Every forecasting model I run on one board: whichever sport is in season sits on top, with its slate, its track record, and the rest waiting their turn.
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.
8h +2★ 644Sports & fitness
Do you operate this project?
Owners of public repositories under a personal GitHub account can sign in to verify ownership and manage the introduction themselves. No administrator approval is needed.
Organization repository · verify with a repository file
Sign in, generate a verification file, and commit it to this repository’s default branch within 30 minutes. Files on other branches or forks do not count. Then check the file below to manage the project without administrator approval. Keep the file to retain your access.
If needed, you will go to sign-in. After signing in, return here and press your chosen verification button again.
Run /nomorevibe verify in your project folder to verify ownership and manage the introduction yourself.
Loading the full README. The saved excerpt is shown below.
Probably not — but let's see. A Python-based sports betting tracker and modeling
system that snapshots bookmaker odds twice daily, tracks line movement across
sportsbooks, and trains models to pick winners (and covers) against the closing
line.
The season is graded and the books are in
(NFL/inventory/INVENTORY_2025.md,
NFL/Weeks.md). Over 4,350 out-of-sample games from 2010-2025,
walking the model forward one retrain per season:
Probably not — but let's see. A Python-based sports betting tracker and modeling
system that snapshots bookmaker odds twice daily, tracks line movement across
sportsbooks, and trains models to pick winners (and covers) against the closing
line.
The 2025 answer: no
The season is graded and the books are in
(NFL/inventory/INVENTORY_2025.md,
NFL/Weeks.md). Over 4,350 out-of-sample games from 2010-2025,
walking the model forward one retrain per season:
model
closing line
straight up, 2010-2025
64.3%
66.4%
straight up, 2025 only
65.5%
65.8%
ATS, 2010-2025
50.5%
— (52.4% needed)
totals, 2010-2025
49.5%
— (52.4% needed)
Blending the model into the market makes the market worse out of sample
(log loss 0.613 vs 0.610), which is the cleanest statement of the result: after
the closing line has spoken, the model adds nothing. Flat-stake ROI is -4.9% on
spreads and -4.5% on totals, both about what the vig costs.
That is the honest baseline every future change here gets measured against.
Details and the two places an edge might actually live are in
NFL/model/v2/README.md.
Features
What each live model actually looks at, per sport:docs/MODEL_FEATURES.md — the quick sheet
(home Elo, away Elo and the sport-specific features, with the tuned
parameters and which features are currently dormant).
docs/MODEL_SOURCES.md — the same features
with the upstream each one is pulled from, its measured worth, and the
training window per model. Plus what an NBA board would need.
Multi-sport odds ingestion — The Odds API (free tier) for NFL and NBA
moneylines, spreads, and totals across major US sportsbooks; World Cup planned
Bookmaker tracking — per-book odds history, sportsbook comparison charts,
bookmaker variance and accuracy analysis (which bookies move first, which are sharpest)
Line movement — opening vs. closing spread aggregation from timestamped
odds snapshots (data/odds_api_data_*.csv)
NFL LightGBM models — win / ATS / total targets on an Elo + form + market
feature set, walk-forward validated one retrain per season, with a flat-stake
betting backtest (NFL/model/v2/; the older single-split v1 is kept in
NFL/model/ for provenance)
Season inventory — end-of-season audit of what was captured, what was
missed, and how the closing line itself performed (NFL/inventory/)
Soccer / World Cup model — custom international Elo (fresh 2006 start,
tiered K-factors, friendlies barely weighted) plus a multinomial outcome
model with host effects and FIFA-rating squad-strength hooks; predicts the
2026 World Cup slate (soccer/, spec in soccer/SPEC.md)
European club Elo (top-5 leagues + their second divisions) — one Elo
pool per country spanning EPL/Championship, Bundesliga/2. Bundesliga,
La Liga/Segunda, Serie A/Serie B and Ligue 1/Ligue 2, so promoted clubs
carry real second-division form (through a tuned winner's-curse blend)
instead of a flat prior; glued cross-country by Champions/Europa/
Conference League results, with Transfermarkt spend features and a
pooled W/D/L outcome model validated on a two-season holdout; a daily
runner predicts each day's slate across both tiers, grades a running
ledger, and Monte Carlos every top-flight table
(soccer/clubs/, spec in soccer/clubs/SPEC.md)
MLB daily model — betting-blind Elo (K=3, +24 home, MOV-weighted)
tested live against the 2026 season: three emails every morning (futures
Monte Carlo, today's slate with simulated scores, yesterday's graded
report card) plus the model card at whosyurgoat.app/vegas/mlb
(mlb/daily/, writeup in reports/mlb_elo_stat_associations.md)
College football daily model — betting-blind FBS Elo over every
FBS-involved game since 2001 (conference-aware season regression so
realignment is handled by construction, one pooled FCS opponent, capped
margin of victory; tuned on 2005-23, 2024-25 holdout log-loss 0.499), a
daily runner that grades the slate against a paired always-pick-home
baseline, predicts the next two days with expected scores, and Monte
Carlos expected wins / bowl / undefeated / conference-title odds; model
card at and a Mon/Thu update email
(, writeup in )
Local Setup
1. Clone the repo
git clone https://github.com/Hijodeagua/Can-Tre-Beat-Vegas.git
cd Can-Tre-Beat-Vegas
2. Install dependencies
pip install -r requirements.txt
Python 3.11+ recommended.
3. Configure
Set your Odds API key (free tier: 500 requests/month at
the-odds-api.com):
export ODDS_API_KEY=your_key_here
No key is needed to work with the committed odds snapshots in data/ —
the models and reports run entirely off those CSVs.
4. Fetch odds
# Snapshot current NFL + NBA odds (h2h, spreads, totals)
python -m data_jobs.odds_api.fetch_odds --sport all
# Just one sport
python -m data_jobs.odds_api.fetch_odds --sport nfl
5. Train the NFL models
# from the repo root
python3 NFL/model/schedule.py --refresh # optional: pull latest nflverse schedule
python3 -m NFL.model.v2.train --target all --save # win + ATS + total, walk-forward
Saved models, per-season metrics, backtests, feature importances, and the full
out-of-sample prediction files land in NFL/model/v2/artifacts/.
A deployable Next.js 14 (App Router) static site — the next-48-hours
homepage. For every upcoming NFL and NBA game it shows the consensus line
(moneyline, spread, total), no-vig market win probability, line movement since
open, model picks with edge-vs-market when available, and an expandable
book-by-book odds table. It reads the static JSON in web/public/data/,
produced by the Python pipeline; it runs no Python server itself.
# 1. Regenerate the static JSON from the committed odds snapshots
python -m data_jobs.export_web_json
# 2. Build / run the front end
cd web
npm install
npm run dev # http://localhost:3000/vegas
npm run build # production build (static, prerendered)
The site is served under the /vegas base path (see web/next.config.mjs)
so it can be proxied at whosyurgoat.app/vegas by the
hub. The unified odds
workflow re-exports the JSON after every odds fetch, so a Vercel project
pointed at web/ redeploys with fresh lines 2x daily.
Automated refresh
GitHub Actions keep the data flowing without manual pulls:
Workflow
Schedule
What it does
unified-odds.yml
2x daily (10:00 / 22:00 UTC)
Fetches NFL + NBA odds snapshots, re-exports web JSON, commits
daily-report.yml
daily (10:00 UTC)
MLB daily pipeline: pulls new box scores, updates Elo, sends the futures / slate / grade emails, commits site data (see mlb/daily/README.md)
soccer-daily.yml
daily (10:00 UTC)
Club-soccer daily pipeline: refreshes top-5 league + UEFA results, regrades the ledger, predicts the slate, Monte Carlos the tables, commits site data (see soccer/clubs/daily/README.md)
cfb-daily.yml
daily (10:00 UTC)
College-football daily pipeline: refreshes the season from cfbfastR-data, replays the FBS Elo, grades the ledger, predicts the slate, Monte Carlos the regular season, commits site data; Mon/Thu update email (see CFB/daily/README.md)
nfl-daily.yml
daily (10:00 UTC)
NFL Elo daily pipeline: refreshes the nflverse schedule, replays the betting-blind Elo, grades the ledger, predicts the next week, Monte Carlos the season and the playoff bracket, commits site data; Tue/Thu update email (see NFL/daily/README.md)
The 2x-daily cadence is deliberate — it captures a morning line and an
evening line before games while conserving the free-tier API quota.
Project Structure
Can-Tre-Beat-Vegas/
├── data_jobs/ # Data ingestion + reporting
│ ├── odds_api/ # The Odds API client
│ │ ├── client.py # HTTP client + quota tracking
│ │ ├── config.py # Sports config, team metadata, stadium coords
│ │ ├── fetch_odds.py # CLI entry point (--sport nfl|nba|all)
│ │ └── processors.py # Raw API response → tidy CSV
│ ├── reports/ # Daily report generators
│ └── export_web_json.py # Next-48-hours slate → web/public/data/
├── NFL/
│ ├── elo/ # Betting-blind NFL Elo engine, tuner, tuned_params.json
│ ├── daily/ # Daily slate/grade/futures runner + email (see daily/README.md)
│ ├── model/ # v1 (superseded, kept for provenance)
│ │ ├── features.py # Rolling feature engineering (79 features)
│ │ ├── schedule.py # nflverse schedule loader
│ │ ├── line_movement.py # Open vs close aggregator (shared with v2)
│ │ ├── train.py # Single temporal split + LightGBM
│ │ └── v2/ # Current model
│ │ ├── elo.py # MOV-adjusted Elo, walk-forward
│ │ ├── dataset.py # 45 features off the nflverse schedule
│ │ ├── train.py # Season-by-season walk-forward + backtest
│ │ ├── predict.py # Score upcoming games → data/predictions/
│ │ └── artifacts/ # Models, metrics, backtests, OOS predictions
│ ├── inventory/ # End-of-season audit
│ │ ├── audit.py # Capture coverage + Vegas' report card
│ │ ├── grade_season.py # Writes the graded weekly ledger
│ │ └── INVENTORY_*.md # Generated season inventory
│ ├── Week_1/ ... Week_22/ # Weekly picks vs Vegas, graded (out of sample)
│ └── Weeks.md # Season-long results tally
├── soccer/ # World Cup / international soccer model
│ ├── SPEC.md # Model spec (Elo + squad-strength adjustments)
│ ├── data/ # International results 1872–present + fixtures
│ ├── model/ # Elo engine, training, fixture predictions
│ └── clubs/ # Top-5 European league club Elo (EPL, Bundesliga,
│ │ # La Liga, Serie A, Ligue 1) — see clubs/SPEC.md
│ ├── data/ # League + UEFA results, transfers, fixtures
│ ├── model/ # Per-league Elo, UEFA glue, outcome model
│ └── daily/ # Daily slate/grade/futures runner
├── CFB/ # College football — FBS Elo + daily pipeline
│ ├── data/ # cfbfastR-data fetcher, team aliases
│ ├── model/ # CfbEloEngine, tuner, tuned_params.json
│ └── daily/ # Daily slate/grade/futures runner + email
├── data/ # Odds snapshots + stats
│ ├── odds_api_data_*.csv # NFL odds snapshots (timestamped)
│ ├── nba/ # NBA odds snapshots + actual game results
│ ├── schedules/ # Cached nflverse games.csv
│ ├── models/ # Pickled model checkpoints
│ └── predictions/ # Model prediction outputs
├── web/ # Next.js front end (basePath /vegas)
│ ├── app/ # Slate homepage + methodology
│ └── public/data/ # Static JSON written by export_web_json.py
├── reports/ # Generated daily HTML reports + charts
└── .github/workflows/ # Automated odds fetch + report generation
Roadmap
The end state is a homepage for the next 48 hours of games — every NFL,
NBA, and World Cup matchup coming up in the next two days, side by side with:
The bookies' view — current odds per sportsbook, consensus line, and
how the line has moved since open
My models' view — win probability and ATS pick, with disagreement vs.
the market highlighted (that's where the edges live)
Track record — running tally of model vs. Vegas, by sport and bet type
Steps to get there:
Static JSON export of the upcoming-48-hours slate (odds + model picks),
same pattern as the election tracker
— data_jobs/export_web_json.py
Next.js front end reading that JSON (web/, served at /vegas)
Deploy web/ to Vercel and point the hub's /vegas rewrite at it
Soccer Elo + outcome model predicting 2026 World Cup fixtures
(soccer/, see soccer/SPEC.md for the full roadmap)
Club Elo for the top-5 European leagues, tuned per league and validated
against a two-season holdout (soccer/clubs/)
College football Elo + daily pipeline (CFB/), the third daily
model alongside MLB and club soccer
NFL Elo + daily pipeline (NFL/elo/, NFL/daily/), the fourth daily
model — betting-blind, so its line can be held up against the books
NCAAF on the odds feed (americanfootball_ncaaf in
data_jobs/odds_api/config.py) so the college model has a market to
lose to — ~180 credits/month in season on top of NFL's ~180, so it
needs the NBA gate or a one-snapshot-a-day cadence first (see
CFB/DATA_PULL_PLAN.md §7)
Odds API soccer keys (soccer_epl, …) and club model picks on the
slate (quota permitting)
World Cup odds ingestion (add soccer_fifa_world_cup to the Odds API
config) and soccer model picks on the slate
Wire NBA model predictions into the slate for current games
(the export already joins data/predictions/*.csv when dates match)
NFL model predictions for upcoming games —
python3 -m NFL.model.v2.predict --season 2026 --write
Point at the v2 prediction files
What the 2025 inventory says to fix next
Ranked by how much they'd move the answer, not by effort:
Capture the opener. 2025 capture began a median of 11 days before
kickoff, and weeks 1-4, 17 and 21 were missed entirely (83 of 285 games have
no odds record). Beating the closing line is close to impossible; beating a
stale opener is the winnable version of this bet, and we can't test it
without openers.
Store the spread number, not just the juice. The legacy schema kept
-110 / -110 but never -3.5, so the entire 2025 season has moneyline-only
line movement. The current schema fixed this; don't regress it.
Log best-available price per book, not the average. The snapshots
already carry 8-10 books. Line shopping is a mechanical edge that doesn't
require the model to be right about anything, and nothing in the repo
measures it yet.
Wire line_movement.py output in as a model feature (now that it
actually returns 2025 games)
NFL Elo daily model — the betting-blind sibling of the LightGBM model:
a franchise-continuous Elo over every game since 1999 (bye-week rest
edge, capped margin of victory, 40% off-season regression; tuned on
2005-23, 2024-25 holdout log-loss 0.624 vs 0.691 for always-pick-home), a
daily runner that grades the ledger against a paired always-home
baseline, predicts the next week with the model's own line and expected
scores, and Monte Carlos the rest of the season through the full playoff
bracket (division / playoff / #1 seed / conference / Super Bowl odds);
model card at whosyurgoat.app/vegas/nfl and a Tue/Thu update email
(NFL/elo/, NFL/daily/, writeup in NFL/elo/README.md)
How the models work — whosyurgoat.app/vegas/models walks every model
from its data source to its graded pick, each step linking to the file that
does it (web/app/lib/modelDocs.ts); the club-soccer sides-vs-differences
study (four outcome models on this year's matches, with KPIs and feature
importances) is at reports/soccer/sides_vs_diffs_2026.html
(soccer/clubs/model/report_sides_2026.py)
Email archive — every report email each pipeline sends is archived as
sent under web/public/emails/<league>/ with a permalink footer (this
email, the previous one of its kind, the league's archive), indexed in
web/public/emails/index.json and browsable by league at
whosyurgoat.app/vegas/emails (data_jobs/email_archive.py); the weekly
models-check email draws its permutation / SHAP importances as email-safe
bar charts (data_jobs/reports/email_charts.py)
Daily HTML reports — automated odds breakdowns, bookmaker performance,
and team odds-history charts (reports/)
Weekly NFL picks — model picks vs. Vegas, graded week by week
(NFL/Week_*/, running tally in NFL/Weeks.md)
Automated refresh — GitHub Actions fetch odds 2x daily and generate
reports, tuned to stay inside the free-tier API quota (~60 requests/month)
