What actually predicts next week's fantasy points?

Every week, fantasy football managers choose between players using recent form, matchups, and Vegas lines. This project tests those instincts against 27 NFL seasons (1999-2025, 144,000 player-games): how the game changed, what never did, a projection model scored on seasons it never saw, a season-long model for drafts, and live projections for the 2026 season (this week's).

Live, phone-friendly version: joshwilliams332.github.io/nfl-historic-analysis

Python · pandas · scikit-learn · matplotlib · Jupyter · Power BI (DAX) · JavaScript · data from nflverse

Findings: the modern game (2021-2025)

1. Volume is a trait. Touchdowns and efficiency are mostly noise.

Averaging a stat over odd weeks and over even weeks, then correlating the two halves, separates role from luck. Targets and carries repeat almost perfectly (r ≈ 0.9). Touchdowns repeat far less (r ≈ 0.4-0.6), and yards per target or per carry hardly repeat at all. When a player's points outrun their volume, expect a pullback.

2. The full season beats the hot hand

The last game is the weakest predictor at every point in the season, and the last three games trail the season-to-date average. Early in the year, last season's average is as informative as this season's.

3. Matchups and Vegas totals matter, but far less than people assume

From the toughest fifth of matchups to the softest, scoring moves about 10% relative to a typical week. The Start/Sit heuristic on my fantasy site (Mingus Sports) assumes about 38% across the same groups. Vegas implied totals add about 5% once a player's average is known.

4. A simple model beats every rule of thumb on unseen games

Trained on 2021-2023, tuned on 2024, tested once on 2025 (2,591 player-games):

Projection	MAE (pts)	Correlation	Start/sit accuracy
Last game	7.53	0.290	58.7%
Last 3 games	6.24	0.400	62.6%
Season average	5.87	0.450	64.1%
Mingus site formula	6.08	0.432	64.0%
Gradient boosting	5.63	0.475	64.9%
Ridge regression	5.67	0.488	65.5%

Start/sit accuracy: across every pair of same-position players in the same week, how often the one projected higher actually scored more (50% = coin flip).

A linear model was enough. Gradient boosting didn't find interactions worth its extra complexity.

5. Know when it's a coin flip

When two players' projections are under 2 points apart, no projection is much better than a coin. At 6+ points, the call is right about 4 times in 5.

What the model weighs

Target share, this season's average, and last season's average carry the most weight. Touchdowns per game is negative once points and volume are known: the model has learned to expect touchdown streaks to fade.

Findings: 27 seasons of history (1999-2025)

6. The game changed...

Wide receivers' share of top-150 fantasy points fell from about 46% to about 37% (3-season average). Quarterbacks and tight ends took most of it.

The 300-carry back went from about 10 a season in the early 2000s to none in 2023, with a small rebound since.

7. ...but what predicts fantasy points didn't

Season average beat the last 3 games, and the last 3 beat the last game, in all 27 seasons. Finding 2 is not a quirk of the modern NFL.

Touchdown spikes fade over full seasons too. The top fifth by touchdowns per touch dropped from 9.8% to 6.4% the next year, while their touches per game held steady. Year-to-year correlation: 0.54 for TD rate, 0.85 for touches per game.

8. Aging curves

Delta method (each player compared only with themselves, year over year). Running backs peak at about 24 and fall off steeply after 26. Receivers and tight ends peak around 25-26 and decline more gradually. Quarterbacks peak around 30 and hold up into their mid-30s. Survivors only, so real declines are steeper.

9. More history doesn't make a better weekly model

Ridge trained on	Rows	2025 MAE	Start/sit accuracy
2021-2023	7,894	5.674	65.5%
2015-2023	23,280	5.668	65.5%
2009-2023	38,186	5.670	65.5%

Five times the data, the same accuracy. The ceiling is the noise in a single game, not a shortage of examples. Better information (injuries, snap counts, depth charts) is what would move it.

Better information: snap counts and injury reports

Players who suit up with an injury tag score less than usual: 11% below a typical week when listed questionable, 15% when they missed practice. But only about 6% of player-games carry a tag, so adding the report barely moves overall accuracy.

Snap share (fraction of offensive plays in earlier games) is a small but real gain. On 2025 it saves 0.06 points of MAE, with a 95% interval of +0.03 to +0.10 from a bootstrap over weeks, and lifts start/sit accuracy from 65.5% to 66.1%. The live model uses both.

Season-long projections (draft prep)

Next season's PPR per game from this season: production, the season before, volume, touchdowns, snap share, and a separate age curve per position. Predicting 2024 and 2025, it cuts the error of "same as last season" from 3.26 to 2.84 points per game and orders players within a position better (rank correlation 0.61 → 0.68).

The honest 2026 check, about seven weeks in: on raw error the model is slightly behind "same as 2025" (3.28 vs. 3.23), mostly because 2026 is the highest-scoring season in the sample and players have declined less than usual. Its read on who changes is weak so far (r ≈ 0.2 between predicted and actual change): the players it expected to rise are up, but its expected fallers haven't dropped more than everyone else. Worth rechecking after the season.

Live: the 2026 season

Trained only on 2021-2025, the weekly model beat the season average in every week of 2026 so far: 65.8% of start/sit calls right vs. 62.3%, and a 5.68-point average miss vs. 6.20.

Week 8 projections → (also as CSV)

Every week: python scripts/project_week.py projects the next unplayed week. Run it again after Friday's injury reports to include the tags.

Interactive site (phone-friendly)

site/ is a static web version of the Power BI report (same seven views, buttons, and sliders) built for phones first, published with GitHub Pages. Refresh its data with:

python scripts/build_powerbi_data.py && python scripts/build_site_data.py

Interactive Power BI report

powerbi/ holds a Power BI project built on the same data: buttons and sliders to test how well any this-season stat predicts next season, find historical seasons that match a profile (and see what happened next), explore 27 seasons of trends and individual players, browse this week's projections, and compare receivers with corners (2025-26) by team, game and side of the field. See powerbi/README.md to open it.

Method

• No look-ahead. Every feature for week t uses only games before t: rolling stats are shifted one game, and a defense's points-allowed figure only counts earlier weeks. Unit tests in tests/ check this on toy data.
• Time-based split, not random: a random split would let the model learn from games after the ones it predicts.
• Sample: QB/RB/WR/TE player-games where the player had 3+ earlier games that season averaging 5+ PPR, which roughly covers anyone a manager would consider starting.
• Data cleaning: player stats use today's franchise codes for every season (LV, LAC, LA) while schedules use the city at the time (OAK, SD, STL), so schedules are mapped to franchises. nflverse has almost no targets for 2003-2008; those are set to missing rather than zero, and model training never starts before 2009.
• Matchup feature: PPR points a defense allowed per game to the position, ÷ the league average over the same weeks. Vegas feature: implied team total = (over/under ± spread) / 2.

Limitations

• Only players who played appear in the data, so injuries and surprise inactives (a big part of real start/sit risk) aren't modeled.
• Snap counts and injury reports are in; routes, depth charts, and news are not.
• The season model predicts points per game if a player plays: about 15% of qualifying players don't get 4+ games the next season, and it doesn't see offseason moves or project rookies.
• Defense-vs-position numbers depend on which players a defense happened to face; an opponent-adjusted version would likely make the matchup effect a bit stronger.
• Scoring is full PPR; results for half-PPR and standard should be similar but weren't tested.

Run it

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
pytest                      # feature leakage tests
jupyter lab notebooks/      # run 01 -> 07 in order; ~250 MB of nflverse data downloads to data/raw on first run
python scripts/project_week.py   # projections for the next unplayed week

Layout

notebooks/
  01_data_and_features.ipynb     load nflverse data, scoring check, features, leakage spot check
  02_what_is_predictable.ipynb   reliability, recent-form windows, matchup and Vegas effects
  03_projection_model.ipynb      baselines vs. ridge vs. gradient boosting, start/sit evaluation
  04_history.ipynb               1999-2025: position shares, workhorse RBs, era checks, TD regression, aging, training window
  05_snaps_and_injuries.ipynb    snap share and injury reports as features, bootstrap confidence intervals
  06_season_projections.ipynb    season-long model for drafts, 2026 preseason projections vs. 2026 so far
  07_live_2026_week8.ipynb       2026 backtest and week 8 projections
scripts/
  project_week.py         rerunnable weekly projections -> reports/projections/
  build_powerbi_data.py   exports the CSVs the Power BI report loads -> powerbi/data/
  build_site_data.py      compact JSON for the web version -> site/data/
site/                     phone-first web version (GitHub Pages)
powerbi/                  Power BI project (semantic model + report), see its README
src/fantasy/
  data.py        download + cache nflverse player weeks, schedules (Vegas lines), snap counts, injuries, birth dates
  features.py    leakage-safe features (incl. snaps, injury tags, upcoming-week rows), the Mingus heuristic
  weekly.py      the weekly projection model: build, fit, project a week
  season.py      season-long projection model
  evaluate.py    MAE / start/sit accuracy scoring
  plots.py       shared chart style (colorblind-checked palette)
tests/           unit tests for the no-look-ahead guarantees
reports/figures/ charts used above
reports/projections/ weekly projection tables