Model vs. Consensus
Our projections benchmarked head-to-head against the expert consensus, graded weekly in-season
We beat the expert consensus
Measured head-to-head vs Sleeper expert consensus projections, 2022-2024 · 7,009 matched player-weeks. Overall MAE 4.749 vs their 5.285 (gap -0.54).
How we grade this| Position | Our MAE | Consensus MAE | Gap | Grade | Verdict |
|---|---|---|---|---|---|
| QB | 4.742 | 6.401 | -1.66 | A+ | Beat Consensus |
| RB | 4.866 | 5.332 | -0.47 | A+ | Beat Consensus |
| WR | 4.910 | 5.042 | -0.13 | A | Beat Consensus |
| TE | 4.001 | 4.455 | -0.45 | A+ | Beat Consensus |
| Overall | 4.749 | 5.285 | -0.54 | A+ | Beat Consensus |
- Matched pairs only. We score a player-week only when both our model and the expert consensus published a projection — same players, same slate.
- Identical actuals. Both sources are graded against the same realised Half-PPR fantasy points.
- Injury-aware. Inactive and ruled-out players are handled consistently on both sides before scoring.
Benchmark
Sleeper expert consensus projections
Seasons
2022-2024 · Wk 3-18
Scoring
Half-PPR
Player-Weeks
7,009
Source artifact
consensus_matched_sleeper_half_ppr_20260823_023112.csv
Every Tuesday during the season we grade the previous week's projections against the expert consensus, and our game lines against the closing Vegas market — then publish the scorecard here. No hindsight, no edits: the numbers land the day after the games.
The first 2026 report lands after Week 1.
Seasons
2022-2024
Weeks
2-18
Scoring
Half-PPR
Player-Weeks
11,358
| Position | Model | MAE | RMSE | Bias |
|---|---|---|---|---|
| QB | Hybrid Residual | 4.84 | 6.28 | -0.52 |
| RB | Hybrid Residual | 4.13 | 5.67 | -0.19 |
| WR | Hybrid Residual | 4.26 | 5.96 | -1.01 |
| TE | Hybrid Residual | 3.16 | 4.41 | -0.23 |
Projections are generated from a heuristic base model: weighted rolling averages combined with usage, matchup, and Vegas implied-total multipliers, with TD regression and per-position recency weighting (v4.2).
QB, RB, WR, and TE projections all add an ML residual correction layer (hybrid heuristic + ML) trained on 2016-2025 data on top of that heuristic base. All models are evaluated on held-out seasons, and ship decisions are confirmed on a sealed 2025 holdout.
Lower MAE is better. A bias near zero indicates well-calibrated projections. Negative bias means the model slightly under-projects on average.