Fig. 4 — enlarged from sheet 01
Fantasy football draft analysis built on a graph: players, coaches, and games as nodes, stints as edges with validity spans, backtested against held-out seasons.
depth-chart predicts weekly fantasy performance from the structure around a player rather than his stats alone. Players, teams, coaches, and games form a graph. Every relationship (a player’s stint on a team, a coordinator’s tenure) has a validity span, so features can be computed as of any week without leaking the future into a backtest.
The bet: change and stability predict better than talent estimates, because talent is already priced into draft position but structural change is priced badly. A receiver whose coordinator and quarterback both left is a different projection than his points per game suggest.
The first backtests train on earlier seasons and score rankings within each position against a season the model never saw. Team changes, quarterback continuity, and years of experience each moved the needle, and what the model learned lines up with football sense: changing teams costs about a tenth of production, young players improve, veterans fade, and a young player who gets moved early usually had reasons.
Running backs resist every continuity feature so far, which points at role rather than stability as their story. Depth charts, arriving draft capital, and simulated game scripts are the next layers, ending in season simulation and a weekly tool for deciding who to start.