How the rankings hold up
On 3,089 games where all four systems picked a winner, Power Elo picked 2,172 correctly. FPI picked 50 more. RElo answers a different question. It ranks a season by giving equal weight to team strength and what the team accomplished.
Who makes the same Top 25?
For each edition we took the teams eligible for that comparison, put them in RElo order and counted how many of the first 25 also made the other system’s Top 25. A shared team counts once regardless of position. Each edition gets one vote in the average.
| Results through | FBS teams | AP | CFP | FPI | SP+ | Massey |
|---|---|---|---|---|---|---|
| Nov 25, 2023 | 133 | 22 | 24 | 20 | 22 | 22 |
| Nov 29, 2025 | 136 | 22 | 23 | 19 | 21 | 19 |
| Dec 6, 2025 | 136 | 22 | 23 | 19 | 21 | 19 |
All three computers ranked the full, identical FBS field on each of these dates. AP and CFP publish only the leading teams. We matched the last completed game date, checked the archived FPI order against ESPN’s dated release, used Connelly’s own modern SP+ tables, and retrieved the individual Massey ballot. No subsequent Division I game had been played before those ballots were collected. The committee shared the most teams with RElo on these three weekends. Three late-season observations across two seasons are too few to generalize that ordering to the whole year.
Release dates for these three comparisons
- November 25, 2023 games. FPI and AP appeared November 26, CFP November 28, and the individual Massey ballot was archived November 30. November 2023 ballots
- November 29, 2025 games. FPI and AP appeared November 30, CFP December 2, and Massey was archived December 4. Week 14 ballots
- December 6, 2025 games. FPI, AP and CFP appeared December 7, and Massey was archived December 11. Week 15 ballots
There is also a longer comparison using Connelly’s earlier S&P+ model. Across 26 shared weekend slates from 2014–18, the average number of teams in common with RElo was 22.46 for AP, 22.73 for CFP, 20.31 for FPI, 18.85 for historical S&P+ and 21.88 for Massey. We used Massey’s own rating. Some archive weeks label the collection date without dating each contributor’s ballot. In particular, two 2016 FPI entries appear to carry over earlier rankings. Excluding both weekends gives 22.50, 22.71, 20.29, 18.79 and 21.75 respectively across 24 editions.
Across 362 AP editions since 2003, RElo averaged 21.05 shared teams and agreed on No. 1 in 54.7% of editions. Across 74 CFP editions since 2014, it averaged 22.65 and agreed on No. 1 in 63.5%. The committee starts later in the season, when more results are available. The two averages cover different dates.
| Ranking and field | Seasons spanned | Editions | Shared Top 25 | Median rank correlation |
|---|---|---|---|---|
| FPI · FBS, dated releases | 2015–25 | 84 | 18.50 | 0.910 |
| SP+ · FBS | 2022–26 | 57 | 18.40 | 0.887 |
| Historic S&P+ · FBS | 2011–18 | 81 | 18.38 | 0.906 |
| Massey · FBS | 2003–21 | 297 | 20.52 | 0.964 |
| SP+ · FCS | 2022–26 | 70 | 18.69 | 0.864 |
| Massey · FCS | 2003–21 | 304 | 17.14 | 0.803 |
Each row above covers its own available weeks and shared teams. Rank correlation measures order across that field. A value of 1 means identical order. Five AP editions included 26 names because of ties at 25. Preseason polls are excluded.
Which games explain the difference?
We removed one result, kept the season’s opening ratings fixed, and recalculated the entire league through the same date. The table shows the actual ranks on three selection weekends. The examples below give the difference between a team’s RElo with and without a game. The public team logs apply the same question to Power Elo. Games share opponents and their effects can overlap.
| Team | RElo | AP | CFP | FPI | Historic S&P+ | Massey |
|---|---|---|---|---|---|---|
| 2016 Alabama | 1 | 1 | 1 | 1 | 1 | 1 |
| 2017 UCF | 7 | 10 | 12 | 21 | 9 | 10 |
| 2018 Clemson | 1 | 2 | 2 | 2 | 2 | 2 |
| 2018 Alabama | 2 | 1 | 1 | 1 | 1 | 1 |
2023 Washington, with modern SP+. Before the Pac-12 title game, Washington was 12–0. Through November 25, AP and CFP ranked it third, RElo fifth, Massey sixth, SP+ 11th and FPI 13th. Its R Score ranked first and its Power Elo ninth. The 36–33 home win over Oregon, fourth in Power Elo at 2094.2, added 52.7 RElo points. Removing that win dropped Washington from fifth to ninth. Only the first Oregon meeting had been played at this cutoff. Washington’s November ballot archive
2018 Clemson and Alabama. Both were 13–0 through December 1. All five benchmarks preferred Alabama. So did Power Elo, 2271.4 to 2239.2. Clemson had the stronger R Score, 9.74 to 9.07, and led the blended ranking 2250.7 to 2240.2. Removing Clemson’s 41–7 win over NC State cost 45.0 RElo points and flipped the order. Alabama’s own 29–0 win over LSU was worth 46.1 points.
2017 UCF. Through December 2, two wins against Memphis helped explain the undefeated Knights’ No. 7 ranking. The 40–13 win was worth 45.6 RElo points and the 62–55 title game was worth 36.7. Removing either one dropped UCF to ninth. Removing its 49–24 win over UConn cost just 5.0 points and left UCF seventh. Memphis had a Power Elo of 1856.2 at the cutoff.
2016 Alabama. All six rankings agreed on No. 1 through December 3. The 52–6 win over USC was worth 48.7 RElo points. Alabama remained first after we removed each of its 13 games in turn. These experiments explain the Index’s order. The rankings alone cannot tell us how an individual voter evaluated a game.
How often did Power Elo pick the winner?
We used each system’s pregame pick on the same completed FBS games. Power Elo favors the higher-rated team after a 20-point Elo allowance for the home team at a non-neutral venue. For FPI and Massey we used their archived game projections, and for SP+ we used Bill Connelly’s weekly selections. Missing, ambiguous or tied picks were excluded from the common sample.
| Season | Games | Power Elo | FPI | SP+ | Massey |
|---|---|---|---|---|---|
| 2022 | 757 | 68.03% | 70.15% | 69.35% | 69.75% |
| 2023 | 783 | 70.11% | 72.54% | 74.07% | 72.29% |
| 2024 | 793 | 70.49% | 70.74% | 69.99% | 68.35% |
| 2025 | 756 | 72.62% | 74.34% | 73.02% | 71.96% |
| 2022–25 combined | 3,089 | 70.31% | 71.93% | 71.61% | 70.57% |
| 2026 through Oct 31 | 492 | 70.12% | 73.78% | 73.58% | 74.19% |
The four-season counts were 2,172 correct for Power Elo, 2,222 for FPI, 2,212 for SP+ and 2,180 for Massey. Some systems have longer archives. Against FPI alone, Power Elo trailed by 1.87 percentage points on 8,219 shared games (95% interval, −2.67 to −1.09). Against SP+ the gap was −1.21 on 3,136 games (−2.50 to +0.16). Against Massey it was −0.45 on 8,886 games (−1.15 to +0.25). We estimated these intervals by resampling whole weeks 10,000 times, keeping predictions for the same game paired. The wider FPI comparison gives the clearest evidence of a forecasting gap.
What we would change
Keep the published 50/50 RElo. Raising the power share to 75% improved AP overlap on 2014–25 editions from 20.96 to 21.30 teams. It reduced CFP overlap from 22.63 to 22.29. Choosing the blend to imitate one poll would change what the Index measures.
We used 2003–06 to establish opening ratings, then tested nine Power Elo settings on 9,745 games in 2007–13. The update sizes were 64, 128 and 256, and the margin weights were 0.25, 0.5 and 1. Each setting carried its own ratings forward. We selected the lowest average log loss, froze that choice, and compared it with the published settings on 17,701 FBS/FCS games in 2014–25. For this experiment every pick in a Monday–Sunday week was fixed before the week began.
| Power Elo settings | Accuracy | Log loss ↓ | Brier score ↓ |
|---|---|---|---|
| Published · 128, 0.5 | 72.30% | 0.53815 | 0.18162 |
| Training choice · 64, 1.0 | 72.40% | 0.53362 | 0.17964 |
The candidate picked 17 additional winners across 12 seasons and improved both measures of probability error. On the specific 3,089 recent FBS games with all four forecasts, it picked 70.22%, just below the published 70.31%. That warrants a prospective comparison before changing Power Elo. Some later seasons had already been examined during development. They were held out of this parameter search, and a truly unseen test begins with future games.
We also tested reducing an opponent’s influence when one unusual game accounted for much of its rating. The old-season search favored zero discount. Applying the best nonzero discount (25%) on the 8,219 FPI matchups cost 15 correct picks. It also lost three correct picks on the 2,131 games where the less experienced team had played three to five games. We kept the original opponent weighting.
The Index’s strongest demonstrated use is explaining a ranking. Readers can see whether strength or résumé puts a team ahead, and which individual results matter. The holdout evidence does not support calling Power Elo the most accurate forecaster.
Coverage, timing and sources
We replayed 33,839 completed Division I games from 2003 through November 14, 2026. All games played on the same Denver calendar date enter together. Pregame ratings use the last completed date before kickoff. Within each season, later evidence can change how the model values an earlier opponent. In the replay, a future result can never change the rating used to make an earlier pick.
For each AP or CFP release we use the most recent ratings strictly before its calendar date. Historical computer archive dates label ranking weeks, sometimes after the games and polls. For SP+ workbooks the cutoff is inferred from the first upcoming matchup. ESPN’s independently dated FPI extract provides an additional check. The final-whistle times, exact posting times for every table and historical revisions to scores are unavailable. Some same-day information can differ even when the field and weekend match. Repeated weekly editions share teams and games.
FCS coverage. FCS games affect opponents’ ratings, game impacts and the tuning test. AP and CFP provide no full FCS field, and our FPI archive covers FBS. A five-system FCS result cannot be estimated from these records. FCS-specific SP+ and Massey tables support the separate rank comparisons above. Limited games between subdivisions also make their relative placement harder to validate.
Score archives and FPI timestamps come from SportsDataverse. The historical ranking columns identify Kenneth Massey’s own system and Connelly’s S&P+ predecessor. Poll release dates and SP+ projections come from the linked archives. Published results here are aggregate comparisons calculated by the Index.
- Scores and schedules · Dated FPI releases
- AP poll archive · CFP weekly rankings
- Kenneth Massey’s historical rankings archive · Paul Kislanko’s archive of individual ballots · Prediction Tracker’s game forecast archive
- Bill Connelly’s SP+ workbooks · 2022 · 2023 · 2024 · 2025