Pages

RPI Plus: A simple fix to the RPI

3/20/2008
Buried deep in Part 2 of my 3-part presentation exposing the flaws of the RPI is a proposal for a simple, common sense fix to some of the major problems with the RPI. I called it RPI+ because it takes the RPI and adds an extra performance component to come up with a more accurate number.

Essentially, all it's doing is asking 'If Team X has a RPI of 100, did they outperform or underperform expectations for team with that RPI and that schedule. A team that won more games than expected for its RPI would have a pretty strong case that its rating was too low. A team that lost more games than expected for its RPI would be viewed as being rated too high. This is basically what I do with the Colton Index, only I start with the win-loss record and do multiple iterations of the expectations and performance relative to expectations until the complete ranking set is as correct as it can be. The RPI+ proposal uses just one iteration of the performance component (although additional iterations would only make it better). You'll see in a minute that just one iteration does have a significant impact. Also, because you can capture the home-road adjustments more accurately in the performance component, you really wouldn't need to do the crude 40% home/road adjustment upfront. You'd be better off just using the old RPI and let the home/road show up in the performance component. However, I used the new RPI here simply because it makes for easier comparison of teams that would be most impacted.

The other proposed change is to finally abandon the deeply flawed, linear RPI SOS measure. Since we are calculating expected winning percentage for the performance component, it's easy to implement here. The proposed SOS captures the home/road influence accurately and recognizes that SOS is not linear (playing the 1st, 2nd and 330th ranked teams on the road IS NOT the same as playing the 110th, 111th and 112th rated teams at home).

Below is the new RPI and RPI+ for the top 90 RPI teams plus 6 teams who would be top 90 under RPI+. The 'Exp Wins' column shows the Expected Number of Wins for a team with that RPI given its schedule. So Tennessee would be expected to win 28.11 with its schedule if it were the #1 RPI team. Since they won 28 games, you'd expect their 'true' ranking to be slightly lower than that. The far right column shows the expect number of wins in the second iteration (if you account for the change in a team's rating for RPI+ and the change in the rating of all its opponents). I only show it here for information purposes, so you can get some sense directionally which way teams would go if you added additional performance iterations. The top 20 overperformers relative to RPI are shaded in yellow. The top 20 underperformers relative to RPI are shaded in orange.

Just adding the one performance components is a huge improvement in the RPI. The wins versus expectations are off by an average of 0.89 wins per team in the new RPI. Adding that one performance component brings that number down to 0.34 (the Colton Index is off by 0.04 per team). The other way I like to look at these rankings is by measuring the amount of informational lift that ranking provides over the most simple rank order you could think of...simply sorting teams by their win-loss record. We use alternative rankings because we know win-loss record doesn't tell the full story. Presumably, the ranking is a more accurate reflection of what's going on during the course of a basketball season, otherwise we wouldn't use it. Therefore, it's easy to measure different rating systems on how much additional value they provide.

Traditionally, the new RPI has been about 5-6% worse than the old RPI, and the Colton Index has provided 40-50% more lift versus the baseline relative to the new RPI. The 2007-08 season is no exception. I know it's hard to believe, but the old RPI is 5.2% better than the new RPI. The Colton Index outperforms the new RPI by 44%. RPI+ is 33.5% better than the new RPI. This seems like an easy fix that would provide the Selection Committee with significantly better information to make their decisions.

A couple teams to note from the table below:

- Vanderbilt seemed to be overseeded as a 4 seed, and we see their performance was over a game worse than expected from that lofty 12th-rated RPI. The 2nd iteration would suggest they would go down even futher with additional iterations. UNLV is another team that seemed to be overseeded as well.

- The Selection Committee doesn't seem to have much faith in very strong RPI numbers for mid-majors (see Southern Illinois last year). Butler seemed to suffer from this bias this year, although the RPI+ numbers suggest they outperformed their 17th-rated RPI.

- Arizona St is one of the biggest movers amongst tourney or bubble teams. They move up 30 spots when you consider they outperformed their RPI rating by 1.73 games. Also, by moving away from the linear nature of SOS, the 'bottom feeders' in their non-conference schedule don't hurt as much and their SOS increases from 77 to 32.

- South Alabama doesn't move in RPI+ (they would move down in future iterations), but just check out that SOS impact. 127 vs 203.

EXP WINS RANK SOS EXP WINS2
TEAM WL RPIRES RPIRPI+ RPIRPI+ RPI+RES2
Tennessee284 28.11-0.111318 28.30 -0.30
North Carolina 322 30.511.49 21 321 31.46 0.54
Memphis 331 31.291.71 32 27115 32.28 0.72
UCLA 303 28.321.68 44 1422 29.42 0.58
Kansas 303 28.901.10 55 5072 29.68 0.32
Texas276 27.25-0.256846 26.97 0.03
Duke275 26.490.5176810 26.89 0.11
Georgetown275 26.440.56873243 26.97 0.03
Xavier276 27.34-0.349112455 27.50 -0.50
Drake264 25.350.6510106892 26.03 -0.03
Wisconsin 294 27.331.67 119 6186 28.50 0.50
Vanderbilt 267 27.12-1.12 1217 4183 26.55 -0.55
Louisville248 24.50-0.501314623 24.02 -0.02
Stanford 267 25.060.94 1413 5629 25.72 0.28
Pittsburgh259 25.03-0.0315152012 24.92 0.08
Michigan St258 25.58-0.5816224358 24.80 0.20
Butler 293 27.401.60 1712 128156 28.88 0.12
Connecticut248 24.10-0.1018213540 23.74 0.26
Clemson239 22.910.0919201719 22.91 0.09
Marquette239 22.500.5020192115 22.81 0.19
Kent St286 27.730.272124111140 28.24 -0.24
Indiana257 24.490.5122235980 25.10 -0.10
Washington St 248 22.721.28 2316 4625 23.61 0.39
UNLV 257 26.17-1.17 2434 63141 25.48 -0.48
BYU27726.78 0.22252597122 27.30 -0.30
Notre Dame 247 22.761.24 2618 8167 23.89 0.11
Oklahoma2211 22.16-0.1627311324 21.88 0.12
Southern California21 1120.680.32282793 20.49 0.51
West Virginia2310 22.640.3629284533 22.85 0.15
Gonzaga 257 23.921.08 3026 93102 24.99 0.01
EXP WINS RANK SOS EXP WINS2
TEAM WL RPIRES RPIRPI+ RPIRPI+ RPI+RES2
Arkansas 2211 23.29-1.29 3140 2559 21.98 0.02
Dayton 2110 22.39-1.39 3242 3378 21.16 -0.16
Illinois St 239 23.85-0.85 3339 7193 22.84 0.16
Miami (FL)2110 21.22-0.2234364248 21.01 -0.01
Davidson256 24.430.573530129150 24.96 0.04
St. Mary's246 23.280.723629138145 24.16 -0.16
Arizona 1814 19.27-1.27 3746 21 17.19 0.81
South Alabama246 24.19-0.193837127203 24.60 -0.60
Western Kentucky256 24.620.383935142201 25.53 -0.53
Mississippi St2210 22.33-0.3340415885 22.26 -0.26
Texas A&M 2310 21.991.01 4133 5256 23.06 -0.06
Massachusetts2110 21.31-0.3142437084 21.17 -0.17
Baylor2010 19.710.2943384954 19.98 0.02
Saint Joseph's2112 21.55-0.5544485360 21.23 -0.23
Purdue 248 22.581.42 4532 11498 23.89 0.11
Creighton2010 20.40-0.4046506679 19.96 0.04
Temple 2112 22.15-1.15 4758 5173 20.65 0.35
Mississippi2110 21.31-0.3148496591 21.07 -0.07
Ohio State1913 19.61-0.6149541827 18.70 0.30
Kansas St1911 19.33-0.3350513451 18.61 0.39
Villanova2012 19.780.2251454739 20.04 -0.04
Virginia Tech 1913 19.98-0.98 5263 3938 18.45 0.55
Oral Roberts238 23.72-0.725359157191 23.56 -0.56
VCU24723.61 0.395444161189 24.45 -0.45
Syracuse1913 19.68-0.6855611034 18.50 0.50
UAB 2210 23.56-1.56 5670 95155 22.12 -0.12
Kentucky1812 18.03-0.0357551941 18.02 -0.02
Oregon1813 18.02-0.025856379 17.36 0.64
New Mexico 248 23.140.86 5947 155165 24.51 -0.51
Florida St1914 19.11-0.1160601516 18.49 0.51
EXP WINS RANK SOS EXP WINS2
TEAM WL RPIRES RPIRPI+ RPIRPI+ RPI+RES2
George Mason 2310 24.79-1.79 6183 126211 23.20 -0.20
Southern Illinois 1714 18.40-1.40 6277 1642 16.09 0.91
Cleveland St 1912 20.40-1.40 6386 85105 18.65 0.35
Stephen F. Austin225 21.210.796452263297 22.42 -0.42
Cornell215 20.620.386562265292 21.35 -0.35
Siena 2210 23.09-1.09 6685 120179 22.21 -0.21
Georgia Tech1517 15.70-0.70677672 14.43 0.57
Charlotte2013 20.17-0.1768716495 20.12 -0.12
IUPUI237 23.19-0.196973238277 23.44 -0.44
Utah St 2310 24.93-1.93 70102 168267 23.44 -0.44
Florida 2111 19.921.08 7157 8794 21.22 -0.22
Miami (OH) 1715 17.93-0.93 7289 5450 16.83 0.17
Nevada 2011 21.53-1.53 73100 132161 20.18 -0.18
Texas Tech1515 15.65-0.65748457 14.46 0.54
Akron2310 22.530.477565135157 23.37 -0.37
Ohio 1912 20.07-1.07 7694 109123 18.94 0.06
Rhode Island2111 20.470.537764107124 21.27 -0.27
Oklahoma St1615 15.940.0678721211 15.85 0.15
Belmont248 24.46-0.467982229293 24.81 -0.81
CS Fullerton238 23.29-0.298079199249 23.24 -0.24
Houston229 21.620.388169146171 22.31 -0.31
San Diego St 1912 19.80-0.80 8290 99118 19.00 0.00
Arizona St 1912 17.271.73 8353 7732 18.58 0.42
Austin Peay 2410 26.00-2.00 84116 205288 25.12 -1.12
Maryland1814 18.05-0.0585803163 17.89 0.11
Wright St2110 20.210.798666149144 21.35 -0.35
Boise St248 23.590.418774196276 24.61 -0.61
UMBC238 23.57-0.578895264302 23.57 -0.57
UC Santa Barbara228 21.620.388975208221 22.33 -0.33
American 2111 22.58-1.58 90120 158237 21.48 -0.48
California1615 15.510.4992812814 15.61 0.39
Nebraska 1912 17.391.61 9668 9469 18.72 0.28
Wake Forest 1713 15.261.74 10067 8652 17.20 -0.20
Minnesota 2013 18.751.25 10178 11387 20.03 -0.03
Providence 1516 13.621.38 11287 4417 15.09 -0.09
Washington 1616 14.061.94 12188 6018 15.78 0.22

0 comments:

 
Wegoblogger #31 © 2011 | Designed by Bingo Cash, in collaboration with Modern Warfare 3, VPS Hosting and Compare Web Hosting