Inspiration
Late-game sports arguments are usually settled by memory, highlight reels, and recency bias. A player hits two huge shots in a week and suddenly becomes “clutch”; another player misses a late free throw and the label sticks. But traditional box scores and season-long averages are not designed to isolate performance in high-leverage moments. We wanted to build something that treats clutch time as its own analytical problem: smaller sample sizes, higher variance, different defensive schemes, and much higher stakes.
The NBA’s official clutch window — roughly the last five minutes of a game with the score within five points — gave us a natural starting point. We asked: can we build a tool that not only ranks players and teams in clutch situations, but also shows why those rankings look the way they do, and how uncertain they really are?
That question became ClutchMetrics.
What it does
ClutchMetrics is an interactive web dashboard and predictive modeling pipeline for NBA clutch performance. It lets users explore:
Player and team clutch leaderboards using win-probability-based impact rather than raw points alone.
Clutch Performance Index (CPI) scores that account for scoring efficiency, turnovers, assists, rebounds, and leverage-weighted win probability added.
Win probability curves for close games, showing how individual plays and lineup changes shifted the estimated probability of winning.
Shot charts and play-type breakdowns for clutch possessions, so users can see whether a player’s late-game value comes from rim pressure, pull-up shooting, free throws, or playmaking.
Uncertainty-aware rankings that shrink small samples toward the league mean, because a player with 18 clutch possessions should not be treated the same as a player with 180.
Team-level insights for late-game lineup construction, foul strategy, timeout timing, and shot selection.
The actionable impact is straightforward: coaches, analysts, fantasy players, and fans can move beyond “who hit the big shot?” and ask “which players and lineups actually improve win probability when the game is on the line, and how confident should we be?”
How we built it
We built ClutchMetrics as an end-to-end open-source Python project.
Data pipeline. We ingested public NBA play-by-play and box-score data, cleaned possession-level events, and reconstructed game state: score margin, time remaining, possession team, lineup, and event type. We defined a clutch possession as one occurring in the final five minutes with the score within five points. Feature engineering. For each possession we created features such as score differential, time remaining, possession indicator, home/away status, player usage, lineup composition, shot location, and prior clutch experience. We also computed leverage as the absolute change in win probability a possession could produce. Dashboard and deployment. We used pandas and NumPy for data processing, scikit-learn for modeling, Plotly for interactive charts, and Streamlit for the web dashboard. The app includes filters for season, team, player, and clutch definition, plus explainability panels showing which features drove a player’s or team’s clutch score. The full source code, data-processing scripts, and model notebooks are public on GitHub.
Challenges we ran into
The hardest part was not building a dashboard — it was making the analytics defensible.
Defining “clutch” is subjective. The five-point/five-minute window is useful, but it excludes overtime, blowouts that become close, and high-leverage moments earlier in the game. We made the window configurable and documented the tradeoff.
Small sample sizes. Many players have only a few dozen clutch possessions in a season. Raw rankings were noisy and unstable. Shrinkage and uncertainty reporting became essential.
Selection bias. Clutch minutes are not randomly assigned. Stars play more late-game possessions, good teams win more close games, and lineups are highly correlated. We had to be careful not to interpret correlation as causation.
Data cleaning. Play-by-play data contains missing events, lineup ambiguities, and occasional score discrepancies. Reconstructing possession state required validation checks and manual spot-checking.
Temporal leakage. We had to ensure the win probability model was trained only on information available at the time of the possession, not future outcomes.
Calibration vs. ranking. A model can rank players well but produce poorly calibrated probabilities. Since the dashboard shows win probability curves, calibration mattered as much as discrimination.
Performance. Recomputing win probability and CPI across every play, player, and lineup combination was slow at first. We cached intermediate tables and precomputed aggregations for the dashboard.
Accomplishments that we're proud of
We built a complete, reproducible pipeline from raw public data to an interactive analytics product.
We created a clutch metric that is leverage-weighted, sample-size-aware, and explainable rather than a black-box ranking.
We prioritized model calibration and uncertainty, which makes the dashboard more honest than a simple “clutch leaderboard.”
We shipped a working prototype with player pages, team pages, win probability charts, and shot charts.
We made the project open source so others can extend it to different sports, seasons, or definitions of clutch time.
What we learned
We learned that clutch analytics is less about finding “the clutchest player” and more about modeling context and uncertainty. Late-game performance is noisy, and the most useful analytics product is often the one that communicates how much we should trust the number.
On the technical side, we learned how much work goes into possession reconstruction, feature engineering, and model calibration. On the product side, we learned that explainability matters: users want to know why a player ranks highly, not just where they rank. We also learned to design around small samples from the start rather than treating shrinkage as an afterthought.
What's next for ClutchMetrics
Next, we want to:
Expand beyond the NBA to the WNBA, MLB, NFL, and soccer, using sport-specific leverage definitions.
Add Bayesian hierarchical models to better estimate player and lineup effects with partial pooling.
Incorporate player-tracking and injury data to separate skill from lineup context and availability.
Add live or near-live data ingestion for in-game win probability and clutch alerts.
Build a lineup optimizer for late-game situations and a fantasy-focused clutch projection mode.
Improve the explainability layer with counterfactual “what if this lineup or shot selection changed?” simulations.
Publish a reproducible benchmark and documentation so other students and analysts can build on ClutchMetrics.
Log in or sign up for Devpost to join the conversation.