiency margins are almost exactly flipped when using the actual score (ND = 18.5, BC = 37.7) and GameScript-derived score (ND = 39.5, BC = 17.5).

The numbers in the "Average" column appear to be closest to what Torvik lists on his site*.

KenPom doesn't report game-by-game adjusted efficiencies, but I think the "actual score" column is closest to what he does. These differences would explain why Duke's KenPom rating dropped after the ND game, but rose after the BC game, while Torvik treated both games similarly.




#########################################

*I suspect that a major difference between the values in the "average" column and the actual values from Torvik's site is that Torvik stops counting when the lead is considered 100% safe. I didn't attempt to do this. For the ND game this shouldn't make much difference as the lead wasn't considered 100% safe until 14 seconds left. On the other hand, Duke achieved a safe lead in the BC game when they went up 25 with 6:02 remaining after Cooper hit his post-technical 3 pointer. Duke played BC even over the remainder of the game, but Torvik's system should be more impressed with a 25 point win in 34 minutes (55-60 possessions) than a 25 point win in 40 minutes (67 possessions).

Either way, something seems a bit "off" about my calculations. The AdjOE's are pretty spot on, but my AdjDE's are about 2 points lower than what Torvik reports for both games.
 
Last edited:
Yeah, isn't adjusted efficiency margin essentially scoring margin adjusted for competition, game location, and tempo?

KenPom appears to use the actual final score to calculate adjusted efficiency margin. Torvik uses "both the actual score and GameScript-derived score". As has been discussed, the GameScript-derived score is based on the average margin over the course of the game. Torvik's FAQ provide more detail, but he doesn't mention how he weighs the actual score and the GameScript-derived score. As far as I can tell, he appears to give ~50% weight to each.

To see how this plays out, consider the ND and BC games.

Duke beat ND by a final score of 86-78 in 65 possessions. Duke's average margin over the course of the game was 11.1, so Torvik's GameScript-derived score was 93.1-70.9.

The table below shows the Adjusted Margins based on a) the actual final score, b) the GameScript-derived score, c) an average of the actual and GameScript-derived scores, and d) the value Torvik reports on his site.
Duke/NDActual
score
GameScript-
derived
score
Average
(actual &
GameScript)
T-Rank
Site
AdjOE130.4141.1135.7135.6
AdjDE111.8101.6106.7108.5
AdjEM18.539.529.027.1
GameScore85989493

For the BC game, the actual score was 88-63. Torvik's GameScript-derived score was 81.2-69.8.

Duke/BCActual
score
GameScript-
derived
score
Average
(actual &
GameScript
T-Rank
site
AdjOE130.9120.1125.1124.5
AdjDE93.2102.697.699.9
AdjEM37.717.527.425.8
GameScore98869493

For these two games, the adjusted efficiency margins are almost exactly flipped when using the actual score (ND = 18.5, BC = 37.7) and GameScript-derived score (ND = 39.5, BC = 17.5).

The numbers in the "Average" column appear to be closest to what Torvik lists on his site*.

KenPom doesn't report game-by-game adjusted efficiencies, but I think the "actual score" column is closest to what he does. These differences would explain why Duke's KenPom rating dropped after the ND game, but rose after the BC game, while Torvik treated both games similarly.




#########################################

*I suspect that a major difference between the values in the "average" column and the actual values from Torvik's site is that Torvik stops counting when the lead is considered 100% safe. I didn't attempt to do this. For the ND game this shouldn't make much difference as the lead wasn't considered 100% safe until 14 seconds left. On the other hand, Duke achieved a safe lead in the BC game when they went up 25 with 6:02 remaining after Cooper hit his post-technical 3 pointer. Duke played BC even over the remainder of the game, but Torvik's system should be more impressed with a 25 point win in 34 minutes (55-60 possessions) than a 25 point win in 40 minutes (67 possessions).

Either way, something seems a bit "off" about my calculations. The AdjOE's are pretty spot on, but my AdjDE's are about 2 points lower than what Torvik reports for both games.
IIRC I read a few years ago that Kenpom also added a slight adjustment for runaway games, but I don’t remember how it works. However (again IIRC) KPs adjustment has less effect on lopsided game ratings than Torvik’s “average lead” approach.

I’ve been too lazy to pour through both KP & Torvik’s scattered documentation of their methodologies to attempt to extract more details. If I get my desktop finally set up I might do that (last few months I’ve been using my iphone for web surfing). But of course my preference is that you, Kedsy, CDu, etc. do it first!
 
IIRC I read a few years ago that Kenpom also added a slight adjustment for runaway games, but I don’t remember how it works. However (again IIRC) KPs adjustment has less effect on lopsided game ratings than Torvik’s “average lead” approach.

I’ve been too lazy to pour through both KP & Torvik’s scattered documentation of their methodologies to attempt to extract more details. If I get my desktop finally set up I might do that (last few months I’ve been using my iphone for web surfing). But of course my preference is that you, Kedsy, CDu, etc. do it first!
Torvik has a pretty good explanation of the differences, but it is fairly technical: https://adamcwisports.blogspot.com/p/every-possession-counts.html
 
IIRC I read a few years ago that Kenpom also added a slight adjustment for runaway games, but I don’t remember how it works. However (again IIRC) KPs adjustment has less effect on lopsided game ratings than Torvik’s “average lead” approach.

I’ve been too lazy to pour through both KP & Torvik’s scattered documentation of their methodologies to attempt to extract more details. If I get my desktop finally set up I might do that (last few months I’ve been using my iphone for web surfing). But of course my preference is that you, Kedsy, CDu, etc. do it first!
Sorry, Torvik, but "average lead" is a very mischievous stat. Lessee... a team starting a game 20-0 gets a higher game score than a team that ends the 1st half 20-0, even though halftime and final scores are same. Perhaps I misunderstand (which happens often).
 
Again, there is less than a 1pt difference between the ORtg and between the DRtg for Duke at Torvik and KenPom. Both sites view Duke almost identically.

The real differences between the two sites appear in their ratings of Auburn (4 points better offensively at BT than at KP) and Houston (2 points better offensively and 1.5 points better defensively at BT than KP). I haven't looked into why BT rates those two offenses as better, but I suspect it's a combination of three things: the "game control" measure; the way the adjusted efficiency is measured in general (KP uses an additive approach; BT multiplicative); and the way they apply their "recency bias" (i.e., how much more they weight recent games than early-season games.
 
Sorry, Torvik, but "average lead" is a very mischievous stat. Lessee... a team starting a game 20-0 gets a higher game score than a team that ends the 1st half 20-0, even though halftime and final scores are same. Perhaps I misunderstand (which happens often).
There's downsides to either approach, but I think Torvik tries to normalize a common end of game phenomenon where the score changes the margin without materially impacting the quality of the performance/closeness of the game. We often see teams up 20+ in the closing minutes put in their bench and perhaps "only" win by 10. Is it more accurate to use the 10-point margin or average it out? From a fan perspective of "how close a game felt," the averaging out actually seems more accurate. Conversely, we sometimes see a game going down to the wire with a back and forth affair with a team up by only 2 with a minute to go and the other team misses and has to foul continuously to give themselves a chance, but end up losing by 6. This was us vs. Kentucky (final margin 5 points there...) when we had a shot to actually win the game in the final 30 second - but we turned it over instead. For a game that was tight for 39 minutes, it seems fair to say the teams performed fairly evenly vs. just looking at the final margin of 6 or whatever it ends up being by garbage time FTs.
 
IIRC I read a few years ago that Kenpom also added a slight adjustment for runaway games, but I don’t remember how it works. However (again IIRC) KPs adjustment has less effect on lopsided game ratings than Torvik’s “average lead” approach.

I’ve been too lazy to pour through both KP & Torvik’s scattered documentation of their methodologies to attempt to extract more details. If I get my desktop finally set up I might do that (last few months I’ve been using my iphone for web surfing). But of course my preference is that you, Kedsy, CDu, etc. do it first!
Here's my understanding.

Torvik
Torvik makes 3 adjustments which impact blowouts:

1. He incorporates average margin (GameScript-derived score) in additional to the actual score in his individual game calculations. His system is more "impressed" when a team that jumps out to a big lead early than a close game where a big lead opens up late(r) in the game.

2. He ignore garbage time in his individual game calculations. Torvik's system completely ignores what happens after a lead is considered 100% safe based on a formula created by Bill James.

3. He "underweights" blowouts between overmatched teams when computing his overall, season-long ratings. The link that MChambers provided describes how Torvik does this. Essentially, the greater the difference in the quality of the teams, the less the game "counts" if it is a blowout. If the game is unexpectedly close (<20 points), it still counts as a full game.

KenPom
This blog post from KenPom suggests that he makes adjustment #3 above. He doesn't provide specific detail, so the exact calculation is likely different from Torvik even if the concept is the same. I can't find any mention of KenPom incorporating adjustments similar to #1 or #2, so it seems that his individual game calculations are based on the final score.

NET
My understanding is that none of these adjustments are included in the NET's efficiency margin calculation.

###################

Example
To see how this (probably) works, consider the Duke-Army game.

Torvik calculates the AdjEM of this particular game using Duke's average margin (~31 points) in addition to the final margin (42 points) and ignores everything that happened after the lead became 100% safe (Duke up 70-43 with 9:00 left). He then discounts the weight of the game by 56% when calculating Duke's season-long AdjEm rating.

KenPom uses the final score (100-58 in 71 possessions) to calculate the AdjEM of this particular game. He then discounts the weight of this game (by some unknown amount) when calculating Duke's season-long AdjEM rating.

The NET seems to consider this as a 42 point win in 71 possessions and weights it equally to all other games Duke plays.
 
There's downsides to either approach, but I think Torvik tries to normalize a common end of game phenomenon where the score changes the margin without materially impacting the quality of the performance/closeness of the game. We often see teams up 20+ in the closing minutes put in their bench and perhaps "only" win by 10. Is it more accurate to use the 10-point margin or average it out? From a fan perspective of "how close a game felt," the averaging out actually seems more accurate. Conversely, we sometimes see a game going down to the wire with a back and forth affair with a team up by only 2 with a minute to go and the other team misses and has to foul continuously to give themselves a chance, but end up losing by 6. This was us vs. Kentucky (final margin 5 points there...) when we had a shot to actually win the game in the final 30 second - but we turned it over instead. For a game that was tight for 39 minutes, it seems fair to say the teams performed fairly evenly vs. just looking at the final margin of 6 or whatever it ends up being by garbage time FTs.
On the other hand, we have games like Uconn vs butler, where they push the lead to 15, and then end up in OT anyway. Determining which possessions matter in any objective way is tricky business.

Consider the WF game. the game was "over" by all accounts...wake was up 9 with a minute left. Was the game closing up because they weren't trying? Or because they weren't good enough to hold a lead?

As you say, there's no perfect answer.
 
Again, there is less than a 1pt difference between the ORtg and between the DRtg for Duke at Torvik and KenPom. Both sites view Duke almost identically.

The real differences between the two sites appear in their ratings of Auburn (4 points better offensively at BT than at KP) and Houston (2 points better offensively and 1.5 points better defensively at BT than KP). I haven't looked into why BT rates those two offenses as better, but I suspect it's a combination of three things: the "game control" measure; the way the adjusted efficiency is measured in general (KP uses an additive approach; BT multiplicative); and the way they apply their "recency bias" (i.e., how much more they weight recent games than early-season games.
My understanding is that there can be fairly significant differences in how KenPom and Torvik view a specific game (Torvik was more impressed than KenPom with Duke's win over ND). However, these differences tend to balance out over the course of the season (Torvik was less impressed than KenPom with Duke's win over BC). So, in the end, the overall ratings tend to be pretty similar.

One major difference would be a team that (intentionally) keeps its foot on the gas well after a lead is 100% safe. KenPom will think more of this team than Torvik. This example has more to do with ignoring "garbage time" than it does average margin.
 
Last edited:
Here's my understanding.

Torvik
Torvik makes 3 adjustments which impact blowouts:

1. He incorporates average margin (GameScript-derived score) in additional to the actual score in his individual game calculations. His system is more "impressed" when a team that jumps out to a big lead early than a close game where a big lead opens up late(r) in the game.

2. He ignore garbage time in his individual game calculations. Torvik's system completely ignores what happens after a lead is considered 100% safe based on a formula created by Bill James.

3. He "underweights" blowouts between overmatched teams when computing his overall, season-long ratings. The link that MChambers provided describes how Torvik does this. Essentially, the greater the difference in the quality of the teams, the less the game "counts" if it is a blowout. If the game is unexpectedly close (<20 points), it still counts as a full game.

It's interesting. I totally get the value of #2.

And to a bit lesser degree, I can see why Torvik (and apparently KP too) choose to "down-weight" less competitive games with #3 - although it seems like if the game is only discounted when it's a blowout, the superior team has a lot more downside than upside in playing inferior teams (welcome to the new ACC?)

But if you are doing #2 and #3 already, I don't see any added value in #1. It becomes strictly a factor for rewarding teams that start strong - and penalizing teams that finish strong.

Is a team that has a lot of depth and excels at making adjustments at halftime and tends to have stronger second halves a lesser team? Torvik, apparently, would say that they are.
 
(Warning - long and ponderous explanation that may or may not be worth the read!)

Here is a thought experiment supporting Torvik's "average lead" approach. Note I am not saying his approach is perfect or even the best possible one - neither is true. It's just one argument why zooming out to an early lead and then holding it might suggest greater superiority than staying close most of the way and then pulling away at the end:

Thought experiment scenario is two runners competing in a 100 meter race, winner take all for $1000.

Situation 1: Runner A is considerably faster than runner B. He/she may not know this, but it is true.

Runners take off and A quickly finds himself pulling away. It would be common for A to then let up a bit, just putting in just the effort needed to keep a safe (say 10 yard) lead until the end.

Situation 2: Runner A is just a bit faster than B. Race starts and is close for first 20 yards, with A maybe pulling ahead a couple yards. It would be common for A to keep pushing hard since his lead isn't safe until the last few meters.

So both races ended with A winning by 10 yards, only difference is when the 10 yd lead is established. Situation 1 race pattern clearly indicates runner A was dominant, while the slowly emerging lead in situation 2 suggests A is slightly faster than B. Torvik's average lead approach would capture these differing patterns, KP's final score method not so much.

Of course, running is different than playing bball. There is very little luck in foot racing (unless you are as old as I am) while there is a lot of luck in bball games. So both talent and luck combine to create scoring streaks in a bball game and therefore even superior bball teams have less control of when their leads emerge than do runners. Sometimes a team starts out hot from 3, other games they hit their hot streak in the middle of the 2nd half. Torvik's method will give more credit to the team in the first game than the second, credit which may be a misread. But when a team gets hot in each game will vary from game to game and should even out over a season. What is left may well be a real effect analogous to the dominant runner effect and therefore reflecting a true dominance captured by the Torvik method. So, to my mind this randomness to bball only moderates the validity of the runner analogy, it doesn't obliterate it.

tl;dr vers: There are valid arguments for Torvik's average lead approach. It may be an overadjustment from the final score approach but in general, consistently taking over games early and maintaining leads is a decent indicator of marked superiority, and a bit more so than letting teams hang around and then pulling away in fourth quarters.
 
Is a team that has a lot of depth and excels at making adjustments at halftime and tends to have stronger second halves a lesser team? Torvik, apparently, would say that they are.
I would agree with Torvik.

Here's a thought experiment. Take an ACC team that beats a mediocre ACC opponent 82-50. Would you think more of that team if the halftime score was 53-26 or 29-24? Me, I'd take the team that wins the first half 53-26.

Who needs halftime adjustments when you're flattening the other team in the first half.
 
Back
Top