최근 완료된 8시간 회차 사이 GitHub 스타 증가량 순
Use Zwift Click, Zwift Play, Shimano Di2, or any Bluetooth controller with MyWhoosh, Rouvy, Zwift, and more. Free to try on iOS, Android, Windows, and macOS.
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기Floor & Ceiling — NFL fantasy projections: PyTorch quantile transformer, walk-forward evals, self-updating GitHub Pages draft board
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
NFL fantasy football projections: a small quantile transformer (trained on a
free Kaggle T4) versus classical baselines, evaluated honestly with
walk-forward backtests, published as a static site that updates itself weekly
during the season.README 전문 보기 ↗NFL fantasy football projections: a small quantile transformer (trained on a free Kaggle T4) versus classical baselines, evaluated honestly with walk-forward backtests, published as a static site that updates itself weekly during the season.
The site is four pages, no backend and no framework: a draft board with a pick-time optimizer that reads a live Sleeper draft, a trade calculator, weekly projections with p10/p50/p90 bands, and an about page carrying the backtest numbers. Models predict raw stat lines; fantasy points are computed from those by pure scoring functions, so PPR, half-PPR, standard and this league's custom rules all derive from one set of predictions.
Start here: docs/case-study.md, a two-page overview of what was built, how the numbers
are kept honest and what they say. The full record is docs/methodology.md.
Live site: megatron-ff.vercel.app ·
Design spec: docs/superpowers/specs/2026-07-09-fantasy-football-model-design.md
python -m venv .venv # Python >= 3.10
source .venv/Scripts/activate # POSIX: source .venv/bin/activate
.venv/Scripts/python.exe -m pip install -e ".[dev]" # POSIX: .venv/bin/python
pytest # unit tests (offline)
pytest -m integration # network tests against live nflverse data
python -m ffmodel.data.pull # cache 2012-2025 data to data/raw/
python -m ffmodel.eval.run # walk-forward backtest -> models/backtests/baselines.json
The live contracts were checked on 2026-09-07. Both leagues use full PPR,
6-point passing TDs, -2 per interception, and an additional -3 for a pick-six
thrown. The YAML files in configs/leagues/ record every nonzero Sleeper
scoring category, including the leagues' different kicking and long-TD rules.
The board displays its league and scoring; navigation preserves the choice
per tab. The in-season pages (Weekly, FAAB & waivers, Trade) are not
tied to these two leagues: they read any Sleeper league's own settings from the
league-neutral data (see docs/remaining-season-projections.md).
Before either draft, run this read-only check from the repository root:
.\.venv\Scripts\python.exe tools/check_draft_readiness.py
It compares both boards with their configs and live Sleeper league/draft settings, then runs the production JavaScript contract and optimizer checks. The live panel also rechecks settings during the draft and refuses a known scoring or roster mismatch. Standalone mocks must be set up to match the league; Sleeper's mock defaults can differ, and mock scoring cannot be verified through a league ID.
Projection limits are shown on each board. QB forecasts include an approximate
pick-six expected-cost adjustment, using the observed NFL pick-six/interception
rate from completed seasons in data_snapshots/pick_six_rates.json. Interception
volume is estimated by integrating a nonnegative, piecewise-linear curve through
the model's interception quantiles (point-only models use their point estimate).
The expected penalty shifts all weekly band endpoints equally before the normal
games-played simulation, including rookie QBs. It is not a trained pick-six head
and does not add discrete pick-six uncertainty or player-specific return rates.
Historical scoring is unchanged: an observed pick-six still costs the ordinary
interception penalty plus the extra penalty; missing actual counts remain zero.
50+ yard TD bonuses, two-point conversions, and return/recovery TDs remain
unprojected. K/DST reminders use ADP, not model point projections.
The manual draft refresh workflow rebuilds both leagues. Weekly refreshes
build Gabagool only; --league fam --week ... is refused before any output is
written. To regenerate a draft locally, use ffmodel.site.generate --draft
with --league fam or --league gabagool and the normal model/artifact arguments.
Training runs on Kaggle's free GPU tier (~30 h/week). The four
notebooks/*_kaggle.ipynb notebooks are thin wrappers — every line of logic
lives in src/ffmodel/, so the same commands run locally on CPU, slower.
*_kaggle.ipynb and Run All, then walk away (~1 hour for a
12-run sweep). The first cell clones or pulls the repo and installs it.Sessions can be cut off at any time, so every training loop checkpoints each epoch: re-running the notebook skips completed runs and resumes an interrupted one from its checkpoint. Nothing is scheduled here — all recurring work runs in GitHub Actions (below).
notebooks/train_studio_lab.ipynb is the original SageMaker Studio Lab
notebook, kept because it still runs. Studio Lab closed to new customers in
July 2026; new work should use the Kaggle notebooks.
Fairness note carried over from the bake-off: the transformer reserves the season right before each test year as an early-stopping validation set, while the baselines are fit through that season with no holdout — a small handicap for the transformer, called out here rather than hidden.
The main workflows in .github/workflows/:
ci.yml runs the Python suite (pytest -W error) and every Node fixture on each push to main and on every
pull request.weekly-update.yml regenerates the site JSON in season. It runs twice daily (23 5,17 * 9-12,1 *, September
through January) so that 72-hour consumers stay fresh, and it can also be triggered manually. The site is
served from site/ by Vercel, which redeploys on every push to main.weekly-accuracy.yml scores the projections the site actually published before kickoff against what happened,
every Tuesday and Wednesday. It publishes the scorecard in models/diagnostics/live_<season>_weekly.json.market-snapshots.yml runs after each site update. It archives third-party market data (projections, game
lines, player props) into a private repository, for future pre-registered tests. Nothing it captures is
committed here.prospective-freeze.yml / prospective-outcomes.yml freeze a pre-registered 2026 forecast, then capture its
outcomes after the season.Fail-safe contract: weekly-update.yml generates the site JSON before it
commits anything. If the data pull or generation step fails for any
reason, the job fails before the commit step runs, so nothing is committed
and nothing deploys — the site keeps serving last week's data, with its
"data as of" stamp still honestly showing when that data was generated.
Deployed model: the env: block at the top of weekly-update.yml
selects the model. It is set to MODEL: transformer with ARTIFACT_ROOT: models/transformer/v1,models/transformer/v1_s43,models/transformer/v1_s44 —
three run roots, not through<year> directories; the predictor appends
through{last-trained-season} to each and averages the quantiles. A single
root deploys a single model; the comma-separated list is what makes it a
3-seed ensemble.
Preseason draft-board refresh: before week 1, run weekly-update.yml
manually with the draft input checked. This regenerates only the draft
board (--draft, no --week), since the target season has no games yet
and requesting its weekly stats would fail; the weekly slate resumes once
the season starts (cron or plain dispatch, which use --week auto).
Walk-forward only: train on seasons ≤ S, test on S+1, held-out years
2023-2025. There are no random splits anywhere — rolling features make them
leak the future into the past. Weekly PPR points, pooled over the three
held-out seasons (models/backtests/bakeoff.json, n = 17,908 player-weeks):
| model | MAE | QB | RB | WR | TE |
|---|---|---|---|---|---|
| naive last-4 average | 4.612 | 6.57 | 4.44 | 4.62 | 3.75 |
| XGBoost | 4.448 | 6.34 | 4.33 | 4.47 | 3.54 |
| transformer (3-seed ensemble) | 4.325 | 6.24 | 4.33 | 4.25 | 3.45 |
The transformer is ahead everywhere, but read the size before the ranking: 0.12 points per player-week overall, and RB is 4.3253 against 4.3294 — a win in the fourth decimal, which is a tie. WR and QB are where the gap is real. The baselines run through the identical eval harness, and the deployed model is whichever one actually won.
Two later experiments were pre-registered, and neither was promoted. Both
verdicts are committed under models/diagnostics/ next to the criteria, which
record recorded_before_results: true. The two failed differently, and the
difference is the point:
air_share, team_pass_att_last4, is_indoor)
passed both declared criteria — MAE below the committed 4.3263 bar,
calibration inside [0.75, 0.85] for every position-season. It was withheld
anyway, on checks that were explicitly post-hoc: a paired test over the
17,908 identical held-out rows puts the gain inside noise, the sign flips
across seasons, and QB regresses significantly (+0.082 MAE, CI excluding
zero). Passing the letter of a gate is not the same as having an effect.The site still serves v1. A negative result that is cheap to run and honestly recorded is worth more than a promotion that would not have replicated.
Scope guards (v1), deliberately: QB/RB/WR/TE only — no K/DST/IDP projections, no DFS optimization, no injury or news signals.