Model ascension-1.0.0 · tested 2026-10-06
How NFL Ascension works, and how well it did
Ascension turns a player’s usage trend into one 0–100 score and a breakout probability. Everything below was fitted on past seasons only and tested on the season after (2023 → 2024, 2023–24 → 2025), so none of the numbers peek at the games they predict.
The score
For each player we compare his last two games with his baseline: his earlier games this season, shrunk toward last season (or toward a fringe role for rookies). The channels are snap share, route share, target share, carry share, red-zone share, goal-line carry share, air-yards share and opportunity points (the PPR points his volume is usually worth).
Each change is first cut down to the part that tends to stick, using the persistence we measured for the usage risers: a starter going down and a backup taking over keeps most of the gain; a target spike at steady snaps keeps about a quarter. The kept changes are then weighted by how much they moved points over the next three games, fitted per position with weights that cannot go negative. Current weights for wide receivers: opportunity points 0.55, snap share 1.4, route share 1.36, target share 0.51 (points per game per unit of share).
The weighted sum is the expected change in PPR points per game. The score is 50 + 50 × tanh(change ÷ scale), with a per-position scale set so about one player in twelve with a role is Surging (75+) or Falling (25 and under). Rising is 60–75, Cooling 25–40.
Breakout probability
A breakout is at least one top-24 week (RB, WR) or top-12 week (QB, TE) in PPR over the team’s next three games. The probability is a logistic model per position on projected opportunity, baseline points, points over the last two games, the Ascension change, projected snap share and whether he missed the last game.
Pre-breakout stashes
Score 60+ (or a 35%+ breakout chance with a score of 55+), points over the last two games at least 2 a game below what his usage usually scores, and rostered in under half of ESPN leagues.
Backtest
15,493 player-weeks with a role, as of weeks 3–15, 2023–25; 10,292 held out (2024 and 2025). Lower log-loss and Brier are better; AUC is how well the model ranks breakouts above non-breakouts (0.5 = coin flip).
| Model | Log-loss | Brier | AUC |
|---|---|---|---|
| Ascension breakout model | 0.438 | 0.141 | 0.864 |
| Season points per game | 0.474 | 0.154 | 0.837 |
| Last week’s points | 0.517 | 0.170 | 0.805 |
| Usage-riser score | 0.622 | 0.215 | 0.644 |
| Base rate only | 0.650 | 0.229 | 0.500 |
| Test season | Fit on | Player-weeks | Base rate | Log-loss | Last-week naive | Brier |
|---|---|---|---|---|---|---|
| 2024 | 2023 | 5,144 | 35% | 0.435 | 0.513 | 0.139 |
| 2025 | 2023–24 | 5,148 | 36% | 0.442 | 0.522 | 0.143 |
| 2025 | Player-weeks | Base rate | Log-loss | Last-week naive | AUC |
|---|---|---|---|---|---|
| QB | 563 | 54% | 0.531 | 0.593 | 0.809 |
| RB | 1,299 | 44% | 0.455 | 0.528 | 0.870 |
| WR | 2,069 | 32% | 0.437 | 0.524 | 0.851 |
| TE | 1,217 | 26% | 0.396 | 0.478 | 0.875 |
Calibration
Close to the diagonal across the range. It runs slightly cautious at 20–40% (observed a few points higher) and slightly confident above 80% (predicted 85%, observed 79%).
| Bin | Player-weeks | Predicted | Observed |
|---|---|---|---|
| 0–10% | 3,090 | 6.0% | 5.0% |
| 10–20% | 1,773 | 14.1% | 14.7% |
| 20–30% | 917 | 24.6% | 29.6% |
| 30–40% | 630 | 34.7% | 39.2% |
| 40–50% | 569 | 45.0% | 45.7% |
| 50–60% | 538 | 55.2% | 58.6% |
| 60–70% | 706 | 65.1% | 66.3% |
| 70–80% | 759 | 74.9% | 74.2% |
| 80–90% | 702 | 85.0% | 78.8% |
| 90–100% | 608 | 94.5% | 91.4% |
Picking from players who were not already starters
Each week, the top 20 by each method among players whose baseline was below the top-24 (top-12) line, a stand-in for “on waivers” (there is no historical rostered %). Hit rate = share with a breakout in the next 3; change = next-3 points per game minus baseline.
| Top 20 by | Picks | Hit rate | Change (PPR/g) |
|---|---|---|---|
| Breakout probability | 520 | 76.0% | +2.39 |
| Last week’s points | 520 | 66.7% | +3.02 |
| Usage-riser score | 520 | 52.7% | +4.28 |
| Ascension score | 520 | 48.8% | +4.03 |
| Whole pool (base rate) | 8,420 | 25.9% | +0.48 |
Honest read: breakout probability finds the most breakouts (it knows who is already good); the Ascension score finds the biggest improvements, about as well as the usage-riser score it is built on (correlation with the next-3 change 0.34 vs 0.34). Use the probability to rank adds and the score to spot role changes early.
| Tier | Player-weeks | Breakout rate | Change (PPR/g) | Beat baseline |
|---|---|---|---|---|
| Surging | 435 | 48.7% | +4.32 | 78% |
| Rising | 1,395 | 39.5% | +2.06 | 66% |
| Steady | 4,835 | 24.1% | +0.38 | 45% |
| Cooling | 1,526 | 13.6% | -1.56 | 22% |
| Falling | 229 | 20.1% | -3.15 | 20% |
Pre-breakout flag (same pool, without the rostered filter): 689 flagged players broke out 43% of the time and scored 0.8 more points a game than in their last two; risers with the same signal whose points had already arrived broke out 44% of the time but scored 3.9 fewer. The flag does not raise the breakout rate; it finds the risers whose price has not moved yet.
Route share (estimate)
nflverse participation data (who was on the field each play) stops after 2025. For 2026 we estimate the share of team dropbacks a player is on the field for from his snap share: a position curve (tight ends and backs block on some snaps) times his own 2025 rate, shrunk toward the curve. Tested on 2025 with 2024 rates (as 2026 uses 2025). Error is in share points (0.058 = 5.8 points).
| Position | Player-games | Snap % as-is | Position curve | Curve + player | Correlation | Season to date |
|---|---|---|---|---|---|---|
| All | 5,484 | 0.072 | 0.064 | 0.058 | 0.967 | 0.038 |
| RB | 1,495 | 0.067 | 0.066 | 0.060 | 0.948 | 0.041 |
| WR | 2,445 | 0.067 | 0.056 | 0.054 | 0.973 | 0.034 |
| TE | 1,544 | 0.085 | 0.075 | 0.062 | 0.960 | 0.043 |
Targets and yards per route run on player pages divide by this estimate, so treat them as estimates too. Using the true 2024–25 route share instead of the estimate did not improve the breakout model, so the 2026 scores lose nothing by it.
How this works
Re-run: npx tsx scripts/nfl/ascension.ts --backtest. Full write-up in the repository (docs/nfl/model/ASCENSION.md). Back to the Ascension leaderboard.
Sources: nflverse, Next Gen Stats via nflverse, ESPN rostered %
Frequently asked questions
What is the NFL Ascension score?
A 0–100 number for how much a player’s role is growing in ways that tend to stick. It combines his recent change in snap share, route share, target share, carry share, red-zone and goal-line share and air-yards share, each weighted by how much of that kind of change carried into later games in 2023–25 (a promotion after an injury sticks more than a target spike at steady snaps). 50 means no change; 75+ is Surging, 25 and under is Falling.
What does breakout probability mean?
The chance of at least one top-24 week (RB, WR) or top-12 week (QB, TE) in PPR scoring over his team’s next three games. It is a calibrated probability: of players we gave 30–40%, about 39% had one in our 2024–25 test.
What is a pre-breakout stash?
A player whose usage is rising (score 60+, or a 35%+ breakout chance with a score of 55+), whose points over his last two games trail what that usage usually scores by 2+ points a game, and who is rostered in under half of ESPN leagues. In testing these players broke out about as often as other risers, but their points went up afterwards (+0.8 a game) while risers who had already scored fell back (−3.8): you get them before the points arrive.
Are route shares real routes?
No. Route share is an estimate of the share of his team’s dropbacks he was on the field for. Through 2025 it comes from nflverse participation data; from 2026 we estimate it from snap share with a position-specific adjustment and his own 2025 rate (typical error about 6 points a game, 4 points over a season). Pass blockers count as being on the field.