Inspiration
The decision behind every fourth down A fourth-down decision can change a game, yet coaches often have only seconds to choose between going for it, attempting a field goal, or punting. I built AQX to make the tradeoffs easier to see: how does each option affect the team’s chance of winning?
How I built it
AQX uses historical NFL play-by-play data to estimate the outcomes of each choice. It combines models for fourth-down conversions, field goals, punts, and win probability to compare the available actions.
The models are trained in Python, then exported into a self-contained web app. The app runs the decision tool in the browser, where users can adjust a scenario and see how the recommendation changes. What I learned I learned that a useful sports model needs more than accurate predictions: it also needs to explain how its estimates lead to a recommendation. I also learned how important it is to check that a model behaves sensibly across different game situations and that the browser version agrees with the original Python implementation.
Challenges
One challenge was comparing different outcomes fairly. A conversion attempt, a field goal, and a punt each change field position and scoring chances in different ways. I addressed this by estimating win probability after each possible outcome and combining those estimates according to their predicted probabilities.
Another challenge was bringing the models into a browser-based app while preserving their behavior. I exported the trained models and checked the JavaScript predictions against Python on 500 test situations.
What it does
AQX helps users compare the three main options on fourth down: go for it, attempt a field goal, or punt. It estimates each option’s expected win probability and recommends the choice with the strongest odds. Users can adjust the scenario and conversion estimate to explore how the decision changes.
Accomplishments that we're proud of
We built an interactive fourth-down decision tool that runs entirely in the browser. It compares all three choices, shows the expected win probability, and lets users test different assumptions. We also used the models to audit fourth-down decisions from the 2025 season.
What we learned
We learned that a useful sports analytics tool needs to make its reasoning understandable, not just produce a prediction. We also learned how to export trained models for browser use and verify that their predictions remain consistent across Python and JavaScript.
What's next for AQX: Fourth-Down Compass
We’d like to incorporate team-specific conversion rates, player availability, and matchup data to make estimates more relevant to each game. We also want to improve the app’s explanations and expand the historical audit as new NFL seasons become available
