Roster Advice
Model → real add/drop/read decisions for a posted roster (composes with Sleeper sync).
Intelligence engines
Does the grade actually predict?
Anyone can publish a rating. We backtest ours. Build the process grade on the first half of the season, then measure how well it ranks second-half production, against the obvious baseline: past production predicting future production. If the grade adds lift, it carries forward signal.
How we read it
Grade ρ: does the grade rank the future?
Spearman rank correlation between the first-half process grade and second-half production, within position. Higher is better.
Lift: does it beat the past?
Grade ρ minus baseline ρ (past production → future production). Positive means the grade adds signal the box score didn't already have.
Call hit-rate: were buy/sell right?
Of first-half buy-lows, how many rose; of sell-highs, how many fell. Read against the 50% coin flip.
Grade ρ (overall)
0.63
vs 0.61 past-production baseline
Lift over the past
+0.02
211 players · 2024
Sell-high hit-rate
63%
n=32 calls · LCB 46% · buy-low 44% (LCB 27%) · vs 50% coin flip
2024 · trained on weeks 1-9 · tested on weeks 10-18
The backtest, by position
First-half process grade ranks second-half production at ρ≈0.63 and adds 0.02 of rank-correlation lift over past production alone. In-sample, single-season: a directional proof, not a guarantee.
| Overall | 211 | 0.63 | 0.61 | +0.02 | 44%LCB 27% · n=27 | 63%LCB 46% · n=32 |
| QB | 32 | 0.57 | 0.58 | -0.01 | n=6 — too few to rate | n=4 — too few to rate |
| RB | 62 | 0.73 | 0.82 | -0.09 | n=7 — too few to rate | n=8 — too few to rate |
| WR | 82 | 0.72 | 0.63 | +0.09 | n=10 — too few to rate | n=13 — too few to rate |
| TE | 35 | 0.33 | 0.24 | +0.09 | n=4 — too few to rate | n=7 — too few to rate |
Split-half backtest on real nflverse weekly data: build the process grade on the first half of the season, then measure how it ranks second-half production vs the past-production baseline. Buy-low/sell-high calls are scored against the coin-flip line. In-sample, single-season: directional proof.
Out-of-sample · trained on 2023 · tested on 2024
The draft test: does last year's grade predict this year?
Grade built on 2023 ranks 2024 production at ρ≈0.63 (it does not beat raw 2023 production this pair). Out-of-sample, across seasons: the real draft test.
| Overall | 251 | 0.63 | 0.68 | -0.05 | n=19 — too few to rate | 52%LCB 34% · n=27 |
| QB | 29 | 0.54 | 0.67 | -0.13 | n=1 — too few to rate | n=6 — too few to rate |
| RB | 66 | 0.61 | 0.70 | -0.09 | n=8 — too few to rate | n=8 — too few to rate |
| WR | 103 | 0.69 | 0.68 | +0.01 | n=5 — too few to rate | n=8 — too few to rate |
| TE | 53 | 0.57 | 0.71 | -0.14 | n=5 — too few to rate | n=5 — too few to rate |
Year-over-year is the harder, more honest test: players change teams, age, and get hurt. Where the grade beats raw prior-season production (positive lift), it carries signal the box score didn't.
Multi-year stacked · out-of-sample · pooled pairs 2021→2022 · 2022→2023 · 2023→2024
The multi-year test: pool several seasons for real statistical power.
Pooling 737 player-seasons across 3 consecutive train→test pairs (2021→2022, 2022→2023, 2023→2024), the prior-season grade ranks next-season production at ρ≈0.60, and it does NOT beat raw prior-season production over the pool. Multi-year, out-of-sample: the strongest, highest-power form of the draft test.
| Overall | 737 | 0.60 | 0.64 | -0.04 | 49%LCB 38% · n=73 | 66%LCB 57% · n=110 |
| QB | 90 | 0.35 | 0.55 | -0.20 | n=5 — too few to rate | 58%LCB 39% · n=26 |
| RB | 186 | 0.57 | 0.60 | -0.03 | n=23 — too few to rate | 76%LCB 59% · n=33 |
| WR | 306 | 0.69 | 0.69 | 0.00 | 44%LCB 27% · n=27 | 69%LCB 52% · n=35 |
| TE | 155 | 0.61 | 0.65 | -0.04 | n=18 — too few to rate | n=16 — too few to rate |
Each train→test pair is normalized within its own seasons (percentiles per pair), then the pairs are pooled, never re-ranked across seasons. Out-of-sample throughout; where it beats the baseline (positive lift), that is the strongest evidence the grade carries forward signal.
Data via nflverse (nflverse-data), licensed CC BY 4.0. How we source data
More engines & APIs5 engines without a standalone boardPOST-only, founder-gated, or rendered on the player boards under /players. Expand to browse them.
Graded Pool
Composes the model + xFP + team environment (real schemeFit from neutral-script offensive EPA) + QB-forward passing signal into a real graded pool that drives every fantasy tool when the founder enables it.
QB Consensus
ESPN QBR (results) vs Next Gen CPOE (accuracy), triangulated. Disagreement surfaced, not averaged.
Rushing Efficiency
RYOE/att vs volume with stacked-box context: bell-cow / buy-low / volume-dependent.
Receiving Opportunity (WOPR)
Air-yards & target share → WOPR, with opportunity-vs-production buy/sell.