Wharton Sports Research Journal

A Comparative Analysis of Rating Systems in the US Junior Tennis Development Pathway

Koray S. Abramson

The United States Tennis Association (USTA) has historically used point-per-round rankings to determine competitive tournament entry and seeding, but this system often rewards participation over quality of play and can be distorted by random draw effects. Alternative systems such as Universal Tennis Rating (UTR) and World Tennis Number (WTN) use algorithmic predictive modeling based on prior head-to-head results to estimate player ability across gender, age, and geography.

Although previous studies (Im, 2023; Kiely, 2025; Krall, 2025; Mayew, 2023) have evaluated predictive accuracy between these two systems using smaller, elite-level samples, large-scale analyses spanning all competitive levels of U.S. junior tennis remain limited.

This study addresses that gap through a comprehensive, multi-level analysis of 70,822 USTA junior matches (scraped from January–July 2024), evaluating UTR, WTN, and USTA rankings for both accuracy and bias. Overall, UTR predicted 78.5%, WTN 74.2%, and USTA 70.1% of matches correctly, respectively, with statistically significant differences.

Quantifying the Drivers of Serve Effectiveness in Men’s Tennis

Yijia Chen, Nicole Lin, Harry Wieand, Jason Zhou

Tennis is one of the few sports in which the server initiates every point, making the serve the only stroke entirely under a player’s control. A powerful, well-placed serve can create immediate advantages, but identifying which characteristics most effectively translate serves into winning points remains unclear.

Using point-by-point data from the U.S. Open men’s singles (2018—2019; 2021—2024), we evaluate how first-serve speed, accuracy, and spatial variation (serve-location entropy) shape point outcomes. Regression results show that speed has the strongest relationship with serve efficiency, while accuracy in the service box and variation in serve location have smaller but still statistically significant associations.

Forecasting NFL Wide Receiver Touchdowns with a Temporal Linear Regression Model

Ruben Chung

Forecasting touchdowns for NFL wide receivers is a challenging but valuable problem in football analytics and player evaluation. Touchdowns are notoriously volatile, influenced by red zone usage, quarterback play, and situational variance, making year-to-year outcomes difficult to predict.

This study develops a temporal linear regression model to project wide receiver touchdown totals using a feature-rich dataset spanning 1990–2024. The dataset incorporates lagged statistics, two-year rolling averages, player age and experience, team offensive strength, and efficiency metrics such as catch rate and touchdowns per target. The model was trained on 1990–2010 player-seasons and tested on 2011–2024 data, with strict time-bounded feature engineering to prevent data leakage.

Results show strong predictive accuracy (R² = 0.803, MAE = 0.82 TDs), demonstrating that systematic patterns can be identified even within a highly volatile statistic. Feature importance analysis indicates that efficiency and usage metrics are more reliable predictors than raw prior year touchdown totals, aligning with football intuition and highlighting regression-to-the-mean effects.

The model generates 2025 projections that identify both elite scorers and likely regression candidates, providing insight into the stability of touchdown production. This work demonstrates that with careful feature engineering, a transparent and interpretable linear model can yield valuable insights in sports analytics. Beyond forecasting, the results underscore the importance of efficiency and opportunity metrics in understanding touchdown outcomes, offering a framework that can inform research on statistical predictability in professional sports.

A Run Expectancy Approach to Lead Distance Optimization in Major League Baseball

Lila Dodson, Zach Sissman, and Jack Whitney-Epstein

A runner’s primary lead off first base creates leverage to steal second but also exposure to pickoffs. We develop a nested sequence of logistic models to estimate (i) pickoff attempts, (ii) pickoff success given an attempt, (iii) steal attempts given no pickoff, and (iv) steal success given an attempt, using 2024 MLB data and Baseball Savant metrics. We map stage probabilities to expected runs via fixed linear weights (+0.20 for a successful steal; -0.45 for caught stealing or picked off) and optimize over lead distance to obtain a context-specific optimal lead L∗.

Empirically, observed leads are modestly larger than optimal on average (+0.19 ft), with a larger gap on steal attempts (+0.67), consistent with unobserved intent to steal. This framework quantifies the central trade-off – greater leads increase steal success but raise pickoff risk – on a common expected-runs scale and yield actionable, interpretable recommendations within the observed support.

PRSS: A New Metric to Quantify Pocket Collapses in the National Football League

Jahan Hafiz, Avi Mandhana, Derek Park, and Sohan Saleem

In the NFL, “quarterback pressure” refers to defensive actions that disrupt a passer’s timing, decisions, and positioning. Numerous quantitative measures have been proposed, with mixed effectiveness.

We introduce the Pocket Reduction Speed Score (PRSS), a geometric, tracking-based metric that quantifies how quickly the quarterback’s pocket shrinks. We apply the metric to player-tracking data from the 2021 NFL regular season, compute PRSS for each play, and examine its association with yards gained.

PRSS offers a continuous, interpretable measure that captures multi-defender effects and the moment of greatest pressure, complementing existing binary and closest-defender metrics while offering applications in pass-protection evaluation, scouting, and scheme design.

Beyond the Hot Hand: Skill, Experience, and Context as Determinants of Elite Badminton Performance

Vaasvi Kuthiala

This study develops a predictive model for elite badminton match outcomes to identify the key performance drivers in the sport. Using a comprehensive dataset of 3,761 men’s singles matches from the BWF World Tour (2018-2021), features have been engineered to capture player skill, via custom Elo rating system, experience, recent form and match context. The Elo was then benchmarked against logistic regression and an optimized XGBoost classifier, with evaluations tested on a held-out test set. The XGBoost model achieved superior prediction accuracy of 76.49%, statistically improving upon traditional methods.

Crucially, beyond predictive accuracy, the model’s feature importance analysis reveals a definitive hierarchy of factors influencing wins across tournaments and varying levels. Long term player skill and career experience are the primary determinants, substantially outweighing short term influences and changes in form and hot streaks, as well as exceeding contextual factors like tournament level and qualification rounds.

These findings challenge the traditional emphasis on “hot-hand” momentum, providing data-driven evidence that sustained skill and accumulated experience are more critical for victory. The results offer a practical framework for strategic decision-making by coaches, talent scouts, and sports analysts, highlighting the value of machine learning not just for prediction, but for generating actionable insights into athletic performance.

Is The NBA Draft Lottery Fair?

Charlotte Lieu, Arik Zhang, and Kevin Zheng

We evaluate the statistical fairness of the NBA Draft Lottery for the first overall pick from 1990 to 2025. Using the official league odds, we compute the likelihood of the observed sequence of winners and benchmark it with 100,000 Monte Carlo simulations. The observed joint surprisal yields a p-value of 0.0962, indicating the observed sequence is compatible with randomness under the posted probabilities.

We also examine market-size effects by aggregating team odds within large, medium, and small tiers and comparing observed No. 1 picks to their odds-based expectations via a Monte Carlo tier count test. Across tiers, differences fall within chance variation, and we find no statistically reliable over- or under-performance. Overall, first-pick outcomes from 1990–2025 are consistent with the official odds.

Using Markov Chains and Statistical Analysis to Model Intentional Fouling Situations in NCAA Division 1 Men’s Basketball

Dr. Susan Mattingly, Anirudh Sengupta

Intentionally fouling a team during their offensive possession is a strategy that basketball teams have employed for many years. Though popularized in the National Basketball Association (NBA), the technique has also made its way into Collegiate Basketball in recent years. The purpose of intentionally fouling is often to slow down opponent scoring and expose poor free-throw (FT) shooters. Despite previous studies using Markov Chains or analyzing intentional fouling in the NBA, there is a lack of research combining these methods with collegiate men’s basketball rules, particularly the one-and-one bonus.

This paper will examine the optimal times to start intentionally fouling when trailing, using Markov Chains and additional statistical analysis of data from the National Collegiate Athletic Association Division 1 Men’s Basketball (NCAA D1 MBB). Additionally, this study will account for the Bonus and Double Bonus rules specific to NCAA Men’s Basketball. Play-by-play data was compiled from over 4,000 games in the 2024-25 season to calculate statistics that would help identify the probabilities of different plays. Transition matrices were created to determine the expected points in a possession with and without fouls. By categorizing data based on the time remaining in the game, the expected points for all Division 1 teams in various situations were calculated, enabling the development of a strategy to identify the optimal time for a team to benefit from fouling.

The specific examples of the Florida Gators, Alabama Crimson Tide, and App State Mountaineers demonstrate that the optimal time to start fouling changes based on the opposing team. Additionally, the matrices were used to calculate the maximum FT% of a player for which it would be optimal to foul, rather than to give up a possession. While situations change on a game-by-game basis, the graphs created in this study will provide a general idea of when teams are expected to score fewer points in free throw situations compared to offensive possessions.

Re-evaluating the Qualifying/Finish Relationship in Formula One: A Replication and Correction of Prior Findings

Dr. Shirley Mills, Joshua Weissbock

Prior academic research in Formula One, most notably Mühlbauer (2010), concluded that starting grid positions were the strongest determinants of race outcomes while only examining the top eight finishers (fewer than 40%) of the competitors over a shortened sample of 4 seasons (2006–2009). This truncated sample approach limited the generalizability of its findings and likely affected the observed relationships.

This study attempts to correct these findings by using a much more comprehensive dataset of over 7,800 driver-weekend pairings, spanning nearly two decades of sporting activity and multiple technical eras, while including the entire racing field. We extended the original analysis, by applying contingency coefficients, as well as rank-order correlations and ordinal logistic regression, to quantify the strength of association between sessions in a race weekend (practices, qualification, starting grid, and finishing position).

This extended work indicates that qualifying performance, rather than the starting grid (often altered by post-qualification adjustments due to penalties), exhibits the strongest and most consistent association with outcomes. By re-evaluating this start-finish relationship with more complete data and transparent methods, this study corrects earlier misinterpretations and reinforces that qualifying is the most accurate indicator of the underlying driver and car performance across a weekend, season, and regulatory eras.

Boosting Draft Accuracy: A Two-Stage Classifier–Regressor Approach for NFL Wide Receiver Prospect Evaluation

Aadi Patangi

The NFL draft is an opportunity for teams to draft new players, address positional needs, and strengthen their roster for the upcoming season. Teams often trade players or compensation packages to secure certain prospects, which can be critical for team success. Wide receivers were selected as the focus of this study because the position is heavily influenced by objective data — such as receiving statistics and combine metrics like speed, agility, and explosiveness — making it well-suited for predictive modeling using machine learning.

This study presents a two-stage machine learning approach to first predict whether a wide receiver (WR) invited to the NFL Scouting Combine will be drafted, and if so, at which overall position. Using physical testing data from the NFL combine, college production statistics, and historical draft result training data from 2000 to 2024, we construct a Gradient Boosting Classifier to predict draft likelihood followed by a CatBoost Regressor to estimate draft position for those predicted to be selected.

This approach provides NFL teams and scouts with a reliable estimate of whether and when a combine-invited wide receiver will be drafted, helping them make more informed decisions and strategically position themselves to select desired prospects. In validation, our classifier reached 89.2% accuracy (F₁ = 0.936), and our regressor yielded a 49.2-pick MAE (ρ = 0.626), demonstrating robust predictive performance.

Modeling Realistic Placement Probabilities in Speedcubing Using Kernel Density Estimation (KDE)

Ryan Saito

This paper offers a simulation-based framework for predicting placements in official World Cube Association (WCA) competitions. Currently, the WCA uses psych-sheet rankings, which only present a competitor’s best average of five (Ao5) and singular solve. Our goal is to create a simulation framework that uses recent performance to estimate realistic placement probabilities. We construct kernel density estimates (Gaussian KDEs with an adaptive bandwidth based on the coefficient of variation) for each competitor using their most recent twenty-five official solves, and optionally csTimer practice solves. Then, we use percentile sampling and bootstrap Ao5s to make synthetic solves across 100-1000 tournament iterations.

We also created an open-source app that shows each competitor’s probability of advancing, their expected rank, and a KDE distribution with 95% confidence and prediction intervals. We tested this tool at the Saint John’s Warm Up 2025, where it correctly predicted the podium of the competition with 80% accuracy (12 out of 15 places) and had 33% of the predictions exactly correct (5 out of 15 places). At the Rubik’s World Championships in 2025, the tool also correctly predicted the entire podium for the 3×3 event. Practically, the tool lets competitors and organizers set expectations, evaluate consistencies, and make decisions about training, seeding, and round cutoffs based on the data. It is also accessible to spectators and enthusiasts.

Using Injury-Risk Forecasting to Quantify Financial Impact in the NBA

Ethan Wang

Injuries in the NBA have become consequential not only for team success but for the financial costs those teams suffer. This study develops a machine learning framework that predicts next-game injury risk using publicly available box-score data, player attributes, and injury history, then translates these probabilities into expected financial costs.

Combining five datasets from 2010-2022, I derived sixteen workload and recency features and trained a Random Forest model optimized with five-fold cross-validation. At a 2% threshold for classification, the model predicts out-of-sample 69% of injuries while correctly ruling out 62% of healthy games, indicating better-than-chance predictive power is possible using solely public data. Feature-importance analysis identified workload shifts and rest as primary predictors.

Extending beyond prediction, this study gives a new way to interpret the financial implications of injuries, looking at how strategic rest decisions can minimize financial loss. This study offers NBA organizations a data-driven tool linking injury prevention with financial optimization, bridging injury forecasting with economic decision-making.