Roster Churn and Transfer Portals: A Sports Analytics Project for Predicting College Athlete Movement

Why roster churn is a perfect semester-long analytics project

The modern NCAA roster is no longer a static list of athletes. It is a moving system shaped by recruiting, coaching changes, playing-time expectations, NIL considerations, injuries, and the transfer portal. That volatility creates a rich, realistic setting for a student data science project because the problem is both predictive and causal: you are trying to forecast who moves next while also asking why movement happens. In that sense, the project resembles other dynamic systems where timing, incentives, and shocks matter, similar to how analysts interpret live sports moments or how operators manage uncertainty through news-shock planning.

The Lady Vols example is especially instructive. On the same day, Tennessee saw a high-profile recruit decommit and a current player enter the portal, leaving the roster extraordinarily thin. For a student project, that kind of event is gold: it invites you to model roster volatility as a sequence of decisions rather than a single outcome. A well-designed semester project can teach data collection, feature engineering, classification, survival analysis, and causal inference in one coherent narrative, much like a disciplined research workflow in trend-based content research or operational planning in portfolio decision-making.

Pro tip: A strong sports analytics project should not only answer “who will transfer?” It should also answer “which factors appear before movement, which factors merely coincide with it, and which interventions might reduce it?”

That distinction is what elevates the project from a prediction exercise to a research-methods lesson. Students can build a baseline model, test more sophisticated algorithms, and then ask whether the transfer portal represents cause, symptom, or both. If you structure the semester correctly, they will come away with a portfolio piece that signals technical ability and analytical judgment, while also developing a sharper understanding of NCAA instability and performance forecasting.

Framing the research question: prediction, explanation, and timing

Define the unit of analysis

The first decision is what exactly you are predicting. In a roster project, the unit of analysis can be the athlete-season, athlete-month, team-season, or even athlete-day during active portal windows. For most undergraduate or master’s students, the best compromise is the athlete-season because it gives enough observations while remaining tractable. You can label each player-season as one of several outcomes: stayed, transferred, entered the draft, decommitted, or retired from collegiate competition. This multilevel framing is more informative than a binary yes/no transfer label because it captures the full range of roster movement.

Once the unit is defined, students should distinguish between descriptive and predictive questions. Descriptive questions ask how roster volatility changed over time, whether portal activity rose after rule changes, and whether certain conferences behave differently. Predictive questions ask which players are most likely to move in the next cycle. Causal questions ask whether playing time, coaching changes, recruiting rank, or team performance actually influence movement after adjusting for confounders. In a research-methods pillar, this three-part structure matters because it teaches students that one dataset can support several analytical lenses, just as a publishing workflow may need both conversational search thinking and formal documentation.

Choose a realistic scope

A semester is short, so the scope should be ambitious but bounded. The strongest version of this project focuses on one sport, usually women’s or men’s basketball, because roster turnover is frequent, recruiting data is accessible, and transfer narratives are often public. A Tennessee-inspired case study works well because it naturally raises questions about elite recruiting, immediate roster gaps, and the pressure created by roster scarcity. Students can compare a focal program like Tennessee to peer programs with similar recruiting profiles, which turns a one-team story into a comparative study.

This is also a great place to teach the difference between a compelling case and a biased sample. The Lady Vols story may motivate the project, but the model should not be trained on one team alone. Use a multi-team NCAA sample, then reserve Tennessee as a “motivating example” or qualitative vignette. That approach protects against overfitting to a single program and mirrors how analysts use a standout case without mistaking it for the whole market, similar to studying customer churn during leadership change or tracking roster effects like a portfolio manager would track risk shifts.

Translate the question into testable hypotheses

Students should end the framing phase with explicit hypotheses. For example: athletes with lower playing time are more likely to transfer; higher recruiting rank reduces transfer probability in the first year but may increase it later if expectations are unmet; coaching changes increase short-term portal entries; and teams with greater roster instability have more variable performance. Hypotheses keep the project anchored in theory instead of turning it into a blind machine-learning contest. They also guide feature engineering and interpretation later in the semester.

To deepen the research-methods angle, require each student or team to submit a one-page pre-analysis plan. That plan should list the outcome, predictors, model family, validation strategy, and one causal inference design. Pre-specification reduces hindsight bias and trains students to think like researchers rather than hobbyists. It is the sports-analytics equivalent of a structured launch checklist, the kind of discipline seen in operational checklists and trust-building when deadlines slip.

Data sources: building a credible NCAA movement dataset

Roster data

The core dataset should combine historical rosters, player bios, class years, positions, height, hometowns, and eligibility status. Roster pages are often the cleanest source of structured information, though students will likely need to scrape or manually compile data from school sites and public archives. For a more sophisticated workflow, encourage versioned snapshots: one roster at the start of the season, one midseason, and one after the portal opens. That lets students observe how the same team evolves across time rather than treating rosters as fixed objects.

Because roster volatility is the phenomenon of interest, students should create explicit status codes. For example: committed, signed, enrolled, active roster, redshirt, transferred out, transferred in, and portal unknown. A clean status schema is essential because messy labels are the most common source of analytical confusion. This is a useful moment to talk about data lineage and auditability, a concern that appears in other domains too, such as delivery disruption tracking and payments risk monitoring.

Recruiting data

Recruiting rankings are the project’s most intuitive signal. Sources like national prospect rankings, position rankings, star ratings, and high-school honors help students operationalize pre-college talent expectations. Recruits with elite rankings may behave differently from under-the-radar players because they face different pressure, different playing-time trajectories, and different transfer incentives. If a freshman is a top-10 recruit but is buried on the bench, the portal may become an exit option sooner than for a lower-profile recruit with modest expectations.

Students should also build recruiting-context features, not just raw rank. Examples include recruiting class rank for the team, whether the player joined a top-25 class, and whether the player’s recruiting profile matched the team’s competitive level. A strong recruit on a struggling team may show different movement patterns than the same recruit on a contender. That contextual framing resembles how analysts interpret relative value in decision matrices for high-demand products or how marketers use comparative positioning in category prioritization.

Transfer portal and news events

The transfer portal itself is a crucial event source, but students should not rely on it alone. News articles, team announcements, social posts, and beat-reporter updates often reveal timing, motivation, and destination. These details matter because transfer behavior is not just a label; it is a process. A player may first lose playing time, then appear in rumors, then enter the portal, then commit elsewhere. Capturing that sequence improves model quality and enables event-history analysis.

For the causal component, collect team-level shocks: coaching changes, postseason bans, scholarship shortages, injury clusters, and conference realignment. These shocks can be coded as binary or time-varying covariates. Students may also use schedule strength, win-loss record, and minutes distribution as contextual controls. The point is not to assemble every conceivable variable, but to create a defensible dataset with enough signals to support both prediction and explanation. If students want an analogy for balancing signal and noise, point them to how analysts separate meaningful market movement from cosmetic changes in sales surge timing or how content teams handle audience discovery through conversational search.

Feature engineering: turning roster facts into movement signals

Player-level features

Player-level features are where students can be especially creative. Minutes played, starts, usage rates, shot attempts, games missed, and role changes are obvious predictors of transfer risk. But the best projects go further by calculating within-team percentile rank for minutes, average age of teammates at the same position, and distance from home to campus. If a player’s role shrinks while peers at the same position absorb minutes, that may indicate a likely exit. If a player is far from home and has no stable role, the transfer signal can become stronger.

Students should also create “expectation gap” variables: recruiting rank versus actual playing time, class year versus role, and preseason projected importance versus realized importance. These are particularly useful because movement often follows disappointment relative to expectations, not just absolute underperformance. That mirrors how consumers judge value relative to promise in areas like limited-release purchasing decisions or how individuals assess service quality versus advertised value in hidden-fee environments.

Team-level features

Roster churn is rarely an individual-only phenomenon. Team-level features such as coaching continuity, offensive system changes, scholarship pressure, and win-loss volatility can strongly influence movement. A bench player on a stable, winning team may stay put even with limited minutes if development pathways look promising. The same player on a team undergoing staff turnover may view the portal differently. That makes team context essential for predictive modeling and causal inference alike.

A useful aggregate measure is roster volatility index, defined as the proportion of players on a roster who enter, exit, or change status in a season. Students can also compute concentration metrics such as the share of minutes played by the top five players. Teams with extreme concentration may produce more dissatisfied role players and therefore more portal activity. This is a great bridge to general analytics principles, including the notion that some systems perform best when managed like a portfolio rather than a fixed list, a lesson echoed in portfolio orchestration.

Temporal and policy features

Because the NCAA environment changes, students should add time-based variables. These include portal window timing, pre- and post-policy eras, and season phase. A transfer decision made in January may have different determinants than one made after the season. If the course is advanced, students can even model event timing by week or month. That creates a richer dataset and supports survival analysis, which is ideal for studying when movement occurs rather than just whether it occurs.

Temporal features also connect to institutional change. Rule updates, scholarship rules, and NIL-related shifts can all reshape incentives. The key pedagogical point is that prediction in sports analytics is rarely timeless. Models trained on last decade’s NCAA landscape may fail if the transfer environment evolves. That is why students should think about distribution shift, a concept as important in sports as in volatile publishing calendars or leadership-transition churn systems.

Modeling strategy: from baseline classification to causal inference

Start with a transparent baseline

The first model should be simple and interpretable. Logistic regression, regularized regression, or a decision tree gives students a baseline to beat and teaches them how coefficients or splits relate to roster churn. Baselines also help reveal whether the signal is strong enough to justify more complex methods. If a simple model performs reasonably well, students may learn that clarity beats complexity for many research questions. If it performs poorly, that opens the door to richer feature interactions and nonlinear methods.

Students should report not just accuracy, but also precision, recall, F1, calibration, and area under the ROC curve. In roster prediction, calibration matters because a 0.70 probability should actually mean about seven out of ten comparable cases lead to movement. Well-calibrated probabilities are especially useful for coaches, compliance teams, and analysts who need risk scores rather than just class labels. If students are new to model evaluation, you can analogize it to evaluating buying decisions with both price and timing rather than relying on one flashy metric, much like advice found in consumer timing guides.

Move to predictive modeling

Once the baseline is set, students can compare random forests, gradient boosting, and XGBoost-style approaches. These models often capture nonlinear interactions such as “high recruiting rank plus low minutes plus coaching change” or “injury plus position crowding plus portal proximity.” However, advanced models should not be treated as magic. They need careful validation, feature importance checks, and fairness review. The best student projects explain why a model worked, not just that it did.

Cross-validation should be done in a time-aware way. Do not randomly split athlete-seasons if the goal is to forecast future movement; that leaks information from the future into the past. Instead, train on earlier seasons and test on later ones. This mimics real-world deployment and makes the project far more credible. Students can frame this as a sports-analytics analog of holding out future campaigns, much like market forecasters do when building trend-based content calendars or when teams manage uncertainty in delayed launch environments.

Use survival analysis for timing

Transfer decisions are not just yes/no events; they occur over time. That makes survival analysis a powerful addition to the project. A Cox proportional hazards model or discrete-time hazard model can estimate how each variable affects the risk of a transfer in any given period. For example, the model might show that reduced minutes accelerate exit risk, but only after a player has completed one full season. This approach is especially good for teaching event history, censoring, and time-to-event logic.

Survival analysis also helps students avoid a common mistake: treating all players who eventually transfer as identical. Some leave immediately, some after a redshirt year, and some after coaching turnover. Timing matters because the reasons may differ. When students see this, they begin to think like methodologists rather than just coders, which is exactly the goal of a research-methods pillar.

Causal inference: separating correlation from plausible mechanism

What can be treated as “treatment”?

Causal inference makes the project more mature, but it requires restraint. Students cannot claim that playing fewer minutes “causes” transfer without careful design, because player ability, injuries, and coach preferences all confound the relationship. Still, they can ask more modest causal questions. For example: does a coaching change increase portal entries among non-starters? Does a recruiting mismatch lead to higher transfer risk after adjusting for team quality and class year? These are plausible and researchable.

The easiest treatment variable is a coaching change or a sudden roster shock. Students can compare pre- and post-event transfer rates using difference-in-differences, matching, or interrupted time series if enough data exist. Another route is propensity score matching, where players with similar recruiting rank, position, class year, and minutes are compared across teams with different stability levels. These designs do not prove causality in the strongest sense, but they bring rigor to a domain often dominated by narrative.

Potential outcomes thinking

Introduce the potential outcomes framework in plain language: what would have happened to this player if the team had not changed coaches? What if the player had received 10 more minutes per game? Students do not need to estimate every counterfactual perfectly, but they should understand that the observed roster path is only one of many possible paths. That conceptual shift is one of the most valuable lessons in the semester.

To keep the project honest, require students to identify assumptions and threats to validity. Selection bias is huge: the most visible players are not random, and transfer portal reporting is incomplete. Reverse causality is also possible, because a player may lose minutes because a transfer decision is already underway. Discuss these issues explicitly, and students will learn scientific humility. That habit of careful interpretation is as important in analytics as in fields that rely on trustworthy signals, such as ethics and sponsored reporting or evidence-based claims in litigation.

Practical causal designs for class projects

A feasible design is a matched comparison of players before and after a coach leaves. Another is a team-level difference-in-differences analysis comparing schools with high transfer inflow/outflow to similar schools without a shock. A third is event study analysis around portal openings or major roster announcements. Each of these can be taught with accessible software and moderate statistical maturity. The key is choosing a design that fits the available data rather than forcing a glamorous method onto a small sample.

If students want a simple, memorable guiding principle, tell them: prediction asks who is likely to move; causal inference asks what changed the odds. Both matter, but they answer different questions. Mixing them without warning is a common error in sports analytics and in many other data-driven fields, from local demand sensing to churn monitoring.

Project workflow, deliverables, and semester timeline

Weeks 1–3: question, scope, and dataset

In the opening weeks, students should finalize the research question, identify target schools or conferences, and build the roster-recruiting-transfer database. This is also when they should draft data dictionaries and decide how to handle missingness. A clean, well-documented dataset often matters more than a complex model, especially in a class environment where reproducibility counts. Encourage students to store code, notes, and data snapshots in a shared repository from the start.

Weeks 4–7: exploratory analysis and feature engineering

Next, students should profile the dataset: transfer rates by position, by recruiting tier, by year, and by team stability. They should visualize roster churn over time and identify outliers like the Lady Vols-style extreme volatility case. This phase should also include feature creation and data-quality checks. Visual storytelling here is powerful because it shows how a roster can change from a crowded pipeline to a nearly empty bench, the kind of pattern that is easy to miss without structured analytics.

Weeks 8–12: modeling and validation

Students can build the baseline, then add more advanced models and compare performance. A good report will include a model table, confusion matrix, calibration plot, and variable-importance discussion. Students should also document time-aware splits and any leakage prevention steps. If the project includes causal inference, this is when matching or difference-in-differences should be estimated and diagnosed. Encourage students to write interpretation in plain language, not just statistical jargon, because the final audience should understand why the results matter.

Weeks 13–15: synthesis and communication

The final deliverable should look like a professional research memo rather than a homework assignment. Ask for an executive summary, methods appendix, limitations, and recommendations for coaches or analysts. Students may also create a short dashboard or poster with the main prediction results and one causal chart. The best submissions will tell a coherent story: what roster volatility looks like, what predicts it, what might reduce it, and how the findings change the way we think about NCAA program management.

How to evaluate model quality and research credibility

Use the right metrics

Accuracy alone is not enough, especially when transfers are relatively rare compared with stays. Students should prioritize recall, precision, F1, ROC-AUC, and calibration. If the class is more advanced, they can also use Brier score and decision curves. These metrics help distinguish a model that simply predicts “stay” most of the time from one that actually identifies at-risk players.

Interpretation should always reflect the use case. If a coach wants to avoid surprise exits, recall on transfer cases may matter more than raw accuracy. If an analyst wants to avoid false alarms, precision becomes more important. This is a useful lesson in decision-making under uncertainty, and it parallels how consumers weigh tradeoffs in areas like travel perks or how buyers choose between flash and durability in performance reviews.

Report uncertainty and limitations

Students should never present transfer predictions as destiny. The portal is shaped by hidden factors such as family needs, coaching relationships, academic progress, and NIL opportunities that are often not fully observable. Good analysis acknowledges these limits and discusses how omitted variables could bias estimates. A thoughtful limitations section is not a weakness; it is proof of mature research design.

Similarly, students should discuss sample bias, reporting bias, and the challenge of labels that change over time. A player may be rumored to transfer before officially entering the portal, and some departures are never fully documented. This makes the project an excellent lesson in imperfect data, a reality found in many domains, including shrinking local news ecosystems and systems outages.

Encourage reproducibility

Require students to submit code, a README, and a clear description of how data were gathered. If they scraped roster pages, they should note the date of each scrape and any manual corrections. If they used public sources for recruiting rankings or portal movements, they should document the source and timestamp. Reproducibility is a research habit that will serve them well beyond sports analytics.

For instructors, reproducibility also allows grading on process, not just outcome. A slightly less accurate model with excellent documentation is often more educational than a high-performing black box. That principle is the same one that underlies trustworthy reporting in fields as diverse as ethics and sponsorship and agent-based software workflows.

Teaching extensions, storytelling, and capstone ideas

Build a team-risk dashboard

A highly effective extension is a simple dashboard that ranks teams by roster volatility risk. Students can display predicted transfer risk by player, roster concentration by position, and historical movement patterns by season. This makes the project feel applied and helps non-technical audiences understand the results. It also mirrors real analytics environments where decision-makers want a concise risk view, not just a spreadsheet.

Compare women’s and men’s basketball

If the class has enough data, students can compare transfer behavior across women’s and men’s basketball. They may find differences in roster size, portal frequency, recruiting structure, and timing. That comparative angle is especially valuable because it teaches them to avoid universal claims based on one sport. It also turns the project into a broader study of NCAA structure rather than a narrow team case.

Use the Lady Vols as a narrative anchor, not the dataset

The Tennessee example should remain the opening narrative, because it is vivid and timely, but the full project should use broader data. That lets students connect headline-driven events to statistical patterns. It also creates a strong final presentation: “We started with a striking roster shock, then tested whether the same drivers appear across NCAA programs.” This storytelling arc is what turns a class assignment into a memorable research experience.

For students interested in broader analytical storytelling, compare the project to how market watchers identify inflection points in major sports events or how content strategists plan around volatility in shifting news cycles. In both cases, the value comes from seeing a signal early, then proving whether it matters.

Conclusion: a research project that teaches the right habits

A semester-long NCAA roster project is powerful because it combines messy real-world data with clear methodological lessons. Students learn how to predict movement, how to interpret volatility, and how to distinguish correlation from causation. They also see that transfer portal behavior is not random noise; it reflects incentives, opportunity structures, and institutional change. That makes the project a rare teaching case that is both accessible and genuinely sophisticated.

If designed well, the assignment can produce a portfolio-ready artifact and a deeper appreciation for sports analytics as research. It teaches students to organize data, select models, validate carefully, and communicate uncertainty. Most importantly, it helps them understand that roster churn is not just a headline—it is a measurable system. And once students learn to study that system well, they are better prepared to analyze any domain where movement, timing, and incentives shape outcomes.

For further inspiration on how to structure analytical thinking across changing environments, students can also explore operational checklists, churn alert design, and trend-mapping methods. Those adjacent frameworks reinforce the same lesson: when the environment is volatile, the best analysts do not merely observe change—they model it.

The best outcome is usually a player-season transfer label, but advanced students can model multiple outcomes, including decommitment, portal entry, transfer destination quality, or time-to-transfer. A multi-class outcome is more realistic because roster movement is rarely just stay versus leave. If the class is beginner-friendly, start with a binary transfer outcome and then expand.

2) How much data is enough?

There is no perfect number, but students typically need several seasons across multiple teams to build a credible model. One school alone is usually too small unless the goal is a deep case study rather than prediction. A multi-team sample from one conference or a national sample by women’s basketball or men’s basketball is often the right scale.

3) Which model should students use first?

Start with logistic regression because it is interpretable and establishes a baseline. Then compare it to a tree-based model or gradient boosting method. If timing matters, add survival analysis. The project should show progression from simple to more advanced methods rather than jumping directly to the most complex algorithm.

4) Can students really do causal inference with public sports data?

Yes, but the claims must be modest and the design carefully justified. Good options include matching, difference-in-differences, or event studies around coaching changes or portal rule changes. Students should avoid overstating causality and instead frame their findings as evidence consistent with a plausible mechanism.

5) How do we avoid turning this into a speculative rumor project?

Use only verifiable roster, recruiting, and portal sources, and document each data point. Distinguish confirmed moves from rumors, and keep a clear status taxonomy. The goal is a reproducible research project, not a gossip timeline. Strong documentation and source discipline keep the analysis credible.

6) What makes the Lady Vols example useful?

It provides a vivid, current illustration of roster volatility and helps students understand why the question matters. But it should be used as a motivating case, not as the sole dataset. The full analytical value comes from studying many teams and comparing patterns across the NCAA landscape.

Related Reading

• Navigating News Shocks - A useful framework for handling volatile, fast-changing environments.
• Real-Time Customer Alerts to Stop Churn During Leadership Change - A helpful analogy for trigger-based movement risk.
• Operate or Orchestrate? - Shows how to think about managing a shifting portfolio of assets.
• How to Mine Euromonitor and Passport - Strong inspiration for building a structured data collection workflow.
• Live Events, Slow Wins - A strategic lens for using major moments to build durable audience value.