What the transfer portal is worth in our CFB model

North Texas brought in the biggest net transfer haul in the country this offseason, plus 16 bodies. Read that the way most August takes do and you would move the Mean Green up. Our college football rating moves them down for it, a little. Not because the portal is worthless, but because the raw count of transfers turned out to be one of the few inputs in our model that points the wrong way.

The rating is a linear composite, a constant plus six inputs, and two of those inputs describe the portal. One is net transfer flow, the plain count of players in minus players out. The other is the average star rating of the class coming in. We fit the whole thing with ordinary least squares on the 2021 through 2023 seasons, then let the weights fall where the data put them. Here is what the two portal terms came out to:

- 0.0947 * portal_net
+ 1.7239 * tin_stars

The count is negative. The quality is positive. Holding everything else in the model constant, an extra net transfer shaves about a tenth of a rating point, and a better incoming class adds points. That sign on the count surprises people, so it is worth being exact about what it does and does not say.

What the negative weight is actually saying

It is a partial coefficient. In a model that already knows your returning production, your four year talent base, and the quality of the players you brought in, the leftover raw count is not measuring reinforcement. It is mostly measuring churn. The teams at the very top of the net inflow board are not loading a finished roster with extras. They are rebuilding one.

Look at who sits there. North Texas at plus 16 returns 2.5 percent of last year’s production. James Madison at plus 10 returns 8 percent. South Florida at plus 13 returns 12.6 percent. These are rosters that lost almost everything and refilled through the portal, and the count is the fingerprint of the teardown, not evidence against it. The model prices the rebuild through returning production and talent, then the raw count picks up the small residual that says a heavy churn year is a little riskier than a stable one.

What it is not is a strong law of nature. Across the 132 teams we have portal data on, the plain correlation between net inflow and returning production is weak, about negative 0.07. Plenty of teams run net negative and keep their best players. Ohio State is minus 22 and returns 68 percent. South Alabama is minus 18 and returns 79 percent, a roster that kept its starters and let depth walk. The count only earns its negative sign once the model has already accounted for everything else, and even then the effect is mild. North Texas at plus 16 loses about 1.5 rating points from the count term. That is a nudge, not a verdict.

Where the portal signal actually lives

The signal is the other term. Average incoming star rating carries a positive weight, and it is the number our receipts should cite when we talk about a portal class doing something for a team.

The cleanest way to see the split is two teams near the top of the same board. North Texas and Ole Miss both ran huge net inflows, plus 16 and plus 13. On the raw count they look alike, and both take a similar knock from the count term, about 1.5 and 1.2 points. But Ole Miss brought its transfers in at a 3.30 average star rating, near the top of the sport, and returns half its production on top of them. North Texas brought its haul in at 2.96, below the average of 3.05, onto a roster with almost nothing back. Same headline volume, and the model separates them on quality and continuity, which is exactly where the difference is.

That is the whole point of carrying two portal numbers instead of one. Bodies are not talent. Stars are, and the star term is how the model says so.

What this does not license

Two limits ride with this, same as everywhere else we describe the rating. First, this is a preseason snapshot. Portal and coaching fields move until kickoff, and the values here are the inputs frozen with the model on the day it was built, so a live intel table will drift past them as classes fill in. Second, the rating validates predicting actual wins, not beating a market. It misses by about 1.9 wins per team on average out of sample, which is why any one team’s number is a lean and not a lock, and why we make no closing line or return claim on it until a full season of posted lines is stored and graded.

So when a team wins the offseason on the portal leaderboard, that is a fact about volume. Whether it moves our number is a separate question, and the answer runs through who they brought in and how much they kept, not how many. The formula that does the sorting is public, printed to the coefficient at edgelabs.bet/research/the-cfb-model-in-seven-numbers, and the full field it produces is at edgelabs.bet/cfb/power-ratings. Founding access is open at edgelabs.bet/join.