Scans thousands of Kalshi markets, cross-references 12 sportsbooks and free data APIs, prices fair value with a normal-CDF model, and executes Kelly-sized limit orders — every bet cleared through 13 risk gates and logged with fill-accurate accounting for closing-line-value tracking.
| Category | Coverage | Data sources |
|---|---|---|
| 🏟️ Sports betting | NFL, NBA, MLB, NHL, NCAAB/NCAAF, UFC, soccer/MLS, F1, NASCAR, PGA, IPL, esports — 27 filters across 18 sports | The Odds API (12 sportsbooks), ESPN, NHL/MLB Stats, NWS |
| 🏆 Championship futures | Super Bowl, NBA Finals, Stanley Cup, World Series, PGA Tour — priced with N-way de-vig | Sportsbook outright odds |
| 📈 Prediction markets | Crypto (BTC, ETH, XRP, SOL, DOGE), S&P 500 + VIX, weather (13 cities), politics | CoinGecko, Yahoo Finance, NWS |
graph LR
A["12 Sportsbooks<br><sub>The Odds API</sub>"] --> D["Normal CDF<br>Fair Value"]
B["Team Stats<br><sub>ESPN / NHL / MLB</sub>"] --> D
C["Signals<br><sub>Weather, Pitchers, Rest, Sharp $</sub>"] --> D
D -->|"compare"| E["Kalshi Price"]
E -->|"Edge >= 3%"| F["Composite Score<br><sub>0-10 scale</sub>"]
F --> G["13 Risk Gates"]
G --> H["Kelly Sizing"]
H --> I["Limit Order + Log"]
style D fill:#8B5CF6,color:#fff,stroke:none
style G fill:#e74c3c,color:#fff,stroke:none
style I fill:#2ea44f,color:#fff,stroke:none
| Signal | Source |
|---|---|
| Normal CDF Model | Sport-specific stdev bell curve probabilities |
| Sharp Book Weighting | Pinnacle 3x, Circa 3x, DraftKings 0.7x |
| Team Stats | ESPN/NHL/MLB win% validates fair value |
| Sharp Money | Open-vs-close odds detect reverse line movement |
| Weather | NWS forecasts for 61 NFL/MLB outdoor venues |
| Pitcher Matchups | ERA, FIP, WHIP, K/9, rest days from MLB Stats API |
| Rest Days | NBA/NHL back-to-back fatigue detection |
| Book Disagreement | >4pt spread range flags injury news |
Important
Every scan defaults to preview mode. No money is risked until you pass --execute. Each scan row shows a Gate column (R18) that previews whether it will pass the static risk gates — ok if all clear, or a short label (score, conf, no-fav, pred-off, etc.) for the failing gate.
Every order must clear gates 1-7 (including 3.5, 4.5, 4.6, 4.7). Gates 8-9 cap sizing instead of rejecting.
| # | Gate | Action |
|---|---|---|
| 1 | Daily loss limit | Reject at -$250 |
| 2 | Position count | Reject at 50 open |
| 3 | Edge threshold | Reject below floor (3% global; 4% MLB/NBA/NCAAB) |
| 3.5 | Market price floor | Reject bets priced below MIN_MARKET_PRICE (default $0.06 — lottery-ticket filter) |
| 4 | Composite score | Reject below 6.0/10 |
| 4.5 | Min confidence | Reject below MIN_CONFIDENCE (default medium) |
| 4.6 | NO-side favorite | Reject NO bets <25¢ unless edge ≥25% AND confidence=high |
| 4.7 | Prediction-market safety | Reject crypto/weather/spx/mentions/companies/politics unless ALLOW_PREDICTION_BETS=true (R25) |
| 5 | Duplicate check | Reject same market |
| 6 | Per-event cap | Reject at 2/game |
| 7 | Series dedup | Reject same matchup bet within window (MLB/NHL 72h, others 48h; per-sport via SERIES_DEDUP_HOURS_<SPORT>) |
| 8 | Bet size cap | Cap at $100 |
| 9 | Bet ratio cap | Cap at 3x batch median |
All limits configurable via .env. See Architecture for the full pipeline.
Why these gates exist — calibration history behind each rule
- Gate 3.5 —
MIN_MARKET_PRICE(R7, 2026-04-22): sub-10¢ bets ran 1W-3L while the model claimed "+50% edge" on 8-10¢ longshots. - Gates 4.5 / 4.6 —
MIN_CONFIDENCE,NO_SIDE_*(2026-04-21): low-confidence bets hit -105% ROI and all 13 high-edge losers were NO-side on heavy favorites. NO bets belowNO_SIDE_KELLY_PRICE_FLOOR(35¢) are additionally sized at half-Kelly. - Per-sport edge floors → 0.04 (2026-06-14): MLB/NBA/NCAAB ran a higher floor (NBA Brier 0.3306 was worst-of-sport) to offset a model that over-claimed ~15% edge. The 06-03/06-05 edge-matching fixes removed that over-claim at the source, so the floor was lowered 0.06 → 0.04 to stop double-correcting and re-admit honest 3-6% edges.
- One-way confidence bumps (R13, 2026-04-24): team-stats, rest/B2B, and sharp-money signals can drop a tier but no longer raise one — upward bumps tracked inflated claimed edge, not better outcomes.
- Gate 4.7 —
ALLOW_PREDICTION_BETS(R25, 2026-04-24): an audit found all 6 prediction modules (crypto/weather/spx/mentions/companies/politics) cached stale data with no TTL and produced nonsense fair values; blocked by default until rebuilt. - Cross-category dedup (R8, 2026-04-29): opt-in
CROSS_CATEGORY_DEDUP_<SPORT>=truecollapses ML+Total+Spread on the same game to the highest-composite row before gating. Default off — cross-category correlation varies by sport.
Bet size scales with edge, divided by batch count to control total exposure. Edge is soft-capped above 15% before sizing (trusted_edge()) to damp Kelly on likely-overstated signals — raw edge remains in gates and reports.
bet = max(unit, (kelly_frac / batch) * trusted_edge(edge) * bankroll)
| Edge | Trusted | 1 bet | 5 bets | 10 bets |
|---|---|---|---|---|
| 3% | 3% | $0.75 | $0.15 | $0.08 |
| 10% | 10% | $2.50 | $0.50 | $0.25 |
| 15% | 15% | $3.75 | $0.75 | $0.38 |
| 25% | 20% | $5.00 | $1.00 | $0.50 |
| 35% | 25% | $6.25 | $1.25 | $0.63 |
Example: $50 bankroll, KELLY_FRACTION=0.50. Capped by max bet ($100) and balance. Soft-cap: KELLY_EDGE_CAP=0.15, KELLY_EDGE_DECAY=0.5.
# 0. Clone repo and enter project
git clone https://github.com/michaelschecht/Edge-Radar.git
cd Edge-Radar
# 1. Create + activate virtual environment
python -m venv .venv
# macOS/Linux (bash/zsh):
source .venv/bin/activate
# Windows PowerShell:
# .venv\Scripts\Activate.ps1
# 2. Install dependencies and create env file
python -m pip install --upgrade pip
pip install -r requirements.txt
cp .env.example .env
# 3. Verify environment (API keys, dependencies)
python scripts/doctor.py
# 4. Preview opportunities (no money risked)
python scripts/scan.py sports --filter nba
# 5. Execute with risk controls
python scripts/scan.py sports --filter nba --execute --unit-size 1 --max-bets 5
# 6. Settle bets and view P&L
python scripts/kalshi/kalshi_settler.py report --detail --saveTip
All scanners share the same flags: --execute, --unit-size, --max-bets, --pick, --ticker, --save, --date, --exclude-open. Use --date tomorrow --exclude-open to avoid double-betting.
| Guide | What it covers |
|---|---|
| Setup Guide | First-time install, API keys + RSA private key generation, .env wiring, safe rollout plan (dry-run → low-stakes → normal), automation, ongoing monitoring, troubleshooting |
Expand any section for copy-paste CLI examples by workflow.
Sports Betting
python scripts/scan.py sports --filter nhl
python scripts/scan.py sports --filter mlb --execute --unit-size 1 --max-bets 10
python scripts/scan.py sports --filter mlb --date tomorrow --exclude-open
python scripts/scan.py sports --filter nba --saveChampionship Futures
python scripts/scan.py futures --filter nba-futures
python scripts/scan.py futures --filter mlb-futures --execute --unit-size 2 --max-bets 5
python scripts/scan.py futures --filter nba-futures --savePrediction Markets
python scripts/scan.py prediction --filter crypto
python scripts/scan.py prediction --filter weather
python scripts/scan.py prediction --filter crypto --execute --unit-size 1 --max-bets 5Portfolio & Settlement
python scripts/kalshi/kalshi_executor.py status --save
python scripts/kalshi/risk_check.py --report positions --save
python scripts/kalshi/kalshi_settler.py settle
python scripts/kalshi/kalshi_settler.py report --detail --saveBacktesting
python scripts/backtest/backtester.py
python scripts/backtest/backtester.py --simulate --save
python scripts/backtest/backtester.py --sport mlb --confidence high --min-edge 0.10Analyzes settled trades for win rate, ROI, profit factor, Sharpe ratio, equity curves, max drawdown, and calibration — broken down by sport, category, confidence level, and edge bucket. The --simulate flag runs what-if scenarios across edge thresholds, confidence tiers, and categories; --save exports reports.
Edge-Radar ships with two slash commands for Claude Code:
| Skill | Definition | Description |
|---|---|---|
/edge-radar |
skills/edge-radar/SKILL.md |
Unified command center — scan, bet, status, settle, risk, detail, backtest across Kalshi sports, futures, and prediction markets. |
/edge-radar-analysis |
skills/edge-radar-analysis/SKILL.md |
Post-hoc performance report — trade ledger + slices by sport, category, side, edge bucket, confidence, price, calibration, longshots, streaks, daily P&L. |
The skill source of truth lives in
skills/. Claude Code loads them from.claude/skills/, which on Windows are directory junctions toskills/(git-ignored, sincecore.symlinks=false). After a fresh clone, recreate the junctions once:pwsh -File scripts/setup/link_skills.ps1.
/edge-radar status # Balance, positions, P&L
/edge-radar scan nba # Preview NBA opportunities
/edge-radar bet mlb --unit-size 1 # Scan + execute on confirm
/edge-radar settle # Settle + P&L report
/edge-radar-analysis 30 --save # 30-day performance report to reports/Performance/
Routes natural language to the correct scanner, enforces all risk gates, always previews before executing. All CLI flags work inline.
Note
Requires Claude Code CLI, Desktop, or IDE extension.
Gemini CLI / OpenAI Codex — add the skill content to your GEMINI.md or AGENTS.md for equivalent functionality.
Pre-built scripts scan all sports, rank by composite score, and execute with Kelly sizing. See the Automation Guide.
# Install all scheduled tasks at once
python scripts/schedulers/automation/install_windows_task.py install all| Task | Schedule | Description |
|---|---|---|
daily-summary |
4:50 AM PT | Morning P&L digest — yesterday settled + open exposure + today pending + 7d context. Emailed at 5:00 AM PT (U2, 2026-04-30) |
scan |
8:00 AM ET | Preview scan — saves report, no bets |
execute |
8:00 AM ET | Scan + execute — places live orders |
settle |
11:00 PM ET | Settle bets, update P&L |
next-day |
9:00 PM ET | Scan + execute tomorrow's games |
calibration |
2:00 AM, 1st of month | 30-day calibration report — Brier, calibration curve, prescriptive recommendations |
Reports save to reports/Sports/schedulers/ with full execution details.
Want the complete pipeline — emails, midday/late runs, weekly calibration/backtest/analysis — beyond the installer's core tasks? See Task-Schedule Reference & Setup for the full 17-task roster with copy-paste .bat/.sh templates and schtasks registration.
| Guide | Description |
|---|---|
| Setup Guide | Install, API keys, .env, safe rollout, automation, and monitoring — the single end-to-end operator guide |
| Automation Guide | Windows Task Scheduler for daily betting — one-command installer for the core tasks |
| Task-Schedule Reference | Full 17-task pipeline roster with .bat/.sh templates and schtasks setup |
| Scripts Reference | Every script, flag, and example |
| Sports Guide | 27 filters, edge detection, daily workflow |
| Futures Guide | NFL, NBA, NHL, MLB, golf championships |
| Prediction Markets | Crypto, weather, S&P 500, politics |
| Architecture | Pipeline, edge models, risk gates, data flow, and project structure |
| MLB Filtering | 10 filter categories for MLB picks |
| Roadmap | All enhancements — completed & pending |
| Changelog | Full project history |
All external data is free. Only Kalshi requires a funded account.
| API | Purpose |
|---|---|
| Kalshi | Market data + order execution (API key + RSA signing) |
| The Odds API | 12 US sportsbook odds (500 free req/mo) |
| ESPN | NBA, NFL, NCAAB, NCAAF standings + line movement |
| NHL Stats API | Standings, goal differential, last 10 record |
| MLB Stats API | Standings, run differential, pitcher stats |
| NWS | Hourly forecasts for 61 NFL/MLB outdoor venues |
| CoinGecko | Crypto prices + 24h volatility |
| Yahoo Finance | S&P 500 + VIX implied volatility |
Setup • Architecture • Scripts • Changelog
Built with Python, scipy, and too many API calls — Back to top