Thursday, March 21, 2013
BPI Projections for NCAA Tournament
By Alok Pattani, ESPN Stats & Info
If you look at ESPN’s Basketball Power Index rankings and pick the higher-ranked team to win each game, you’ll end up with a chalky bracket -- other than third-seeded Florida making it to the title game.
A more nuanced way to look at things is to determine the percentage chance of each team getting to each round. This can help not only by taking into account the effects of opponent strength round-by-round (beyond just “Better BPI”), but it helps illuminate what might be considered smart upset picks and undervalued/overvalued teams.
This season, we used a unique way to predict individual tournament matchups that went beyond looking at team and opponent’s BPI numbers. Instead, we look at every team’s distribution of single game BPIs (0-100 scale), accounting for not only how good they are, but also how consistent or inconsistent their BPI game scores are. You can get an idea of this by looking at the Variation column on BPI rankings.
For example, Louisville’s 91.3 BPI rating is the highest in the country, and the Cardinals' 7.0 variation is the lowest of any team in the field. That means Louisville has been the most consistent good team this year.
Contrast that with Illinois, which has a BPI rating of 69.7 (which ranks 63rd) and a 20.5 variation. That's the highest in the field thanks to mix of great and terrible game scores.
Inconsistent teams, like Illinois, generally have a greater chance of pulling off an upset and being upset. While a season ratings only-based system like Log5 would give Illinois an 18.0 percent chance of beating Louisville, the method used here accounts for variance and would give the Illini a 22.2 percent chance because Illinois is more likely to play up to its competition. Of course, this method would also give Illinois a greater chance of being upset by a lower-ranked team for similar reasons.
These variance effects get washed out as all possible matchups are projected, but are important to keep in mind when considering certain outlier teams -- especially in earlier rounds. After taking individual matchup probabilities and accounting for every possible way the tournament could unfold, you can calculate the probability of each team reaching any given round.
If you compare these values to what the public is picking, you can find “value” picks or teams to avoid. For example:
• Pittsburgh has a 26.6 percent chance of making the Sweet 16. People are picking the Panthers to get that far only 12.7 percent of the time, but they have a decent shot of upsetting Gonzaga (assuming they beat Wichita State).
• Florida has a 10.2 percent chance of winning the national title (second-highest behind Louisville). Only 3.0 percent of brackets have the Gators winning it all, so they are undervalued.
• Miami (FL) has a 16.0 percent chance of reaching the Final Four; however they are overvalued because 36.2 percent of brackets have the Hurricanes getting there.
Using the calculated percentages to pick the most likely teams to reach each round results in a bracket that is usually extremely similar to the “Better BPI Wins” bracket (that's the case this year). But if you use BPI data from the past six seasons (unofficial before last year), this approach has outperformed the simple “Better BPI Wins” method in terms of ESPN.com’s Tournament Challenge scoring system by more than 100 points per season.
This system is based on Selection Sunday BPI and the structure of the bracket. It has no additional factors and is not fit to past years’ data as would be done in a traditional predictive model.
Given this and the fact that BPI isn’t designed to predict the tournament, this system of using all of each team’s Game BPIs to project the tournament does quite well.