How the Model Works

A full technical breakdown of OverADP's machine-learning pipeline — how we project fantasy points, how we quantify uncertainty, and how we validate without leaking future information into the past.

The one-sentence version

We train one CatBoost point model per position on completed NFL seasons, test it chronologically on the next season, and pair the point estimate with a separate split-conformal quantile range targeting 80% marginal coverage.

1. Walk-forward validation (not random splits)

Fantasy football data is temporal: player stats in 2023 are not independent of stats in 2022. A standard 80/20 random train-test split would leak future information — the model would see a player's 2024 season during training and then be "tested" on his 2025 season, but it already knows the player's career trajectory.

Walk-forward validation prevents this. The published point-model results use two folds: train through 2023 and test on 2024, then train through 2024 and test on 2025. The aggregate includes every eligible QB/RB/WR/TE row in those folds.

2. One CatBoost model per position

QB, RB, WR, and TE are trained separately because their production scales and useful features differ. The production point model is CatBoost for all four positions, with position-specific feature lists and temporal sample weighting so newer training seasons matter more.

3. Conformal quantile regression (CQR) for honest 80% intervals

Point predictions alone are dangerous in fantasy — every projection is wrong, the question is by how much. We train separate quantile CatBoost models at the 10th and 90th percentiles, then calibrate on the most recent held-out season using split-conformal CQR.

The final adjustment is learned from 2025 calibration rows and targets 80% marginal coverage. Because that same season is used for final calibration, its post-adjustment coverage is a calibration diagnostic, not an independent test-set guarantee. Coverage for an individual player is never guaranteed.

4. Depth-chart awareness (Week 1 snapshot)

One of our biggest 2026 feature additions. We pull pre-season depth charts from nflverse (Week 1 snapshot for 2019-2024; nearest-to-September-5 snapshot for 2025+) and encode each player's depth rank (1=starter, 2=backup, 3+=depth), plus binary is_starter and is_backup flags.

Used for QB, WR, and TE projections. Excluded from RB because RBBC (running-back-by-committee) makes a nominal depth rank less reliable than actual prior workload and teammate carry competition.

5. Target-competition features (prevent phantom breakouts)

Using prior-season teammate targets, we compute each WR's teammate_targets_prev and teammate_rec_yards_prev on their current team (so if Chase Claypool signs with the Jaguars, his projections reflect BTJ's 200+ targets ahead of him). We also compute teammate_carries_prev for RBs.

The feature uses prior-season teammate production on the current roster, never the target season's outcomes. This re-runs after free-agent and draft changes are reflected in the projection-season roster.

6. Conservative monotonic constraints

Aggregate production lags (prior-season fantasy points, targets, receptions, carries) should never have a negative marginal effect on projections. We encode positive monotonic constraints on exactly these features and leave everything else unconstrained. This adds sanity guardrails without overfitting — MAE stays within noise, but the model can't produce pathological projections where scoring more the prior year makes you project lower.

7. College + draft capital features for rookies

For rookies and second-year players, we merge draft picks, combine metrics, college production, and interaction features (college_x_rookie, draft_cap_x_rookie, athletic_x_rookie). Athletic score is a position-weighted composite of combine z-scores. These features give the model signal before an NFL stat line exists.

The 2024, 2025 walk-forward results

All numbers below are aggregated from the exported validation results and weighted by the number of player-season predictions in each fold. The exact-cohort market check uses true FFC ADP in 2024 but an explicitly labeled ESPN preseason-rank proxy in 2025, so no ADP improvement percentage is published here.

Position	MAE	RMSE	R²	Held-Out N
QB	57.15	72.43	0.61	134
RB	43.39	60.47	0.51	269
WR	31.90	43.78	0.57	426
TE	23.79	33.35	0.56	223
Player-weighted summary	36.34	51.14	0.56	1,052

What the model doesn't do

Honest limitations:

It can't predict injuries. Malik Nabers finishing 2025 with 4 games played wasn't a model call — it was a bone bruise. We DO model injury rates from prior-season games-missed features, but week-to-week injuries are noise.

It's only as good as the data. UDFAs and late-round rookies with missing college data get wider intervals and lower confidence. Coaching-change features were tested and rejected after walk-forward validation showed them adding noise rather than signal.

Fantasy football is high-variance. Even a strong aggregate model misses individual players. The interval pipeline targets 80% marginal coverage after calibration, but that target is not a player-level promise and still needs monitoring on future untouched seasons.

What's next

Roster, depth-chart, ADP, and rookie inputs continue to change through training camp. We refresh the board as those sources stabilize and will report interval coverage again only after a future season remains untouched through evaluation.

See the current results in the free War Room, or dive into the top sleepers and top busts.

For the decision-system evidence, read the focused ADP vs model simulation study. For a fast two-player check, use the Who Should I Draft comparator.

Related Rankings