In a landscape dominated by raw spreadsheets and opaque analytics, Ray Carpenter makes a bold claim: the chaotic, emotional world of college football transfers can be tamed with a specific blend of logistic regression and gradient boosting. He doesn't just offer a dashboard; he offers a predictive engine built for a market that has exploded in volume but lacks sophisticated tools. This matters now because the 2026 transfer window is closing, and for the first time, a free resource attempts to quantify the "why" behind a player's move, not just the "where."
The Architecture of Prediction
Carpenter acknowledges the gap in the market immediately. "There are tons of amazing, free resources out there that track the global soccer transfer market, but I don't see any out there yet for college football that I feel are quite up to snuff." This admission sets the stage for his project, TransferPortl, which aims to replicate the depth of soccer analytics for American football. The core of his argument rests on the idea that transfer behavior isn't random noise; it follows patterns that can be modeled. He builds a database of 30,000 players, pulling from APIs and web scraping to create a foundation that includes everything from high school star ratings to coaching staff directories.
"In the age where anyone can create a machine learning model, I think it's really important to explain my steps thoroughly, and tell you all what I include in the model and what I don't."
This transparency is the piece's greatest strength. Rather than hiding behind a "black box" algorithm, Carpenter details a two-step modeling process. He uses logistic regression to establish a baseline probability of a transfer based on core features like position and conference, then layers XGBoost on top to catch the non-linear interactions—like how a coach's departure might impact a specific position group differently than another. Critics might note that relying on historical data from 2021 to 2025, a period of unprecedented volatility in college football rules, could skew future predictions. However, Carpenter's willingness to expose the model's mechanics allows users to weigh the evidence themselves.
Beyond the Numbers: The Human Element
The most compelling aspect of Carpenter's coverage is his attempt to quantify the intangible: loyalty and geography. He introduces the Transfer Vulnerability Score (TVS), a metric designed to predict how likely a player is to leave their school. This isn't just about stats; it's about context. "I created a multi-ring buffer around college campuses to create a 'player home team' variable after noticing that hometown players tend to stick around through coaching changes." This heuristic attempts to capture the emotional tether a player has to their local community, a factor often ignored by purely statistical models.
"The longer a player stays in the portal, the more likely he is to return to his school."
This observation flips the common narrative that portal entry guarantees a transfer. Carpenter's data suggests that time in the portal acts as a filter, where players who don't find a fit quickly often return home. He also tackles the tricky business of Name, Image, and Likeness (NIL) valuations. Using Bayesian regression, he estimates market value based on production and recruiting ratings, acknowledging that a five-star recruit's value can deteriorate if they don't play, while a walk-on's value can skyrocket with performance. "High school star rating interacts in a way where it deteriorates over time, to try and account for scenarios where a 5 star freshman recruit who didn't play much in his first year of eligibility should still be valuable." This nuance is vital for understanding the modern player market, where past glory counts for less than current production.
The Limits of the Model
Carpenter is refreshingly honest about the flaws in his own work. He admits that football positions are "highly bucketed," meaning quarterbacks and wide receivers have rich data, while offensive linemen and kickers are underserved. "You'll observe grades, market values, and transfer vulnerability scores that feel 'better' for positions that have more tracked metrics... The kicker and offensive line departments are currently lacking." He also confesses to the limitations of his "hometown factor," noting that a SQL workaround for missing data might misclassify players who traveled far for high school. "I'm not sure I'll ever be able to properly loyalty context though. There are tons of Michigan alumni who live in Columbus."
"I felt guilty of just slapping a gradient boost on this thing, but I can always go back and tweak it later."
This self-deprecating honesty builds trust. He acknowledges that the rush to launch before the portal window closed forced some compromises, such as the choice of SQLite over DuckDB due to dependency errors. While a counterargument could be made that a rushed model is dangerous for decision-making, Carpenter frames this as an iterative process. He invites users to request data edits and acknowledges that the search bar and individual player pages are still works in progress. The model's current .73 AUC score is a starting point, not a final verdict.
Bottom Line
Ray Carpenter's TransferPortl represents a significant leap forward in democratizing college football analytics, moving beyond simple rosters to predictive modeling that accounts for both data and human behavior. Its greatest vulnerability lies in the uneven quality of data across different positions and the inherent difficulty of quantifying loyalty, but Carpenter's transparent methodology and willingness to iterate make it a tool worth watching. For the busy analyst or fan, the real value isn't just the predictions, but the framework for understanding why the game is changing.
"In the age where anyone can create a machine learning model, I think it's really important to explain my steps thoroughly, and tell you all what I include in the model and what I don't."