MAX EV Sports

Real-time sports analytics platform for identifying and tracking value across major US sportsbooks. Built solo as a first major Python/ML project.

What It Does

Monitors six sharp sportsbooks in real-time, runs ML ensemble predictions across four sports, and surfaces edge opportunities based on line movement and model output. Originally included an automated trading layer for Kalshi prediction markets — that module is archived as Kalshi/prediction markets changed the economics of the approach mid-build.

Stack

Backend: Python 3, FastAPI, SQLite
Frontend: React 18, TypeScript, Vite, Tailwind CSS, Recharts, Electron
ML: XGBoost, LightGBM, Random Forest (ensemble)
Infrastructure: Linux VPS, systemd services, Nginx, cron

Architecture

React/Electron Frontend (TypeScript)
        ↓
FastAPI Backend (Python, port 8000)
        ↓
Real-Time Services (systemd, always-on)
  ├── Sharp Totals Logger     — polls 6 books every 10 seconds
  ├── Sharp Value Trader      — scans for mispriced lines hourly
  ├── ML Predictions API      — serves model output via FastAPI
  └── Kalshi Bot (archived)  — automated trade execution
        ↓
Scheduled Tasks (cron)
  ├── ML Predictions          — daily 6 AM regeneration
  ├── Strategy Alerts         — situational edge detection (daily 10 AM)
  ├── Game Results Scraper    — daily outcome collection
  └── Model Grading           — accuracy and ROI calculation post-game
        ↓
SQLite Databases
  ├── predictions.db          — ML model output and accuracy tracking
  ├── sharp_totals_history.db — historical book line movement
  ├── sharp_value_trades.db   — value trade log
  ├── positions.db            — open/settled position tracking
  ├── strategy_alerts.db      — situational opportunity log
  └── odds_history.db         — full historical odds archive

Data Sources

Source	Purpose	Volume
The Odds API	Live odds from 6 sharp books	~26k calls/day
Kalshi API	Prediction market data + trade execution (archived)	Real-time
BallDontLie API	NBA player statistics for prop modeling	Daily
ESPN / CBS RSS	Sports news, injuries, briefings	Every 4 hours

Sharp books tracked: Pinnacle, Circa, BetOnline, Bookmaker, DraftKings, Bovada

Sports covered: NBA, NHL, NCAAB, NFL

ML Pipeline

Three models run in ensemble for each prediction type:

• XGBoost — gradient boosting, primary model for totals and spreads
• LightGBM — fast boosting alternative, used for player props
• Random Forest — ensemble baseline, improves robustness on small samples

Features: historical game stats, rest days, home/away, injury context, line movement velocity, revenge/letdown situational flags

Output per prediction: probability vs market line, edge percentage, Kelly criterion bet sizing, confidence score

Grading: outcome comparison runs nightly, accuracy and ROI tracked by sport, model, and bet type

Frontend Pages

Page	Description
Home	Daily briefing — news, injuries, trending stories
Markets	Live value opportunities with ML prediction overlay
Positions	Full trade history with P&L
Analytics	Win rate, ROI, model performance over time
Props	Player prop predictions (partial)
Tools	Kelly calculator, edge analysis utilities

What I'd Do Differently

This was my first large Python project. A few things I'd change with what I know now:

• PostgreSQL instead of SQLite — multiple always-on services writing to the same SQLite files created contention. PostgreSQL with connection pooling is the right call at this data volume.
• Separate the ML training pipeline — training and serving are mixed in the same codebase. A proper feature store and scheduled retraining job would be cleaner.
• Start with one sport — building NBA, NHL, NCAAB, and NFL simultaneously from day one spread the model training data too thin early on. Better to nail one sport's prediction pipeline before expanding.
• The Kalshi layer — prediction markets emerged mid-build and changed the arbitrage economics significantly. If restarting today I'd build the analytics layer only and treat execution as a separate, later problem.

Status

Archived. The core analytics pipeline (line movement tracking, ML predictions, strategy alerts) works. The Kalshi automated trading module is preserved but not active. This repo is kept as a portfolio reference for the data pipeline and ML architecture.