Predictions Audit

I've added a new link to a spreadsheet which contains all of the TAPE predictions from the current season, from December 6 through last night's games, along with the results of each game. In short, TAPE has been about as accurate as I could possibly hope for. 

• TAPE picked the correct winning team 73.1% of the time, matching its confidence level almost exactly. Over the course of 3,237 predictions, TAPE expected to be correct 2,365 times; in actuality, it has been correct 2,366 times.
• The actual results have tracked closely with the confidence level, as summarized in the second sheet of the audit workbook.
• The average error for possessions per game has been 4.15 (3.75 when overtime games are excluded.
• The average error for predicted team points per game has been 8.30 (8.04 for games decided in regulation).
• The average error for the predicted margin has been 8.70 (8.82 for non-overtime games).

I'll try and update the audit spreadsheet once a week or so, and then revisit the results, probably in a good bit more depth, after the season is over with.

Predictions Postmortem: December 1 - December 7

TAPE, obviously, is still learning. Overall, the system correctly predicted the winner of 204 of the 307 games between Division I teams last week. While a 66.4% success rate isn't horrible by any means, it's nowhere near the 74% rate that the system predicted for itself. The main reason for TAPE's underperformance was the overperformance of road teams this week. TAPE predicted that visiting teams would win 86 of the 297 contests played on non-neutral courts. They wound up winning 145.

This doesn't appear to be a case where the road teams squeaked out an inordinate number of close games, either. In addition to outperforming their win expectancy, road teams scored an average of 2.06 points per game more than TAPE had predicted (65.16 actual, 63.10 predicted), while the home teams hit their expected points almost on the nose (69.94 actual, 69.96 predicted).

Those numbers would seem to indicate that perhaps the home-court adjustment built into the prediction model is a little bit too strong, but I'm hesitant to change it just yet, for a couple reasons. First, from a theoretical perspective, TAPE is, to a large extent, self-correcting. The home-away adjustment is based on the season-to-date data and changes with each daily update. If road teams continue to do well, in other words, the bonus given to home teams and the penalty applied to road teams in the predictions will become smaller. Second, from a practical standpoint, the home-court adjustment appears to be spot-on when projecting forward. In today's season projection, home teams are expeced to win 61.29% of the inter-conference games remaining on the schedule, which is awfully close to the 61.68% of inter-conference games won by home teams last season. If this keeps up for another week, I'll revisit the situation, but right now I'm going to write it off as just a bunch of road teams having good nights for whatever fluky reason.

TAPE went 88-83 against the spread this week, but this, too, seems to be a case of road teams overperforming relative to just about everyone's expectations. In the 165 non-neutral games on which OddsShark had a point spread posted, home teams were favored by an average of 5.3 points, and TAPE predicted that the home teams would win by an average of 6.1 points, but they wound up winning by only 3.9 points.

We'll see if we can't do better the next time.