Inside the Versus Sports Simulator College Football Playoff and FCS Bracket Projections

By: Steve Pugh, Founder, Versus Sports Simulator
October 9, 2026

Twenty years ago I built a college football rating system for one reason: to take human bias out of it. Polls rank teams partly on reputation, so a famous program gets credit its results have not earned and a small school gets less than its results deserve. I wanted a number built from results and nothing else.

That system has run every season since. It was featured in Phil Steele's College Football Preview in 2015, and from 2020 through 2024 it was one of two computer ranking systems used in the selection process for the NAIA College Football Championship Series.

Even there, the ratings answered one question, how good is each team, and people answered the other one, who gets in. A rating can tell you who deserves a place. It cannot tell you who will get one, because a committee weighs things a rating is built to ignore. So projecting a bracket before it is announced means modeling the committee rather than the teams, and a model of what people will choose is only as good as the ratings underneath it. Two decades of those is what makes the attempt worth making now, and it also means putting human bias back in, deliberately, after twenty years of engineering it out.

So there are two models here. The ratings say how good each team is. A second model, built separately and sitting on top, predicts what a committee will do with them.

The Rating Model and the Selection Model

Everything below describes the selection model, which decides nothing about how teams are rated. The rating model is untouched by any of it and works the way it always has.

That separation matters because the selection model does things no sound rating system should. It counts how many top-25 teams a side has beaten. It gives a conference champion a nudge over a league rival with the same record. It charges a team for playing fewer games than everybody else.

None of those belong in a rating. Strength of schedule and strength of record are already inside the ratings implicitly, in the sense that beating good teams moves your number and beating bad ones does not. Counting your top-25 wins on top of that is double counting. It is also what a human does when scanning a resume, which is why it belongs in the selection model and not in the ratings.

The Four Decisions Behind a Playoff Field

Producing a bracket takes four decisions. A committee makes them in a room. The model makes them in code.

It decides who gets in, which means ranking every candidate against every other candidate. It judges how hard each team's record was to earn, because 9-3 against a brutal schedule and 9-3 against nobody are different achievements. It decides who travels where in the opening round, because in the FCS the format does not settle it. The bracket that results then settles a fourth question on its own: who is likely to win it.

The third of those exists only in the FCS. In the twelve-team College Football Playoff every team in the field is seeded and the format fixes the opening matchups, so nobody chooses who travels where and there are three decisions rather than four.

The model has to make all of them. The first is a judgment call it approximates and cannot verify. The rest are computation.

Selecting and Ordering the Field

Every team in contention gets a selection score. Lower is better, and the team with the lowest score is the number one seed.

\[ S_t \;=\; w_R\,r_t \;+\; w_S\,s_t \;-\; w_Q\,q_t \;+\; D\cdot\max\!\left(0,\; G - g_t\right) \]

Everything in that line is measured in places, which is the unit that makes the terms addable. \(r_t\) is the team's projected divisional rank. \(s_t\) is its resume rank. \(q_t\) is credit for who it beat. \(g_t\) is games played, against \(G\), the number of games most of the division played.

The weights are \(w_R = w_S = w_Q = 0.5\) and \(D = 10\). Those four numbers are the only invented quantities in the entire system. They are an educated guess at how a committee weighs a rating against a resume. Everything else is measured or derived.

Half the score is how good the model thinks a team is. Half is what it actually did. The credit for beating good teams is deliberately small, a thumb on the scale rather than a second record: a committee notices who you beat, but it does not seat a two-loss team above an undefeated one because of it.

The Quality-Win Credit

Beating a top-25 team earns credit on a sliding scale, so that beating the best team in the division is worth considerably more than beating the 25th.

\[ q_t \;=\; \sum_{j \,\in\, \text{wins}(t)} \frac{26 - \text{rank}(j)}{25}\;,\qquad \text{rank}(j) \le 25 \]

A win over the top-ranked team scores 1.0, which at \(w_Q = 0.5\) moves a team half a place. A win over the 25th scores 0.04. A team with several of those wins might climb a place or two. That is enough to settle an argument between two similar resumes, and not enough to lift a team over one with a better record.

Teams That Played Fewer Games

Not every team plays the same number of regular-season games. In 2025 the FCS split 106 teams at twelve, 14 at eleven and 9 at ten, so a ten-game team is being compared against a field that mostly played two more.

Fewer games means less evidence, and less evidence should cost a team something. So the score charges a team for each game it is short of what most of the division played. Part of that is already carried by strength of record, since fewer games means fewer chances to build a record a strong team would struggle to match, and this term sits on top of it.

The charge is a subtraction rather than a ratio because the rest of the score is measured in places, and it falls on any team short of the field's game count rather than on one conference. A team that loses a game to weather takes the same hit.

The term adjusts the Versus ordering for thin evidence, and that is the whole of its scope. Whether a selection committee does anything similar sits outside what this method can observe.

One subtlety. Mid-season the comparison cannot be against twelve, because nobody has played twelve and the charge would hit every team for six or seven games. When the question is "if the playoffs started today," the benchmark becomes the median game count in the division instead.

Strength of Record

Strength of schedule and strength of record sound alike and measure different things. Strength of schedule asks how hard a team's opponents were. Strength of record measures how unlikely it would be for an ordinary team to come out of that same schedule with that same record.

We compute it exactly. Take the 25th-ranked team in the division as a benchmark. Walk the real team's schedule game by game and ask what the benchmark's chance would have been in each game, at the same venue, home or away or neutral. That gives a list of independent probabilities \(p_1,\dots,p_n\), each one different.

The number of games the benchmark would win is then a Poisson binomial random variable, which is a sum of independent coin flips with different biases. There is no closed form, but there is an exact recurrence. Let \(f_j(k)\) be the chance of exactly \(k\) wins after the first \(j\) games:

\[ f_j(k) \;=\; f_{j-1}(k)\,\bigl(1-p_j\bigr) \;+\; f_{j-1}(k-1)\,p_j \]

starting from \(f_0(0) = 1\). Run that over the whole schedule and the strength of record is the upper tail at the team's actual win total \(w\):

\[ \mathrm{SoR}_t \;=\; \Pr\!\left[X \ge w_t\right] \;=\; \sum_{k \ge w_t} f_n(k) \]

A low probability is an impressive record. If an ordinary team would almost never have gone 11-1 against that schedule, then going 11-1 against it means something.

The two can point in opposite directions. A team with a soft schedule can still own a record that was hard to produce, because an ordinary side dropped into that same schedule would have lost more often than it did. The live pages rank both columns separately so the gap is visible for every team in the field.

The Blended Resume Term

The resume half blends two orderings rather than using either alone. The reason is structural: the rating and strength of record are both built from schedule quality, so using one for each half of the score counts the same evidence twice, and a side with a thin record but a brutal schedule climbs past teams that simply won more.

Raw record is the independent signal, and no strength measure reproduces it. A committee weighs it heavily and so does this:

\[ s_t \;=\; \mathrm{rank}\!\left(\frac{\mathrm{rank}_{\text{win\%}}(t) \;+\; \mathrm{rank}_{\mathrm{SoR}}(t)}{2}\right) \]

That keeps the schedule correction without letting it swamp the record.

First-Round Travel

Of the four decisions, this is the one that is a true optimization problem, and it exists only in the FCS.

In the FCS, 24 teams make the field and only the top 16 are seeded. Seeds 1 through 8 get a bye. Seeds 9 through 16 host a first-round game. The other eight teams carry no seed at all, and nothing in the format says which of them travels to which host. The model works to two aims, both taken from what past brackets have looked like: nobody opens against a team from their own conference, and the field travels as few miles as it can.

Eight hosts, eight visitors, each visitor assigned to exactly one host: that is a minimum-cost perfect matching on a bipartite graph.

\[ \min_{\sigma \in S_8} \; \sum_{i=1}^{8} c\bigl(h_i,\, v_{\sigma(i)}\bigr) \]

The two rules are not equally important, and trading one away for the other would be wrong, so the cost function layers them:

\[ c(h, v) \;=\; d(h, v) \;+\; 10^5\cdot\mathbf{1}\!\left[d(h,v) > 400\right] \;+\; 10^7\cdot\mathbf{1}\!\left[\text{same conference}\right] \]

where \(d\) is the great-circle distance between the two campuses. The magnitudes keep the layers from being traded against each other. The worst total travel eight games can produce is well under 25,000 miles, so a 100,000 charge for breaking the 400-mile guideline can never be bought back by shorter trips elsewhere. Eight of those still come to under 10,000,000, so a conference rematch is always the last thing given up. The rules stay in priority order, and the assignment never fails: if avoiding every rematch is impossible, the solver gives up the cheapest one.

Campus coordinates come from the US Department of Education's college directory, which is public domain.

Why Brute Force Does Not Scale

Eight hosts and eight visitors is \(8! = 40{,}320\) arrangements. That is small enough to brute force once and far too slow to run inside a thousand simulated seasons.

A greedy pass is fast but not correct: give the best host its nearest visitor, repeat down the list, and the last host can be left with a trip across the country.

A subset walk solves it exactly. Let \(\mathrm{dp}[M]\) be the cheapest way to fill the first \(|M|\) hosts using exactly the visitors in the set \(M\):

\[ \mathrm{dp}\bigl[M \cup \{v\}\bigr] \;=\; \min_{v \,\notin\, M} \; \Bigl( \mathrm{dp}[M] \;+\; c\bigl(h_{|M|},\, v\bigr) \Bigr) \]

with \(\mathrm{dp}[\varnothing] = 0\). There are \(2^8 = 256\) subsets and 8 candidate visitors at each one, so the whole thing is about two thousand steps instead of forty thousand, and the answer is the true optimum rather than a good guess. In general it is \(O(2^n n)\) against \(O(n!)\).

The saving against naive seed pairing runs to thousands of miles in a typical FCS bracket. Every trip, its host site and its mileage are printed on the live page, so the arithmetic is checkable rather than asserted.

Championship Odds

Once the bracket exists, each team's chance of reaching each round follows from it. The bracket is simulated 20,000 times, resolving every game from the same prediction model that powers the Versus game simulator.

Two numbers on the page measure different things, and multiplying the first set will not reproduce the second. The percentage beside a team in the bracket is its chance of beating the opponent shown in that slot. Those percentages do not multiply into a title chance, because the opponent shown still has to get there.

Take a team shown at 60% in the quarterfinal, 55% in the semifinal and 50% in the final. Those multiply to 16.5%, and its real title chance will be higher. The product answers a narrower question: the chance that team wins the title and every favorite ahead of it also wins. That is one path through the bracket. The semifinal and final figures are each quoted against one specific opponent, and that opponent may not arrive. When a weaker team arrives instead, the chance is better than the figure shown, and all of those paths count toward the title number as well.

The bracket is small enough to solve exactly, which is how the simulation was checked. Carry a full distribution over who occupies each slot. For a game between slots \(A\) and \(B\):

\[ \Pr[\,x \text{ wins}\,] \;=\; \sum_{y} \Pr[A = x]\,\Pr[B = y]\,\Pr[\,x \text{ beats } y\,] \]

Recurse to the final. Run against the 20,000-season simulation, the exact answer agrees to within 0.10 percentage points across all 22 teams with a nonzero chance, and the exact probabilities sum to 1.0000.

Scoring the Method Against the Real Brackets

Our own scores database holds the brackets the committee actually produced, so the method can be scored against them.

Run this selection over the 2024 and 2025 seasons and it picks 90% of the teams the committee really took, and places about two thirds of them in the correct group of eight: the byes, the hosts, or the unseeded. Measuring the resume by winning percentage alone scores worse on both counts, which is the evidence that made strength of record part of the blend.

The travel solver gets the same treatment. Neither the 2024 nor the 2025 first round put two teams from the same conference in the same game. Across every first-round game since 2013 we can place on a map, the median trip was 404 miles, which is where our 400-mile threshold came from.

The eligibility rules are handled at the same level of detail. The MEAC and SWAC send their champions to the Celebration Bowl instead of the playoffs, so the model holds those teams out of the field, along with the SWAC runner-up, whose conference title game falls after the bracket has started. Everybody else in those two leagues remains eligible for an at-large place, and they take them: North Carolina A&T in 2016, Florida A&M in 2021 and North Carolina Central in 2023 all played first-round playoff games while a different team from their league was in the Celebration Bowl.

Where the Projections Live

The projection runs for both subdivisions and rebuilds every week from that week's ratings.

• FCS playoff bracketology, a 24-team field
• FBS playoff projection, the 12-team College Football Playoff

Each page carries four views behind two toggles, because a single answer was being asked to settle two separate questions. Picked by chooses between a selection committee and the Versus ratings on their own. Field as of chooses between the season simulated to its end and the body of work as it stands today. The gap between views can be wide. A team with a strong rating and a thin record can sit near the top of one view and well down another in the same week, and both readings are defensible: one asks what the full season is likely to produce, the other asks only what has been earned so far.

Every week is kept at its own permanent address and never edited afterward. And the whole projection is published as data as well as a page, at playoff-projection.json, so an analyst can pull the field, the seeds, the bids, the ranks and every first-round trip without scraping anything.

None of this is a ruling on who deserves to be in. It is a model of a human process, and the teams where a rating and a resume disagree sharply are the ones any projection, this one included, is most likely to get wrong. Those are the seeds worth arguing with, and the method behind every one of them is published here each week.

Frequently Asked Questions

How accurate is the Versus Sports Simulator playoff projection?

Run over the real 2024 and 2025 FCS brackets, the method selects 90% of the teams the committee actually took and places about two thirds of them in the correct group of eight: first-round byes, hosts, or unseeded. Scored against the 2025 College Football Playoff it selected 10 of the 12 teams and placed every one of those in the right part of the bracket.

What is the difference between strength of schedule and strength of record?

Strength of schedule measures how hard a team's opponents were. Strength of record measures how unlikely it would be for an ordinary team to produce that same record against that same schedule. A team can have an easy schedule and still have an impressive record, and the two columns are ranked separately on every Versus playoff projection page.

How many teams make the FCS playoffs and how many get a bye?

Twenty-four. Eleven places go to conference champions as automatic bids and thirteen are at-large selections. The top sixteen teams are seeded, and the top eight of those receive a first-round bye. The remaining eight teams carry no seed at all.

How does the model predict which unseeded team travels to which host?

It solves a minimum-cost assignment problem over all eight first-round games at once, using campus coordinates from the US Department of Education. Two aims are ranked by cost: no team opens against an opponent from its own conference, and the field travels as few miles as the geography allows, staying within 400 miles of home where possible. The committee makes the real pairings; the model estimates them.

Does winning a conference improve a team's seed?

In the FCS model, slightly. Every conference champion receives an automatic place in the field, and beyond that the selection order gives a champion a small edge over a team from its own conference with a similar record. The FBS model gives no seeding credit for a title, because the twelve-team format seeds straight by ranking.