I’ll go ahead and chalk up the opening weekend of the NCAA tournament as a moderate success for the RotoViz March Madness app. There are some things that I was happy with and some things that I think can be improved. I think the app did give a hint on some matchups that were probably closer than people thought they would be, which was good. In general I was happy with the way that SOS ended up impacting the predictions of the app, although I do think the app overweighted the quality of the ACC and that’s likely going to be important for the app’s pick of Duke as the winner of the tournament. Also, my general theory on the tournament is that the lack of a home court advantage is the source of a lot of what we might perceive to be upsets, so that’s actually one area that I have marked down for further work. In any case, for the first time around the block, I think the app was in general a success and it could use more work. That work will have to wait for next year because the kind of changes that I think might be in order are high level tweaks to the model, like the aforementioned home court advantage.
Below I’ve updated the app with the results of the opening weekend, in case you wanted to play with it any more. Recall that at the heart of the app’s methodology is a linear model which takes the entire college basketball season and attempts to explain it best by assigning quality to each team using a generic point spread. That point spread is the number that best explains every single game during the season by minimizing squared errors. So while the opening games did add data to the model that changed some predictions for the Sweet 16, keep in mind that by this point each team has played 30 or so games. Perhaps the model should be weighted more towards the end of the season, but that’s an assumption that will have to wait for future iterations. Interestingly, I’ve checked the model vs. the Las Vegas lines and it seems to be generally in agreement with Vegas on the outcome of the Sweet 16 games.
[edit: The first version of this post went up when there was an error in the app’s code. Instead of the 13 point margin that the post originally mentioned for the Florida-FGC game, it’s actually predicting an 18 point margin. Perhaps that’s more evidence that the model needs to weight things in favor of more recent games.]
[edit2: After I posted the app I noticed that the home/away adjustment that I wanted to add was easier than I thought. There was always home/away adjusting taking place in the app when predictions were made. But the home/away adjustment didn’t actually make its way into the generic point spread, which was problematic. I added a H/A adjustment to the generic point spread and it ended up benefiting FGC. The problem for FGC is that the H/A adjustment also benefited Florida, who lost all of their games playing either away or on a neutral court. This adjustment also nudged Florida ahead of Duke based on the app’s projections.]