한국어

로그인 프로젝트 공개

Gimme a W

축구·NFL·대학 미식축구의 점수·순위·통계·예측 제공

원문 보기

Scores, standings, stats and predictions for soccer, the NFL and college football.

스포츠·피트니스Python · TypeScriptGitHub @bjhr730온라인운영자 미확인

제품 방문하기 GitHub

써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.

가동 상태: 온라인; 30일 가동률 99.1%
최근 코드 업데이트: 10월 28일; 활성 저장소
최신 버전: 없음
기여자: 1명; 포크 0
nomorevibe 등록: 9월 22일; 저장소 생성 2026년 9월

언어 구성

Python 59.8%
TypeScript 38.4%
JavaScript 0.9%
HTML 외 0.9%

업데이트0건

전체

아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.

최신순

···

GitHub 이슈 수신 설정

로그인 상태를 확인하는 중입니다.

댓글을 불러오는 중입니다.

Gimme a W — nomorevibe

Gimme a W

축구·NFL·대학 미식축구의 점수·순위·통계·예측 제공

원문 보기

Scores, standings, stats and predictions for soccer, the NFL and college football.

스포츠·피트니스Python · TypeScriptGitHub @bjhr730온라인운영자 미확인

제품 방문하기 GitHub

써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.

가동 상태: 온라인; 30일 가동률 99.1%
최근 코드 업데이트: 10월 28일; 활성 저장소
최신 버전: 없음
기여자: 1명; 포크 0
nomorevibe 등록: 9월 22일; 저장소 생성 2026년 9월

언어 구성

Python 59.8%
TypeScript 38.4%
JavaScript 0.9%
HTML 외 0.9%

업데이트0건

전체

아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.

최신순

···

GitHub 이슈 수신 설정

로그인 상태를 확인하는 중입니다.

댓글을 불러오는 중입니다.

소개

My betting little helper

저장소 README에서 가져옴

README

gimme-a-W

Live: https://gimme-a-w-web.vercel.app

A sports prediction web app. It collects statistics, team and player form, betting lines and expert picks from public sports sites (ESPN, Sports-Reference, SI.com and structured sports APIs), then predicts game outcomes and player performance for three verticals:

⚽ Soccer, worldwide — top domestic leagues, continental cups, international tournaments
🏈 NFL
🏈 College football (NCAA FBS)

Before any game, open one page and see who is likely to win or draw, who is likely to score, and what the totals should look like: goals, corners, shots on target, passing, rushing and receiving yards.

Tech stack

The app is a hybrid TypeScript + Python monorepo. The web layer is TypeScript because it deploys anywhere and renders fast; the data-collection and modeling layers are Python because that is where the sports-data and statistics ecosystems live.

Layer	Choice	Why
Web app	Next.js 15 (App Router) + TypeScript + Tailwind CSS + shadcn/ui, mobile-first and installable as a PWA	Server rendering and ISR for score/stat/prediction pages, responsive layouts that work on a phone at the stadium, one-click deploy on Vercel
Collectors	Python 3.12+ with `httpx`, `selectolax` (HTML parsing), `pydantic` (validation), `Playwright` only when a page needs JavaScript	Existing libraries already wrap the sources we care about: `soccerdata` (FBref, ESPN, Understat, FotMob, Sofascore, WhoScored), `nflreadpy` (nflverse), `cfbd` (CollegeFootballData)
Prediction models	Python: `polars`/`pandas`, `numpy`, `scipy`, `statsmodels`, `scikit-learn`, `lightgbm`	Poisson and negative-binomial models for goals, corners, shots; gradient boosting for yards; everything back-testable with standard tooling
Database	PostgreSQL on Neon	Relational data with lots of joins (team × season × game × stat); free tier is enough for v1
Schema / migrations	Drizzle ORM (`packages/db`) is the single source of truth; Python writes through `psycopg`	One schema definition, typed queries on the web side, plain SQL upserts on the collector side
Scheduling	GitHub Actions cron for v1; move to a small worker (Railway / Fly.io + APScheduler) if we need sub-hourly runs	Zero infrastructure, logs live next to the code
HTTP cache	On-disk cache inside the fetcher + a table storing fetched HTML/JSON

Alternative considered: an all-TypeScript setup (Node worker with cheerio + BullMQ + Redis). Simpler to maintain in one language, but it means re-implementing the scrapers and the statistical models that the Python ecosystem already provides. Not worth it for a prediction product.

Data sources

Each source is wrapped by one adapter under collectors/gimme_collectors/sources/. Adapters share one interface (fetch → parse → normalize → upsert) and emit the unified domain model below.

Primary sources

Source	Covers	How we access it	Limits / notes
ESPN	Soccer (dozens of leagues), NFL, college football — scores, fixtures, standings, rosters, box scores, FPI predictions	Unofficial but stable JSON endpoints under `site.api.espn.com/apis/site/v2/sports/{sport}/{league}/…` (e.g. `soccer/eng.1`, `football/nfl`, `football/college-football`)	No key required. Undocumented, so pin response shapes with tests. First adapter to build: one integration covers all three verticals
Sports-Reference (FBref, Pro-Football-Reference, CFB Reference)	Deep historical and advanced stats: shots on target, xG, per-player shooting, NFL/CFB player game logs	Throttled HTML scraping of stat tables; `soccerdata` handles FBref throttling for us	Hard limit: ~20 requests/minute or IP ban. Terms forbid commercial use without a license. Cache every page
SI.com and other editorial sites	Previews, expert picks, injury news, form commentary	RSS feeds + article HTML, honoring `robots.txt`	We store a short summary, the URL and the teams mentioned — never the full text (copyright)

Structured sources the models depend on

Attribution and terms. This is a personal, non-commercial project. The data belongs to the people who publish it: football-data.org issues a personal token and asks to be credited; football-data.co.uk is free to use and asks that it not be resold; nflverse is open data. The deployed site sends noindex and a robots.txt that disallows everything, because an indexed public site is redistribution rather than personal use. Nothing here is betting advice, and none of it is affiliated with any league, club or provider.

Source	Covers	Key	Free tier	Feeds which predictions
nflverse	The whole NFL: schedule and results with closing lines, weekly player stats, the league injury report, team divisions	No	Unlimited (static files)	Win probability, anytime TD, passing/rushing/receiving yards, the NFL table
football-data.co.uk	Historical CSVs since 1993 for 22+ leagues: results, shots on target, corners, closing odds	No	Unlimited (static files)	Back-testing of 1X2, totals, corners and shots models
football-data.org	All soccer fixtures, results and tables — one request per competition returns a whole season	Yes (free)	10 req/min	Every soccer market
The Odds API	Bookmaker lines for all three sports	Yes	500 req/month	Market baseline every model is measured against

Architecture

            ┌──────────────────────── GitHub Actions (cron) ────────────────────────┐
            │  python -m gimme_collectors run espn --sport nfl                      │
            │  python -m gimme_collectors run fbref --competition ENG-Premier-League │
            │  python -m gimme_predict   run --sport soccer --date today            │
            └───────────────┬───────────────────────────────────────────────────────┘
                            │ fetch (cached, rate-limited)
   ESPN · Sports-Reference · SI · CFBD · nflverse · football-data.co.uk · Odds API
                            │ parse → pydantic models → normalize ids → upsert
                            ▼
                   PostgreSQL (Neon)  ◄──── Drizzle schema + migrations
                            │
                            │ features → models → predictions (per game, team, player, market)
                            ▼
            Next.js 15 (Vercel) — server components read via Drizzle, ISR pages
            /  /soccer  /nfl  /cfb  /games/[id]  /teams/[id]  /players/[id]  /models

Key design rules:

Raw first, parse second. Every fetched page/JSON is stored in raw_page with a hash and timestamp. Parsers run against stored raw data, so a parser bug never forces a re-scrape.
One id to rule them all. Internal ids are ours; every source id lives in the external_id mapping table. Cross-source joins (ESPN roster + FBref stats + odds) go through that table.
Provenance everywhere. Every row carries source_id and collector_run_id; every prediction carries model_run_id, so a bad run can be rolled back.
Idempotent upserts. Collectors and model runs can be re-run for the same day/week without duplicating rows.
Point-in-time features. A prediction for a game only uses data available before kickoff. This is what makes back-tests honest.

Prediction engine

The engine lives in collectors/gimme_predict/. Every market is a small, explainable model built on the collected data, and every model is measured against the bookmaker line as the baseline.

Markets

Market	Sport	Output	Model
Match result (home / draw / away)	Soccer	Three probabilities	Dixon-Coles bivariate Poisson on team attack/defense strengths, time-decayed, blended with bookmaker implied probabilities
Win probability and spread	NFL, CFB	Home win probability, expected margin	Elo with margin of victory + LightGBM on efficiency features (EPA per play, success rate, rest days, injuries), calibrated against closing lines
Total goals	Soccer	Expected goals per team, over/under probabilities for 1.5 / 2.5 / 3.5	Same Poisson strengths as match result; xG-adjusted
Anytime goalscorer	Soccer	Probability and expected goals per player	Share of the team's expected goals, from the player's recent goals and shots per appearance and the minutes they are likely to play; Poisson
Anytime assist	Soccer	Probability and expected assists per player	The team's expected goals scaled by the share that are assisted, split by a player's assist share shrunk toward their share of minutes
Player shots on target	Soccer	Expected count per player, 1+ and 2+ probabilities	Share of the team's expected shots on target, from the player's recent rate per appearance; Poisson
Shots on target	Soccer	Expected count per team, over/under lines	Negative-binomial regression on team and opponent rates, home advantage, game state
Total corners	Soccer	Expected count, over/under lines	Negative-binomial regression on team corner rates for and against, possession, tempo
Anytime touchdown scorer	NFL, CFB	Probability per player	Red-zone target and carry share × team expected touchdowns; Poisson
Passing / rushing / receiving yards	NFL, CFB	Expected yards per player with 25th–75th percentile range	LightGBM quantile regression on usage (target share, carries, snap %), opponent defense EPA allowed, expected game script (spread, total), weather

Two corrections keep player markets honest early in a season, when a squad has played two or three games. Per-appearance rates are shrunk toward the squad average, so one goal in one appearance is not read as a goal a game. Shares are diluted when only part of a squad has appeared, so the players on record do not split a total that belongs to the whole team.

Availability

player_status holds one current report per player per source: the source's own wording ("Questionable", "Injured Reserve"), a normalized verdict, and the share of games players carrying that verdict have historically been active for. ESPN publishes a full report for the NFL, a thin one for college football, and none for soccer, so a player with no row means no report and never fit.

The prop models act on it. A player reported out is left out of the markets entirely: publishing 68 rushing yards for a player on injured reserve is worse than publishing nothing. A doubt keeps their projection, which is conditional on them taking the field, but has the probability markets scaled by their chance of playing. Each run records how many players it withheld and flagged.

How a prediction is produced

Feature snapshot. features.py builds one row per game (and per player) using only data timestamped before kickoff: last-N form, season rates, ratings, injuries, odds.
Model run. Each market model reads the snapshot, writes rows to prediction with mean, probability, line, quantiles and model_version.
Ensemble. When third-party numbers exist (ESPN FPI, SP+, closing odds) they enter as features or as a blend weight, never as the answer.
Scoring. After results land, evaluate.py scores every prediction: log loss and Brier for probabilities, MAE and coverage for counts and yards, and always side-by-side with the bookmaker baseline. Results feed the /models page.
Explanation. Each prediction stores its top feature contributions so the UI can say why (e.g. "Arsenal 1.9 xG: strong home attack, opponent conceding 1.6 xG per game, no defensive injuries").

What "good" looks like

Match result and win probability: log loss within 1–2 % of the closing line after one season of back-tests.
Totals, corners, shots: MAE at or below the market's implied line.
Yards: 50 % of outcomes inside the predicted 25th–75th percentile range.

Where we do not clear the bar: NFL totals. Measured over 2,494 games with a closing line, the model's total is unbiased (mean error −0.04 points) but less accurate than the line: MAE 10.87 against 10.51. Blending the two helps at no weight — the best mix is 100 % market — and anchoring the model on the line reproduces it, landing within 0.42 points of it and beating it in 50.1 % of games. Scoring rates and Elo are already in the price. Beating an NFL total needs inputs we do not collect: pace, pass rate over expected, personnel groupings, and injuries to specific skill players. Until then the published total is shown with its middle-half range and the game page says the line is the sharper number, so a gap reads as disagreement rather than an edge.

Predictions are probabilities, not promises. The UI always shows the uncertainty.

Domain model

Entity	Purpose
`sport`	`soccer`, `american_football`
`competition`	Premier League, La Liga, NFL, NCAAF FBS, Champions League… (has `sport`, `country`, `level`)
`season`	A competition-year (`2025-26`, `2026`)
`team`	Club or program, with `competition` membership per season
`player`	Person; roster links to `team` + `season`
`game`	Fixture or result: kickoff, home/away, score, status, venue, week/round, weather
`team_game_stat`	Per-team box score (possession, xG, shots on target, corners, yards, turnovers…) — JSONB `stats` keyed by a per-sport dictionary
`player_game_stat`	Per-player box score: goals, shots, minutes, targets, carries, yards, touchdowns
`standing`	Snapshot of table/rankings per competition-season-date
`team_form`	Derived: last-N results, points-per-game, xG diff, rest days — recomputed after each run
`rating`	Team or player rating per date: Elo, attack/defense strength, xG per 90
`odds`	Bookmaker lines per game and market with timestamp (opening and closing)
`feature_snapshot`	Point-in-time feature row per game/player used by a model run
`prediction`	`model_run`, `game`, `market`, `subject_type` (game/team/player), `subject_id`, `probability`, `mean`, `line`, `quantiles`, `explanation`
`model_run`	Which model version ran, when, on which snapshot, with which metrics
`pick`	External picks: `source`, `game`, `pick`, , , (ESPN FPI, SI experts)

Repository layout

gimme-a-W/
├── apps/
│   └── web/                     # Next.js 15 app (App Router, Tailwind, shadcn/ui)
├── collectors/
│   ├── pyproject.toml           # managed with uv; both packages below
│   ├── gimme_collectors/
│   │   ├── sources/             # espn.py · sports_reference.py · si.py · cfbd.py · nflverse.py · fdcouk.py · odds.py
│   │   ├── models/              # pydantic: Team, Player, Game, TeamGameStat, Pick, Article…
│   │   ├── pipeline/            # fetch.py (cache + throttle) · normalize.py · upsert.py
│   │   └── cli.py               # python -m gimme_collectors run <source> [--sport …]
│   └── gimme_predict/
│       ├── features.py          # point-in-time feature snapshots
│       ├── ratings.py           # Elo, attack/defense strengths
│       ├── markets/             # match_result.py · totals.py · goalscorer.py · shots.py · corners.py · td_scorer.py · yards.py
│       ├── evaluate.py          # log loss, Brier, MAE, coverage vs bookmaker baseline
│       └── cli.py               # python -m gimme_predict run|backtest
├── packages/
│   └── db/                      # Drizzle schema, migrations, seed
├── .github/workflows/           # ci.yml · collect-daily.yml · predict-daily.yml · backtest-weekly.yml
├── assets/logo/                 # gaw-logo.svg — the logo; the header, the favicon and the app icons are all built from it
├── data/                        # local raw cache and model artifacts (git-ignored)
├── README.md
└── .gitignore

Local setup

Prerequisites: Node 20+ (24 recommended), pnpm, Python 3.12+, uv, PostgreSQL (a free Neon branch is the easiest).

git clone https://github.com/bjhr730/gimme-a-W.git
cd gimme-a-W

Web app and shared packages:

pnpm install
cp .env.example .env            # fill in DATABASE_URL
pnpm --filter @gimme/db migrate # apply Drizzle migrations
pnpm --filter web dev           # http://localhost:3000

Collectors and models:

cd collectors
uv sync
uv run gimme-collect run espn --kind all --league nfl --league eng.1 --date yesterday --days 3
uv run gimme-collect run espn --kind roster --league nfl
uv run gimme-collect derive --what form,elo
uv run pytest

--kind all collects scoreboards, teams, standings and game summaries (box scores, lineups, ESPN FPI picks). Rosters are separate because they cost one request per team.

Predictions (the daily workflow runs these after collecting):

uv run gimme-predict backtest --competition nfl     # season-by-season vs the closing line
uv run gimme-predict run --sport american_football  # writes predictions for the next 8 days
uv run gimme-predict run --sport soccer --dry-run   # needs ~30 finals per league to fit

History loads (one-off, then refreshed weekly by the workflow):

uv run gimme-collect run nflverse --seasons 2015-2026 --what games
uv run gimme-collect run nflverse --seasons 2024-2025 --what players
# soccer seasons from ESPN: one request per calendar month, results + shots + corners
uv run gimme-collect run espn --kind scoreboard --league eng.1 --date 2019-08-01 --days 2555
# closing odds from football-data.co.uk (when the site is up)
uv run gimme-collect run fdcouk --league eng.1 --league esp.1 --seasons 2015-2025

Environment variables

Variable	Used by	Required	Notes
`DATABASE_URL`	web, collectors, models	Yes	Neon pooled connection string
`DIRECT_URL`	Drizzle migrations	Yes	Neon direct (non-pooled) connection
`FOOTBALL_DATA_API_KEY`	collectors	Optional	football-data.org
`ODDS_API_KEY`	collectors	Recommended	the-odds-api.com; the models use lines as baseline
`COLLECTOR_CACHE_DIR`	collectors	No	Defaults to `./data/raw`
`MODEL_ARTIFACT_DIR`	models	No	Defaults to `./data/models`
`COLLECTOR_USER_AGENT`	collectors	No	Identify the bot politely, with a contact email

Never commit .env. Commit .env.example with empty values instead.

Roadmap

Phase	Deliverable
0 — Scaffold (done)	README, `.gitignore`, logo, monorepo skeleton, tooling
1 — Data spine (done)	Drizzle schema + migrations · ESPN adapter for soccer/NFL/CFB · CLI · CI and daily collect workflows
2 — Web v1 (done)	Mobile-first Next.js app · today's games per sport with a date strip · game pages with stats and lines · team and standings pages · search · PWA manifest and icons from the logo. Player pages wait for roster data in phase 3
3 — Depth (done, ESPN part)	ESPN game summaries (full team box scores, per-player passing/rushing/receiving and goals/shots, soccer lineups, FPI picks) · ESPN rosters · `team_form` and Elo `rating` derivations · player pages, box scores on game pages, form/Elo/roster on team pages · weekly roster workflow
3b — Historical depth (done)	nflverse: NFL games since 2015 with closing lines, weather and rest days, weekly player stats with EPA and target share · football-data.co.uk: league seasons with shots on target, corners and Bet365/Pinnacle opening and closing odds, matched to ESPN teams by name · Elo with season regression · weekly refresh workflow. CFBD when a key is added; Sports-Reference last, under its rate limit
4 — Game predictions (done)	`gimme_predict`: point-in-time features (Elo, rest, form), logistic win probability + ridge spread and total for football, Dixon-Coles Poisson for soccer (1X2, total goals, team goals, both teams to score), 60/40 blend with the market, season-by-season back-test vs the closing line stored in `model_run`, daily prediction run · "Who gets the W" panel on game pages, model pick chips on score cards, `/models` page
5 — Team and player markets (done)	Passing/rushing/receiving yards with middle-half ranges and anytime TD (usage share, opponent allowed, market game script; holdout back-test), shots on target and corners per team and per match (negative-binomial GLM with over/under lines), anytime scorer from expected-goals share · markets panel on game pages · daily run
Research search (done)	Header search that understands teams, matchups, players, leagues and days, with keyboard navigation; `/api/search`
7 — Availability and parlays (done)	ESPN injury feed (, in the daily run) into · prop models withhold players reported out and scale probability markets for doubts · availability shown on game, team and player pages and as badges on market rows · soccer goals, assists and shots on target per player with expected quantities, shrunk rates and squad-coverage dilution · page, 2 to 20 legs, one per game · mobile back button · daily job re-reads the last three weeks so a missed run cannot leave a gap in the season

Legal and etiquette

Respect each site's robots.txt and terms of service. Sports-Reference explicitly bans scrapers above ~20 requests/minute and restricts commercial use; ESPN's JSON API is unofficial and may change without notice.
Identify the collector with a descriptive User-Agent that includes a contact email.
Cache aggressively; never fetch the same page twice in a day.
Store facts (scores, stats, lines) and short summaries with attribution links. Do not republish article text.
This project is for personal/educational use. Check licensing before any commercial use of the data.
Predictions are statistical estimates for information only, not betting advice.

README 전문 보기 ↗

README

gimme-a-W

Live: https://gimme-a-w-web.vercel.app

⚽ Soccer, worldwide — top domestic leagues, continental cups, international tournaments
🏈 NFL
🏈 College football (NCAA FBS)

Tech stack

Layer	Choice	Why
Web app	Next.js 15 (App Router) + TypeScript + Tailwind CSS + shadcn/ui, mobile-first and installable as a PWA	Server rendering and ISR for score/stat/prediction pages, responsive layouts that work on a phone at the stadium, one-click deploy on Vercel
Collectors	Python 3.12+ with `httpx`, `selectolax` (HTML parsing), `pydantic` (validation), `Playwright` only when a page needs JavaScript	Existing libraries already wrap the sources we care about: `soccerdata` (FBref, ESPN, Understat, FotMob, Sofascore, WhoScored), `nflreadpy` (nflverse), `cfbd` (CollegeFootballData)
Prediction models	Python: `polars`/`pandas`, `numpy`, `scipy`, `statsmodels`, `scikit-learn`, `lightgbm`	Poisson and negative-binomial models for goals, corners, shots; gradient boosting for yards; everything back-testable with standard tooling
Database	PostgreSQL on Neon	Relational data with lots of joins (team × season × game × stat); free tier is enough for v1
Schema / migrations	Drizzle ORM (`packages/db`) is the single source of truth; Python writes through `psycopg`	One schema definition, typed queries on the web side, plain SQL upserts on the collector side
Scheduling	GitHub Actions cron for v1; move to a small worker (Railway / Fly.io + APScheduler) if we need sub-hourly runs	Zero infrastructure, logs live next to the code
HTTP cache	On-disk cache inside the fetcher + a table storing fetched HTML/JSON

Data sources

Each source is wrapped by one adapter under collectors/gimme_collectors/sources/. Adapters share one interface (fetch → parse → normalize → upsert) and emit the unified domain model below.

Primary sources

Source	Covers	How we access it	Limits / notes
ESPN	Soccer (dozens of leagues), NFL, college football — scores, fixtures, standings, rosters, box scores, FPI predictions	Unofficial but stable JSON endpoints under `site.api.espn.com/apis/site/v2/sports/{sport}/{league}/…` (e.g. `soccer/eng.1`, `football/nfl`, `football/college-football`)	No key required. Undocumented, so pin response shapes with tests. First adapter to build: one integration covers all three verticals
Sports-Reference (FBref, Pro-Football-Reference, CFB Reference)	Deep historical and advanced stats: shots on target, xG, per-player shooting, NFL/CFB player game logs	Throttled HTML scraping of stat tables; `soccerdata` handles FBref throttling for us	Hard limit: ~20 requests/minute or IP ban. Terms forbid commercial use without a license. Cache every page
SI.com and other editorial sites	Previews, expert picks, injury news, form commentary	RSS feeds + article HTML, honoring `robots.txt`	We store a short summary, the URL and the teams mentioned — never the full text (copyright)

Structured sources the models depend on

Source	Covers	Key	Free tier	Feeds which predictions
nflverse	The whole NFL: schedule and results with closing lines, weekly player stats, the league injury report, team divisions	No	Unlimited (static files)	Win probability, anytime TD, passing/rushing/receiving yards, the NFL table
football-data.co.uk	Historical CSVs since 1993 for 22+ leagues: results, shots on target, corners, closing odds	No	Unlimited (static files)	Back-testing of 1X2, totals, corners and shots models
football-data.org	All soccer fixtures, results and tables — one request per competition returns a whole season	Yes (free)	10 req/min	Every soccer market
The Odds API	Bookmaker lines for all three sports	Yes	500 req/month	Market baseline every model is measured against

Architecture

            ┌──────────────────────── GitHub Actions (cron) ────────────────────────┐
            │  python -m gimme_collectors run espn --sport nfl                      │
            │  python -m gimme_collectors run fbref --competition ENG-Premier-League │
            │  python -m gimme_predict   run --sport soccer --date today            │
            └───────────────┬───────────────────────────────────────────────────────┘
                            │ fetch (cached, rate-limited)
   ESPN · Sports-Reference · SI · CFBD · nflverse · football-data.co.uk · Odds API
                            │ parse → pydantic models → normalize ids → upsert
                            ▼
                   PostgreSQL (Neon)  ◄──── Drizzle schema + migrations
                            │
                            │ features → models → predictions (per game, team, player, market)
                            ▼
            Next.js 15 (Vercel) — server components read via Drizzle, ISR pages
            /  /soccer  /nfl  /cfb  /games/[id]  /teams/[id]  /players/[id]  /models

Key design rules:

Raw first, parse second. Every fetched page/JSON is stored in raw_page with a hash and timestamp. Parsers run against stored raw data, so a parser bug never forces a re-scrape.
One id to rule them all. Internal ids are ours; every source id lives in the external_id mapping table. Cross-source joins (ESPN roster + FBref stats + odds) go through that table.
Provenance everywhere. Every row carries source_id and collector_run_id; every prediction carries model_run_id, so a bad run can be rolled back.
Idempotent upserts. Collectors and model runs can be re-run for the same day/week without duplicating rows.
Point-in-time features. A prediction for a game only uses data available before kickoff. This is what makes back-tests honest.

Prediction engine

The engine lives in collectors/gimme_predict/. Every market is a small, explainable model built on the collected data, and every model is measured against the bookmaker line as the baseline.

Markets

Market	Sport	Output	Model
Match result (home / draw / away)	Soccer	Three probabilities	Dixon-Coles bivariate Poisson on team attack/defense strengths, time-decayed, blended with bookmaker implied probabilities
Win probability and spread	NFL, CFB	Home win probability, expected margin	Elo with margin of victory + LightGBM on efficiency features (EPA per play, success rate, rest days, injuries), calibrated against closing lines
Total goals	Soccer	Expected goals per team, over/under probabilities for 1.5 / 2.5 / 3.5	Same Poisson strengths as match result; xG-adjusted
Anytime goalscorer	Soccer	Probability and expected goals per player	Share of the team's expected goals, from the player's recent goals and shots per appearance and the minutes they are likely to play; Poisson
Anytime assist	Soccer	Probability and expected assists per player	The team's expected goals scaled by the share that are assisted, split by a player's assist share shrunk toward their share of minutes
Player shots on target	Soccer	Expected count per player, 1+ and 2+ probabilities	Share of the team's expected shots on target, from the player's recent rate per appearance; Poisson
Shots on target	Soccer	Expected count per team, over/under lines	Negative-binomial regression on team and opponent rates, home advantage, game state
Total corners	Soccer	Expected count, over/under lines	Negative-binomial regression on team corner rates for and against, possession, tempo
Anytime touchdown scorer	NFL, CFB	Probability per player	Red-zone target and carry share × team expected touchdowns; Poisson
Passing / rushing / receiving yards	NFL, CFB	Expected yards per player with 25th–75th percentile range	LightGBM quantile regression on usage (target share, carries, snap %), opponent defense EPA allowed, expected game script (spread, total), weather

Availability

How a prediction is produced

Feature snapshot. features.py builds one row per game (and per player) using only data timestamped before kickoff: last-N form, season rates, ratings, injuries, odds.
Model run. Each market model reads the snapshot, writes rows to prediction with mean, probability, line, quantiles and model_version.
Ensemble. When third-party numbers exist (ESPN FPI, SP+, closing odds) they enter as features or as a blend weight, never as the answer.
Scoring. After results land, evaluate.py scores every prediction: log loss and Brier for probabilities, MAE and coverage for counts and yards, and always side-by-side with the bookmaker baseline. Results feed the /models page.
Explanation. Each prediction stores its top feature contributions so the UI can say why (e.g. "Arsenal 1.9 xG: strong home attack, opponent conceding 1.6 xG per game, no defensive injuries").

What "good" looks like

Match result and win probability: log loss within 1–2 % of the closing line after one season of back-tests.
Totals, corners, shots: MAE at or below the market's implied line.
Yards: 50 % of outcomes inside the predicted 25th–75th percentile range.

Predictions are probabilities, not promises. The UI always shows the uncertainty.

Domain model

Entity	Purpose
`sport`	`soccer`, `american_football`
`competition`	Premier League, La Liga, NFL, NCAAF FBS, Champions League… (has `sport`, `country`, `level`)
`season`	A competition-year (`2025-26`, `2026`)
`team`	Club or program, with `competition` membership per season
`player`	Person; roster links to `team` + `season`
`game`	Fixture or result: kickoff, home/away, score, status, venue, week/round, weather
`team_game_stat`	Per-team box score (possession, xG, shots on target, corners, yards, turnovers…) — JSONB `stats` keyed by a per-sport dictionary
`player_game_stat`	Per-player box score: goals, shots, minutes, targets, carries, yards, touchdowns
`standing`	Snapshot of table/rankings per competition-season-date
`team_form`	Derived: last-N results, points-per-game, xG diff, rest days — recomputed after each run
`rating`	Team or player rating per date: Elo, attack/defense strength, xG per 90
`odds`	Bookmaker lines per game and market with timestamp (opening and closing)
`feature_snapshot`	Point-in-time feature row per game/player used by a model run
`prediction`	`model_run`, `game`, `market`, `subject_type` (game/team/player), `subject_id`, `probability`, `mean`, `line`, `quantiles`, `explanation`
`model_run`	Which model version ran, when, on which snapshot, with which metrics
`pick`	External picks: `source`, `game`, `pick`, , , (ESPN FPI, SI experts)

Repository layout

gimme-a-W/
├── apps/
│   └── web/                     # Next.js 15 app (App Router, Tailwind, shadcn/ui)
├── collectors/
│   ├── pyproject.toml           # managed with uv; both packages below
│   ├── gimme_collectors/
│   │   ├── sources/             # espn.py · sports_reference.py · si.py · cfbd.py · nflverse.py · fdcouk.py · odds.py
│   │   ├── models/              # pydantic: Team, Player, Game, TeamGameStat, Pick, Article…
│   │   ├── pipeline/            # fetch.py (cache + throttle) · normalize.py · upsert.py
│   │   └── cli.py               # python -m gimme_collectors run <source> [--sport …]
│   └── gimme_predict/
│       ├── features.py          # point-in-time feature snapshots
│       ├── ratings.py           # Elo, attack/defense strengths
│       ├── markets/             # match_result.py · totals.py · goalscorer.py · shots.py · corners.py · td_scorer.py · yards.py
│       ├── evaluate.py          # log loss, Brier, MAE, coverage vs bookmaker baseline
│       └── cli.py               # python -m gimme_predict run|backtest
├── packages/
│   └── db/                      # Drizzle schema, migrations, seed
├── .github/workflows/           # ci.yml · collect-daily.yml · predict-daily.yml · backtest-weekly.yml
├── assets/logo/                 # gaw-logo.svg — the logo; the header, the favicon and the app icons are all built from it
├── data/                        # local raw cache and model artifacts (git-ignored)
├── README.md
└── .gitignore

Local setup

Prerequisites: Node 20+ (24 recommended), pnpm, Python 3.12+, uv, PostgreSQL (a free Neon branch is the easiest).

git clone https://github.com/bjhr730/gimme-a-W.git
cd gimme-a-W

Web app and shared packages:

pnpm install
cp .env.example .env            # fill in DATABASE_URL
pnpm --filter @gimme/db migrate # apply Drizzle migrations
pnpm --filter web dev           # http://localhost:3000

Collectors and models:

cd collectors
uv sync
uv run gimme-collect run espn --kind all --league nfl --league eng.1 --date yesterday --days 3
uv run gimme-collect run espn --kind roster --league nfl
uv run gimme-collect derive --what form,elo
uv run pytest

--kind all collects scoreboards, teams, standings and game summaries (box scores, lineups, ESPN FPI picks). Rosters are separate because they cost one request per team.

Predictions (the daily workflow runs these after collecting):

uv run gimme-predict backtest --competition nfl     # season-by-season vs the closing line
uv run gimme-predict run --sport american_football  # writes predictions for the next 8 days
uv run gimme-predict run --sport soccer --dry-run   # needs ~30 finals per league to fit

History loads (one-off, then refreshed weekly by the workflow):

uv run gimme-collect run nflverse --seasons 2015-2026 --what games
uv run gimme-collect run nflverse --seasons 2024-2025 --what players
# soccer seasons from ESPN: one request per calendar month, results + shots + corners
uv run gimme-collect run espn --kind scoreboard --league eng.1 --date 2019-08-01 --days 2555
# closing odds from football-data.co.uk (when the site is up)
uv run gimme-collect run fdcouk --league eng.1 --league esp.1 --seasons 2015-2025

Environment variables

Variable	Used by	Required	Notes
`DATABASE_URL`	web, collectors, models	Yes	Neon pooled connection string
`DIRECT_URL`	Drizzle migrations	Yes	Neon direct (non-pooled) connection
`FOOTBALL_DATA_API_KEY`	collectors	Optional	football-data.org
`ODDS_API_KEY`	collectors	Recommended	the-odds-api.com; the models use lines as baseline
`COLLECTOR_CACHE_DIR`	collectors	No	Defaults to `./data/raw`
`MODEL_ARTIFACT_DIR`	models	No	Defaults to `./data/models`
`COLLECTOR_USER_AGENT`	collectors	No	Identify the bot politely, with a contact email

Never commit .env. Commit .env.example with empty values instead.

Roadmap

Phase	Deliverable
0 — Scaffold (done)	README, `.gitignore`, logo, monorepo skeleton, tooling
1 — Data spine (done)	Drizzle schema + migrations · ESPN adapter for soccer/NFL/CFB · CLI · CI and daily collect workflows
2 — Web v1 (done)	Mobile-first Next.js app · today's games per sport with a date strip · game pages with stats and lines · team and standings pages · search · PWA manifest and icons from the logo. Player pages wait for roster data in phase 3
3 — Depth (done, ESPN part)	ESPN game summaries (full team box scores, per-player passing/rushing/receiving and goals/shots, soccer lineups, FPI picks) · ESPN rosters · `team_form` and Elo `rating` derivations · player pages, box scores on game pages, form/Elo/roster on team pages · weekly roster workflow
3b — Historical depth (done)	nflverse: NFL games since 2015 with closing lines, weather and rest days, weekly player stats with EPA and target share · football-data.co.uk: league seasons with shots on target, corners and Bet365/Pinnacle opening and closing odds, matched to ESPN teams by name · Elo with season regression · weekly refresh workflow. CFBD when a key is added; Sports-Reference last, under its rate limit
4 — Game predictions (done)	`gimme_predict`: point-in-time features (Elo, rest, form), logistic win probability + ridge spread and total for football, Dixon-Coles Poisson for soccer (1X2, total goals, team goals, both teams to score), 60/40 blend with the market, season-by-season back-test vs the closing line stored in `model_run`, daily prediction run · "Who gets the W" panel on game pages, model pick chips on score cards, `/models` page
5 — Team and player markets (done)	Passing/rushing/receiving yards with middle-half ranges and anytime TD (usage share, opponent allowed, market game script; holdout back-test), shots on target and corners per team and per match (negative-binomial GLM with over/under lines), anytime scorer from expected-goals share · markets panel on game pages · daily run
Research search (done)	Header search that understands teams, matchups, players, leagues and days, with keyboard navigation; `/api/search`
7 — Availability and parlays (done)	ESPN injury feed (, in the daily run) into · prop models withhold players reported out and scale probability markets for doubts · availability shown on game, team and player pages and as badges on market rows · soccer goals, assists and shots on target per player with expected quantities, shrunk rates and squad-coverage dilution · page, 2 to 20 legs, one per game · mobile back button · daily job re-reads the last three weeks so a missed run cannot leave a gap in the season

Legal and etiquette

Respect each site's robots.txt and terms of service. Sports-Reference explicitly bans scrapers above ~20 requests/minute and restricts commercial use; ESPN's JSON API is unofficial and may change without notice.
Identify the collector with a descriptive User-Agent that includes a contact email.
Cache aggressively; never fetch the same page twice in a day.
Store facts (scores, stats, lines) and short summaries with attribution links. Do not republish article text.
This project is for personal/educational use. Check licensing before any commercial use of the data.
Predictions are statistical estimates for information only, not betting advice.

README 전문 보기 ↗

raw_page

win_probability

spread

published_at

--kind injuries

player_status

/parlay

Gimme a W

언어 구성

업데이트0건

댓글

스포츠·피트니스 분야의 다른 프로젝트

Gimme a W

언어 구성

업데이트0건

댓글

스포츠·피트니스 분야의 다른 프로젝트

이 프로젝트의 운영자인가요?

소개

README

gimme-a-W

Table of contents

Tech stack

Data sources

Primary sources

Structured sources the models depend on

Architecture

Prediction engine

Markets

Availability

How a prediction is produced

What "good" looks like

Domain model

Repository layout

Local setup

Environment variables

Roadmap

Legal and etiquette

README

README

gimme-a-W

Table of contents

Tech stack

Data sources

Primary sources

Structured sources the models depend on

Architecture

Prediction engine

Markets

Availability

How a prediction is produced

What "good" looks like

Domain model

Repository layout

Local setup

Environment variables

Roadmap

Legal and etiquette