StressFreeFantasy: Lineup Optimization with TabPFN and Real NFL Data

This project for the Hacktoberfest: Build for a Friend DEV Challenge focuses on **Best Use of TabPFN**. The author and their brother compete in a competitive fantasy football league, so they frequently root for the opponent team to struggle. However, watching their brother agonize over flex player selections with only 0.3 points separating them became tiresome. The existing platform projections had limitations, such as slow adaptation to game scripts and ignoring opponent defensive strength against specific positions.

To address these issues, the author created **StressFreeFantasy**, a tool that automates lineup decisions. It syncs with private ESPN leagues, conditions a tabular foundation model called TabPFN on historical NFL data, and outputs optimized lineups with floor/ceiling estimates. The tool is built using Streamlit and relies on three data sources: NFL historical data from **nflverse/nflreadpy**, private ESPN league data, and the user's private league roster.

TabPFN is an in-context learning model that can adapt to specific scoring rules without retraining. It provides calibrated P10 (floor) and P90 (ceiling) values, which offer more accurate guidance for starting or sitting players. The project's architecture includes feature engineering using pre-game rolling historical signals and game environment factors like opponent defensive rankings, dome status, and Vegas implied team totals.

The tool's performance was benchmarked against ESPN's default projections. The results showed a significant improvement in key metrics, such as lower MAE, higher correlation, and better pairwise Start/Sit accuracy. The author and their brother tested the tool in their own league, with the brother now using it as a sanity check before each weekly lineup submission. This change has made it harder for the author's brother to beat him in league games.