Kickoff: Conference Realignment, the Portal, and the Coaches Who Got Fired: Joseph Richardson

Georgia Tech lost to Colorado at home on Thursday night, the opening night of the 2026 season, 14 to 13. Tech led 13 to 7 late in the fourth quarter. Colorado scored with 37 seconds left, and Tech's 45 yard field goal to win was blocked as time expired. Vegas had Tech favored by six and a half points, which is about a 68% chance to win. My own prediction model had Tech at 75%.

One play decided that game. The score says almost nothing about which team was better or whether 68% or 75% was the better number. Every prediction system in this essay, and there are four, is scored on its probabilities across thousands of games and never on one result.

Where did the numbers come from? In week one, no model had seen either team play a down this season. My model's 75% prediction was built on last season. Only 12.6% of the yards and points Georgia Tech produced last season came from players still on the roster, one of the twenty thinnest returning rosters in FBS. None of the passing yards came from current quarterbacks, because Haynes King is gone. Colorado had 13.7% of its 2025 production return. It rebuilt its roster through the transfer portal in 2023 and moved from the Pac-12 to the Big 12 in 2024. Both numbers, the six and a half points and the 75%, were built on players who are no longer there.

Many things change a team between seasons. Here are three worth considering: The league it plays in, the roster it plays with, and the head coach. Twenty-three programs changed leagues in 2013 and fifteen in 2024. More than two thousand players entered the portal before the 2023 season. One program in five changes its head coach in an average year. This essay measures what that churn does to college football prediction systems. The four are Vegas, the AP poll, the ESPN Pick'em Crowd, and my model, which we will call the Machine. The Crowd only picks the ten College Pick'em games each week, so that is where it is scored. Essay one set the Machine against the other three.

I have no control over the first three prediction systems. I'm just an observer. But the Machine can be taken apart.

What prediction systems know in September

The Machine is made up of six machine learning models: Four gradient boosted trees, an elastic net and a small neural net. All six read the same thirty features, and their six answers are weighted into one probability for each game. How the Machine is built, and how it's used to predict College Pick'em winners this season, are on another page.

The thirty features come in six groups, and the biggest group is team strength, measured three ways: An Elo rating run over every game since 2001, a second rating that measures how efficiently a team moves the ball and stops the other side, adjusted for who it played, and a preseason prior. Each is there for the home team, the away team and the difference between them. The prior is the number that exists before a snap is played. It blends last season's rating, how much of last season's production is back, recruiting, roster talent, and net transfer portal flow, the stars of the players who came in minus the stars of the players who left. Then come the smaller things: Days of rest, travel, elevation, weather, a stadium roof, whether the game is inside a conference, and each team's AP rank.

Figure 1. UNLV at Hawai'i, as the Machine reads it. Thirty inputs in six groups. A difference is home minus away. The dashed boxes are blank in week 1.

In the season's opening week, one of those groups in the figure above is blank. The form columns, which track how a team has played over its last few games, have nothing to track. So a week-one prediction is last season's team against this week's opponent. The first few games of each season prove almost nothing about many teams: Until about week four there are too few results to chain one team's games to another's, so most pairs of teams cannot be compared through anyone they have both played.

That is why the three churns do the most damage to the prediction systems in September. The new league, the new roster and the new coach are there all season, but in September the prediction systems have nothing to go on except last season, and a few out-of-conference games, easy or hard, do not tell them much about how the team will play in league. It takes a few weeks of tougher opponents before the prior catches up to the team on the field. Until then, a new league changes who a team plays, a new roster means last season's stats belong to players who are gone, and a new coach often means a new roster too, partly because his system needs different players and partly because the players were recruited by, and loyal to, the man who just left.

For Vegas, the Crowd and the AP poll, the only question I can ask is whether they did worse on the churned teams and seasons than on the rest. That is an observation, not an experiment: I cannot choose which leagues realign and which stay put and then see what happens, so every result below states what its test could and could not have seen. The Machine allows a second kind of question. I can remove one of its inputs, rerun twenty-one seasons without it, and measure what the input was worth in each era. I could not find a published study that measures any of these three churns against a prediction system and the betting market together, so the numbers below are, as far as I can find, the first.

We score the models in this essay on log loss, where lower is better: A model that gives a team a 90% chance and is right is barely penalized, and one that gives a team a 90% chance and is wrong is penalized heavily. To read it in College Pick'em points, where a week has 55 possible total points, I measured the exchange rate on the 71 real College Pick'em weeks from 2021 through 2025: Each 0.01 of log loss is worth 0.74 points a week on the ballot, and that is the conversion used below.

Conference realignment

Figure 2. Every conference move since 2005. Press play: Each team that changed leagues draws a line from the league it left to its own stadium, this season in red, older moves in grey; a ring with no line is a program that came up from FCS. The buttons rank the twelve conferences by the system you pick; click a bar to draw that conference.

Press play in the figure above and the map draws every conference move since 2005. The SEC went from twelve members to fourteen in 2012, by adding Texas A&M and Missouri, and to sixteen in 2024, adding Texas and Oklahoma. The Big Ten added UCLA, USC, Oregon and Washington in 2024, and the Pac-12 collapsed to two schools in 2024 and is back to eight schools this season, five of the six newcomers from the Mountain West. Kennesaw State, where I currently work and study, is the grey dot beside Georgia Tech through 2023. In 2024 it moved up to FBS and joined Conference USA, where it won 2 games and lost 10.

The panel beside the map is the Machine's record, all conferences, one season at a time. Across the twenty-one seasons it picks between 69% and 77% of games correctly. 2024, the season of the biggest moves, was one of its four worst seasons, at 71%. The season summary gives the same record for Vegas and the AP poll on the movers' games.

Does the Machine do worse on a league's games in the season right after that league changed members? The answer is a small maybe. On those games it fell 0.005 of log loss further behind Vegas than usual, and the interval runs from −0.005 to +0.016, so the true number could be zero. Statisticians would say there's no significant difference. In College Pick'em points, the most realignment could be costing the Machine is a quarter of a point a week or less. Leagues that changed more did not do worse in any way I could measure.

Whether the teams that moved were harder to pick depends on who moved. Across every mover season since 2013, Vegas and the Machine both picked movers' games right a few points less often than everyone else's, and the reason is who does most of the moving. In 2013 it was twenty-three programs, most of them mid-majors stepping up into the American Athletic or Conference USA. Vegas and the Machine rated them on what they had done in their old leagues, and the new schedules were harder. In 2024 it was Texas, Oklahoma, Oregon and USC, teams that had already proven they could win in a power league, and a move to another one did not change that.

Marshall is different. It had been one of the better teams in Conference USA, moved into a stronger Sun Belt in 2022, and the Machine missed every September game it scored, including the week Marshall went to South Bend and beat Notre Dame. Marshall was not overrated or underrated so much as it was untested.

On the College Pick'em ballot, none of the four systems was measurably worse in 2024 than in other seasons, and the AP poll was the weakest of the four however the games were divided. The Crowd backed the movers more often than Vegas or the Machine did, and the movers won only 41% of those games. Most of the ground the Machine lost to Vegas over five College Pick'em seasons was lost on games with a mover, but more than half of those games came in 2024, one of the Machine's worst seasons anywhere, so the ballot cannot say whether that was the churn or the year. 2024 was the biggest realignment year on record, and my guess is that it broke the comparisons the Machine leans on: Half the Pac-12 was suddenly playing a Big Ten and Big 12 schedule, and results against the old neighbors no longer said much about the new ones.

Figure 3. Where the players go. Every transfer between FBS leagues in one offseason, as a band from the league a player left to the league he joined; width is players. Only transfers with an FBS origin and an FBS destination are drawn, about half of all portal entries. Hover a band or a league to highlight it; type a team to highlight the bands carrying its players.

In the figure above, set the year to 2025 and hover the SEC. In that year it put more players into the portal than any league and took back fewer than it lost. The Big Ten did the same. Now type Auburn and set the year to 2022: Seven players to other SEC schools, three to the ACC and three to the Pac-12. One of those three was Bo Nix, to Oregon.

The rosters

Figure 4. Who lost players and who gained them. Every FBS program, players out along the bottom and players in up the side; the diagonal is break-even, and dot size is the share of last season's production that came back. Type a team or click a dot to follow it, two at a time, with a trail as the seasons play; when the seasons run out the panel sums up the teams you followed, and it always gives the season's totals and how the three systems did in September on the teams that turned over the most.

In Figure 4, above, type Washington State and press play. Before the 2025 season it lost 70 players, the most any program has lost in the six seasons the portal has been counted, after the Pac-12 collapsed around it and its coach left. Then type Colorado and watch 2023: 59 out, 48 in, and 8% of the previous year's production on the field. The dashed diagonal is break-even: A dot below it lost more players than it brought in, and had to fill the gaps with recruits; a dot above it came out ahead in the portal. The dots at the far right are the programs rebuilt from scratch, and you would expect them to be the hardest teams in the country to predict.

The instinct is that a team like that is harder to predict. We found it to be easier. In weeks one to four, games involving a team in the top quarter of roster turnover were picked right 78% of the time by the Machine and 71% for all other games, a gap of 6.8 points with an interval from 1.6 to 12.2. Vegas gained 6.1 points on the same split. A team that lost forty players over the offseason is usually a bad team, and a bad team loses the way everyone expects. After week four the gutted-roster teams are picked at the same rate as everyone else.

Bo Nix entered the portal on December 12, 2021. He had started at Auburn since his true freshman year in 2019, when he beat Alabama in the Iron Bowl and was named SEC Freshman of the Year, and he kept the job for three seasons and two head coaches. Twenty-three more Auburn players entered the portal that offseason, and Auburn's net talent flow that winter was the eighth worst in the country. It opened 2022 with 45% of the previous season's production, went 5 and 7, and fired coach Bryan Harsin on Halloween. The Machine picked Auburn's games right 82% of the time that season. The churn made Auburn bad, and a bad team is easy to read.

Oregon is the other half. It had a new head coach, Dan Lanning, Nix at quarterback, and 16% of the previous season's production back. The prior rated the Ducks the fortieth best team in the country. They went 10 and 3 with Nix throwing 29 touchdowns, and the Machine went 1 for 3 on Oregon's September games. The next year Nix threw 45 touchdowns, completed 77% of his passes, finished third in the Heisman voting and went twelfth in the draft. One transfer made one team easy to pick and another team hard to pick, and the hard one was hard for exactly one month. A month of real games was all it took; from October on the Machine correctly picked seven of Oregon's last eight.

The teams that did two churns at once are where the prior misses. Arizona State moved to the Big 12 in 2024 with 28 players out, the prior expected it to win about a third of its games, and it won the league at 11 and 2. USC in 2022 had Lincoln Riley, a roster rebuilt through the portal and 22% of its production back, the prior said 40%, and it went 11 and 1. Sam Houston in 2025 had a new coach, the prior said an even season, and it went 2 and 10. Across 2021 to 2025 a team with two churns was a bad team on average, five wins in twelve, and the prior missed its record by about half again as much as it missed everyone else's. The prior can handle one change at a time. Two at once and it is guessing.

The third churn is the head coach, and on its own it does nothing to the numbers. One program in five changes coaches in an average year, and in September their games were picked right at the same rate as everyone else's, by the Machine, by Vegas and by the AP poll. A new coach matters because of what comes with him: Teams that change coaches are almost twice as likely to be in the top quarter of roster turnover, and teams that change leagues change coaches more often than teams that stay. When all three happen at once the team sinks. It has happened four times, Cincinnati and Charlotte in 2023, Arizona and Washington in 2024, and none of the four had a winning season.

Taking the Machine apart

The Machine's weakest month is September. Its early-season misses cluster by conference: Across 6,988 team-games in weeks one through four, SEC teams won 4.5 percentage points more often than the model expected, Big Ten teams 4.1 points more, and Conference USA teams 6.5 points less. This pattern vanished after week four. A correction was built, an offset by conference tier for the first four weeks only. It improved held-out early-season log loss by 0.0112 with an interval clear of zero, about a point a week in September, and it was adopted into the Machine’s prediction algorithm.

Figure 5. What two inputs were worth, by era. Remove one input from the Machine, rerun twenty-one seasons without it, and the loss is what the input was earning, in points a week at the exchange rate. Red is the preseason prior, grey is the conference flag; the band is the 95% interval. Weeks 1 to 4 is the default; the buttons switch to week 5 on and to all weeks.

Start with the grey line, the conference flag, the input that says whether a game is inside a conference. It hugs zero in every era. Without it the Machine loses about a tenth of a point a week in September and nothing after, and on movers' games since 2018 it did slightly better without the flag than with it. The name of a league is close to worthless to the Machine, and after a realignment the Machine did slightly better without it.

Now the red line, which is the preseason prior. It climbs: The Machine loses two thirds of a point a week in September without it, and that loss has grown to almost four times what it was in the 2000s. It is worth as much on movers' games as on anyone else's. Then click week 5 on and watch both lines drop to zero, because by then the season's own games have taken over. Realignment and the portal did not break the games. They broke the habit of judging a team by its league.

The Machine is six models taking a vote, and taking it apart lets us ask whether churn fooled any one of them. It did not. A gutted roster made September easier for all six, a new league made it harder for all six, and a new coach did nothing to any of them. Where they differ is what they lean on. The tree models lean hardest on the prior when a team has changed leagues, and the elastic net, which carries most of the weight in our ensemble, is the one model that never got anything out of the conference flag and was hurt by it on movers' games.

What it all means

Every number in this essay so far has been an average over thousands of games. Here are the interesting anecdotes: The single seasons the numbers got most wrong. The list below is the ten churned teams each system misread the most, ranked by how far the team finished from what the system expected, with every game in the window drawn as a cell filled to the system's confidence. Pick the Machine, Vegas or the AP poll, pick a kind of churn, and switch between September and the full season.

Figure 6. The ten teams each system misread the most. Ranked by how far the team finished from what the system expected, expected wins being the sum of the system's win probability over the team's games. Each cell is a game, filled to the system's win probability for the team, red where the call was wrong, W or L for the result. Hover a cell for the score, the ranks and the number. Three games minimum in September, eight over a season; Vegas from 2013; the AP poll has no probability, so its cells are full or empty, its list ranks by calls missed, and beat the number or fell short is the net of wins against the poll's side and losses with its backing. Every game counted is FBS against FBS; games against FCS teams have no number from any system and are shown dashed.

Three of the Machine's ten worst September misses are from 2012. Houston had gone 13 and 1 with Case Keenum at quarterback, Arkansas 11 and 2, Southern Miss 12 and 2 with a conference title, and each lost its head coach that winter: Kevin Sumlin to Texas A&M, Bobby Petrino fired in April, Larry Fedora to North Carolina. The prior carried the best season each program had had in years into a September with a new coach and, at Houston, no Keenum. Houston opened by losing at home to Texas State, a program in its first FBS season. Arkansas, ranked eighth, lost at home to Louisiana Monroe. Southern Miss went 0 and 12. A coaching change does nothing to the numbers in the average season. These were not average seasons. Each program had just had its best year in a long time, and the coach who built it was gone.

Switch to the AP poll. Notre Dame in 2005 was unranked in week one, coming off 6 and 6 with a new coach in Charlie Weis, and beat 23rd-ranked Pittsburgh and 3rd-ranked Michigan in its first two games with the poll on the other side both times. Louisville in 2007 was ranked eighth after Bobby Petrino left for the NFL, then lost at Kentucky and at home to Syracuse in back-to-back weeks; the Machine had Louisville at 96% against Syracuse. And Georgia Tech in 2024, after 3 and 9, 5 and 7 and 7 and 6 in the three seasons before and with 33 players out through the portal, opened in Dublin against 10th-ranked Florida State and won 24 to 21, with the poll and Vegas both on Florida State.

Prediction is a question of what you know before the first snap, and between seasons more of that stops being true every year. Colorado is a great example: The portal in 2023, the Big 12 in 2024, a roster this season that is 14% of the one before, and on Thursday night a win over Georgia Tech. The systems are not broken by that. They are merely humbled for a month. The only honest number in week one is a guess about a roster nobody has seen play, and the season is what corrects it.

How the Machine was built

Code: Python. The store is bitemporal: Every row carries the time it became knowable, and the model may only read rows knowable before kickoff. The check runs before every prediction.
Training data: 7,161 FBS games over the ten seasons before each prediction. The record by season is walk-forward on 14,735 games, 2005 through 2025, each season predicted from the seasons before it.
Models: LightGBM, XGBoost, CatBoost, NGBoost, an elastic net and a neural net, each fit five times and averaged, then weighted into one probability.
Churn data: Conference membership from the schedule, 2005 to 2026. Transfer portal entries and returning production from CollegeFootballData, 2021 and 2014 on. Head coaches from the same source, 2001 on; a change is a different head coach from the season before. Closing lines from 2013. The AP poll scored only on games where at least one team was ranked.
Tests: Every churn test was written down before it was run, with its minimum detectable effect stated before the result. Bootstraps are paired over seasons or conferences, 10,000 draws. Two ablations rerun all twenty-one seasons with one input removed, five seeds each, and are scored for the Machine and for each of its six models.
News: Headlines read daily by qwen3.6:35b-a3b running locally. Adjustments capped at 0.2 log-odds per fact and 0.3 per game, and every applied adjustment carries the sentence that caused it.
Hardware: An ASUS Ascent GX10 desktop for training, the ablations and the daily news run, a laptop for the rest.
Help: Built with Claude. Fable 5 ruled on what was worth testing and closed each item. Opus 5 wrote the code. The decisions, the errors and the choice of what to publish are mine.

References

Elo, A. E. (1978). The Rating of Chessplayers, Past and Present. Arco. The rating system behind the strength inputs.
Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society B 67(2), 301 to 320. The linear model that holds most of the Machine's weight.
Ke, G. et al. (2017). LightGBM: A highly efficient gradient boosting decision tree. NeurIPS 30. Chen, T. and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. KDD. Prokhorenkova, L. et al. (2018). CatBoost: Unbiased boosting with categorical features. NeurIPS 31. Duan, T. et al. (2020). NGBoost: Natural gradient boosting for probabilistic prediction. ICML. The four tree models.
CollegeFootballData.com. Schedules, results, transfer portal entries, returning production and head coaches, 2001 to 2026.