How QuantRoster Works

Ensemble Model Architecture

We train separate models for each position (QB, RB, WR, TE, K, DEF) because the features that predict a quarterback's fantasy output are fundamentally different from those that predict a tight end's. Each position gets its own dedicated ensemble of multiple model types that attack the prediction problem from different angles.

The ensemble approach reduces overfitting and improves reliability. Some components of the ensemble excel at capturing nonlinear relationships (like how a player's value shifts against different defensive schemes), while others provide stability. The final projection blends them in a way that consistently outperforms any single model.

What the Model Sees

The model trains on multiple seasons of NFL game-level data. For each player-game, we extract features including passing/rushing/receiving stats, snap count percentage, red zone usage, target share trends, and opponent defensive metrics (DVOA, points allowed by position, sack rate).

Matchup adjustment is a big part of this. A running back facing the league's worst run defense is a fundamentally different proposition than the same player facing the best. The model captures these interactions through opponent-specific features rather than simple strength-of-schedule adjustments.

Confidence Intervals, Not Just Point Estimates

A point projection alone isn't enough information to make a good start/sit decision. If two players both project at 12 points, but one has a range of 8-16 and the other has a range of 4-22, those are very different decisions. We train separate models specifically to estimate the 10th and 90th percentiles, giving you a floor and ceiling for every projection.

League-Specific Projections

When you connect your league, we pull your exact scoring settings. The model recomputes fantasy points using your league's rules — PPR, half-PPR, standard, custom scoring, whatever your platform uses. This means our projections reflect what a player is actually worth in your league, not some default scoring format.

User-Controlled Model Inputs

Every player's projection card has adjustable sliders for the inputs most relevant to their position: pass attempts for QBs, rush attempts and receiving targets for RBs/WRs/TEs, kick attempts for kickers, and sacks/turnovers/points allowed for defenses.

Think a running back returning from injury will see a reduced workload? Dial down his rush attempts and watch the projection adjust in real time. Under the hood, we pre-compute projection curves at multiple input levels during the weekly pipeline, so your adjustments feel instant with no live model calls required.

Real-Time Game-Day Pipeline

Our morning pipeline generates baseline projections from overnight data, but fantasy decisions depend on information that changes right up until kickoff. A second layer polls official NFL active/inactive lists, weather APIs, and Vegas line movements every 60 seconds on game day, applying rule-based adjustments to existing projections.

The third layer is the differentiator: an AI-powered news interpreter that reads curated fantasy news feeds throughout the day, classifies each update by affected player, assesses materiality, and decides whether to alert you or auto-adjust a projection. Beat reporter says the coach plans to ease a player back in? We catch it before the official lists do.

Cryptographic Verification

Every projection you receive comes with a unique verification hash that locks in the prediction before game time. After the game, you can use that hash to confirm that your projection wasn't changed after the fact. No other fantasy platform lets you independently verify their predictions this way.