apexquant
Paper analysis only
- Language
- Python
- Stars
- 0
- Last updated
From the README
Paper analysis only. No staking logic is connected to money. Analytical tool only; no outcome is guaranteed. Bet responsibly and only where legal.
Two apps live in this repository:
| App | Folder | What it is |
|---|---|---|
| ApexQuant v10.0 suite | backend/, frontend/, tests/ |
Football and tennis analysis suite from the master build prompt (reference/): sharp vs soft prices, de-vig, Dixon-Coles / Markov model, blend, EV, four veto rules, OCR screenshots, news, weather, Roman Urdu report, audit vault, Fluent-style UI. |
| Phase 1 forecast feed | webapp/, backtest/, docs/ |
The original Dixon-Coles backtest (gate failed, see backtest/RESULT.md) and the keyless GitHub Pages forecast feed. |
ApexQuant v10.0 suite
Install and run
pip install -r requirements.txt
copy .env.example .env # fill in keys you have; every key is optional
uvicorn backend.main:app --reload --port 8010
Open http://127.0.0.1:8010. API docs at http://127.0.0.1:8010/api/docs.
Data sources
| Source | Key | Used for | Cache |
|---|---|---|---|
| ESPN public scoreboard | none | fixtures, live scores, venue, records; DraftKings prices as a labelled fallback soft book | 15 min |
| The Odds API | ODDS_API_KEY |
Mode A: Pinnacle (sharp) and 1xBet (soft), 1X2 / spreads / totals; quota shown in the app bar | 20 min per league |
| Gemini Flash / Groq vision | GEMINI_API_KEY / GROQ_API_KEY |
Mode B screenshot OCR, strict "only what is visible" | none |
| Google News RSS + DuckDuckGo news | none | injuries, lineups, press quotes with source and timestamp | 30 min |
| Open-Meteo | none | geocoding + hourly forecast at the match hour | 1 h |
football-data.co.uk (via webapp/ratings.py) |
none | Dixon-Coles ratings for the five leagues | refit on demand |
Without any key the app still runs: fixtures, model, news and weather work; EV needs a Pinnacle price, which you enter by hand (Odds & Value tab) or through screenshots. Missing data is shown as "Not available" and lowers the confidence badge. Nothing is invented.
How a match is analysed
- Pinnacle's margin is removed with the power method (
sum (1/O)^k = 1), multiplicative as a cross-check. Both are shown. - Model: football
lambda/mufrom fitted ratings (league averages with a Low badge when a team is not in the data), Dixon-Colesrho = -0.13, 10x10 scoreline matrix, every market derived from it. Tennis: Barnett-Clarke Markov chain point -> game -> tiebreak -> set -> match from serve/return inputs, Elo, or (flagged Low) the sharp price. P_master = 0.70 x P_pinnacle + 0.30 x P_model(weights in Settings).EV% = (P_master x O_soft - 1) x 100, push-adjusted for quarter lines (half-win, half-loss, push handled exactly). Quarter-Kelly stake with a 2% cap.- Rules: VR-01 edge gate, VR-02 injury/steam trap, VR-03 vig and variance cap, VR-04 model vs market divergence. Verdict SNIPER / CAUTION / VETO with an EC tag.
- Roman Urdu report assembled from fetched facts only, each sentence citing its section. No LLM is involved in any number or verdict.
- "Copy Prompt to Verify in Claude" produces a self-contained audit prompt with all data.
Pages
Overview (KPIs + Match Radar), Football, Tennis, Match detail (Summary, Odds & Value, Model, Ground Reality, Risk Audit, Raw Data), Screenshot Analyzer, Audit Vault (results, ROI, CLV, CSV export), Settings, Glossary & Help. Every metric has a (?) callout with its formula and a worked example from the live numbers.
Files
backend/config.py settings, thresholds, key status (keys never leave the process)
backend/schemas.py Pydantic v2 models
backend/data_fetcher.py ESPN, Odds API, OCR, news, weather, SQLite cache
backend/quant_engine.py all maths, standard library only, pure functions
backend/analysis.py one match end to end, Roman Urdu report, verification prompt
backend/auditor.py VR-01..VR-04, EC tags, Memory Vault CSV, ROI / CLV
backend/main.py FastAPI endpoints, serves frontend/
frontend/ index.html, styles.css (design tokens), app.js + js/ modules
tests/ test_quant_engine.py, test_auditor.py, test_api.py (offline)
vault/ audit_memory_vault.csv, api_cache.sqlite3, settings.json (git-ignored)
reference/ the build prompt PDF and the two UI references
Endpoints
GET /api/health, GET|PATCH /api/settings, GET /api/fixtures?sport&date&league&radar,
GET /api/odds/{match_id}?sport, POST /api/ocr (multipart soft, sharp),
POST /api/analyze, GET|POST /api/vault, PATCH /api/vault/{id}, GET /api/vault/export,
POST /api/refit.
Tests
python -m pytest -q test_quant_engine.py backtest webapp tests
Phase 1 forecast feed (unchanged)
uvicorn webapp.server:app --reload --port 8000
webapp/build_static.py writes the GitHub Pages site into docs/;
.github/workflows/feed.yml refits every 6 hours. Pages setup: Settings, Pages, Source
"Deploy from a branch", branch main, folder /docs. The backtest result stands: the
Dixon-Coles model alone does not beat the Pinnacle closing line (backtest/RESULT.md).
That is why the v10.0 suite blends it at 30% behind the sharp price instead of using it alone.
Environment variables
| Variable | Default | Meaning |
|---|---|---|
ODDS_API_KEY |
unset | The Odds API (Mode A prices, both apps) |
GEMINI_API_KEY, GEMINI_MODEL |
unset, gemini-2.0-flash |
OCR |
GROQ_API_KEY, GROQ_VISION_MODEL |
unset, meta-llama/llama-4-scout-17b-16e-instruct |
OCR fallback |
SHEETS_WEBHOOK_URL |
unset | optional Google Sheets mirror of saved audits |
ODDS_REGION |
eu |
Odds API region |
APEX_BANKROLL, APEX_KELLY_FRACTION, APEX_KELLY_CAP, APEX_BLEND_SHARP |
1000, 0.25, 0.02, 0.70 | paper sizing defaults |
APEX_DATA_DIR, APEX_VAULT_DIR, APEX_RATINGS_CACHE |
./data, ./vault, webapp/cache/ratings.json |
paths |
APEX_AUTO_REFRESH, APEX_REFRESH_HOURS |
1, 6 |
Phase 1 feed only |