Sunday, January 14, 2018

Teamcast tip: note the Projected T-Rank variable

One of the newer features over at the T-Rank site is a tool I call "Teamcast," which lets you select the results of a team's games and shows how those results would affect the T-Ranketology forecast.

I'm not gonna lie, I think this is a pretty awesome tool and I play around with it a lot. But of course there are some pretty important caveats to keep in mind:

1) The underlying T-Ranketology algorithm is not a crystal ball. Although it performed really well last year, last year was a pretty predictable year. In retroactive investigations, T-Ranketology does "pretty good" for prior years too. But overall it's safe to say that you should consider any T-Ranketology forecast to have at least a +/- one-seed margin of error.

2) There's reason to believe that this year will be much less predictable than prior years, because the Committee will be considering new and different information (most importantly, the new quality win quadrants). We can only presume how the Committee will use the new information. My guess is that they will use it to make some ... unpredictable ... choices.

That said, the Teamcast is still a fun tool, and at the very least it should be pretty good for looking at relative changes. If it's predicting your team to be an 8-seed, and picking it to win a game moves them up 4 spots to a 7-seed, it's reasonable to say the win was worth about one seed-line—even if it was actually from a 10 to a 9, or a 6 to a 5.

One important thing that gets overlooked, I think, by people playing around with the tool is that the current T-Rank rating plays a role. First, it is obviously how predictions for future games are calculated. Second, it is an independent variable in the T-Ranketology algorithm, and a not insignificant one. (Not because I think the committee is using it, but as a good-enough proxy for other ratings like Kenpom that the committee is using.)

So particularly if you are exploring an extreme scenario, like say, "What would happen if my team won the rest of its games?" you might get a weird result if you do not change the projected T-Rank as well. Because in the scenario where your team wins out, its T-Rank is probably going to improve a lot.

Fortunately, the Teamcast tool lets you change the projected T-Rank (right at the top of the page), which fixes this problem. In fact, simply changing the projected T-Rank and doing nothing else is a pretty good way to use Teamcast earlier in the season. E.g., keep upping the projected T-Rank until your team gets in the field, and that shows how well your team will have to play the rest of the way to likely make the tournament.

ADDED: One way I've attempted to automate this is the "DynamaRank" feature. Basically, this estimates changes to power ratings (including NET) based on the changed results. So this box should be checked for any extreme scenarios where a team loses or wins a lot more games than expected.

If you get any funky results with the Teamcast, or have ideas for improvement, please let me know.


  1. This is a great feature, Thanks! I wish we could see odds on the actual result for instance the likelyhood of the chosen teamcast!

    1. Interesting thought. One thing that strikes me is that the likelihood of any particular result (meaning having the rest of the season go exactly as chosen) is always going to be very small, at least when there are more than a few games left.