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 gradient-boosted model on completed NFL seasons (2014 onward), test it season by season on years it never saw, then blend each projection 50/50 with the points implied by current ADP, because that combination tested better than either the model or ADP alone.

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 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. Every test season is predicted by a model trained only on earlier seasons. We tested seven seasons this way (2019 to 2025) and chose the design using 2019 to 2024 before looking at 2025.

2. One model, three scoring formats

The point model is a bag of gradient-boosted trees averaged with a two-stage version that predicts games played and points per game separately, so a season shortened by injury is not automatically read as lost talent. One model is trained across QB, RB, WR and TE with position as an input, which tested better than separate position models, especially at TE. Standard, half-PPR and PPR are each trained directly.

3. What goes into it

Production and usage: three seasons of stats, per-game rates, target share, air-yards share, snap share and route participation (routes run, targets and yards per route).

Opportunity quality: expected fantasy points from play-by-play, red-zone and goal-line work, deep and end-zone targets, and touchdowns over expected to catch regression.

Role: the preseason depth chart, last season's closing depth chart and each player's order in his position room. The depth chart is the strongest single signal after past production.

Rookies: draft capital, combine results, age and college production (share of team receiving yards and touchdowns, breakout age, yards per team pass attempt, recruiting rating).

Team context: the new team's pace, pass rate and scoring, vacated targets and carries, returning competition at the position and the quality of the team's quarterback.

4. Blending with ADP

ADP carries information public stats do not, such as camp news and injury recovery timelines. We convert each player's ADP into implied points (a log-ADP curve fitted by position on past seasons) and average it with the model. On 2019 to 2025, across three scoring formats, the blend had lower error than ADP alone in 19 of 21 season-format tests and ranked players better in 14 of 21. Players without real ADP use the model alone.

5. 80% ranges

Every projection is wrong; the question is by how much. Each range comes from how far past projections at the same position and projection level missed, measured only on earlier seasons. In 2024 and 2025 about 77% of outcomes landed inside the 80% range (lower at TE and QB), so treat the ranges as a guide, not a guarantee.

6. What we tested and left out

Injury-report features, coaching changes, week-1 betting lines, a model that predicts yards and touchdowns separately, and a learned (rather than 50/50) ADP combination were all tested season by season. None beat the current design by more than noise, so none were added.

The 2024, 2025 walk-forward results

These are the live board's numbers (model blended with ADP, half-PPR) for players with a prior NFL season or a top-150 draft pick, each season predicted using only earlier seasons. Because the player pool differs from the previous model's published table, the error values are not directly comparable to older versions of this page. On the same true-ADP players in 2023 to 2025, the previous model had a rank correlation of 0.29 and the new blend 0.43.

Position	MAE	RMSE	R²	Held-Out N
QB	42.93	62.45	0.71	184
RB	31.40	47.85	0.68	328
WR	25.21	38.32	0.67	515
TE	17.99	28.14	0.67	276
Player-weighted summary	27.74	43.30	0.68	1,303

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. The model does learn from games played in past seasons, but week-to-week injuries are noise.

It's only as good as the data. Undrafted and late-round rookies with thin college data carry more uncertainty. Players not in the model's roster snapshot keep the previous model's projection.

Fantasy football is high-variance. Even a strong aggregate model misses individual players. Stars coming back from injury are the biggest known blind spot for public stats, which is one reason the board leans on ADP too.

What's next

For 2027 we plan to rebuild the board from draft-day rosters and depth charts through training camp and to re-test the full draft simulator before August. This page will be updated with the results.

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

Draft simulations

We also test full drafts against ADP-first drafting. All results, including a re-run that did not reproduce an earlier study, are on the ADP vs Model page.

For a fast two-player check, use the Who Should I Draft comparator.

Related Rankings