MySports AI

Crawls and normalizes betting odds across multiple sportsbooks (DraftKings, FanDuel, BetMGM, etc.) in real-time, converting heterogeneous line formats into a unified data model for comparative analysis. Uses scheduled ETL pipelines to detect line movements, identify sharp vs soft books, and flag arbitrage opportunities. Normalizes American, decimal, and fractional odds into a canonical representation for downstream ML models.

Trains ensemble ML models (gradient boosting, neural networks, or hybrid approaches) on historical sports data (team stats, player metrics, weather, rest days, injury reports, public betting volume) to predict game outcomes and generate probability distributions. Models output point estimates with calibrated confidence intervals, allowing users to assess prediction uncertainty. Likely uses feature engineering pipelines to extract predictive signals from raw sports data and cross-validates on holdout test sets to estimate generalization performance.

Compares model-predicted probabilities against sportsbook implied probabilities (derived from odds) to identify bets where the model believes the line is mispriced. Generates ranked recommendations based on expected value (EV) calculations: EV = (model probability × potential payout) - (1 - model probability × stake). Filters recommendations by confidence threshold and minimum EV threshold to surface only high-conviction opportunities. May apply Kelly Criterion or fractional Kelly sizing to suggest bet amounts.

Monitors for triggering events (line movement exceeding threshold, new recommendation generated, odds at target level, injury report published) and delivers notifications via push, email, or SMS. Likely uses event-driven architecture with message queues (Kafka, RabbitMQ) to decouple alert generation from delivery. Allows users to configure alert preferences (sports, bet types, minimum EV threshold, notification channels) and quiet hours to avoid spam.

Logs all user bets (placed through the platform or manually logged) and tracks outcomes (win/loss/push) against predicted probabilities. Computes aggregate metrics: win rate, ROI, Sharpe ratio, maximum drawdown, and calibration curves (comparing predicted vs actual win rates across probability buckets). Generates performance dashboards and reports to help users assess whether recommendations are generating positive returns and whether model predictions are well-calibrated.

Implements a freemium business model with tiered access: free tier provides limited predictions and odds data (likely delayed or aggregated), while premium tier unlocks real-time alerts, specific pick recommendations, advanced analytics, and priority support. Uses feature flags and API rate limiting to enforce tier boundaries. Likely uses subscription management (Stripe, Paddle) to handle billing and tier upgrades.

Ingests injury reports, roster transactions, and player status updates from official sources (ESPN, NFL.com, NBA.com, etc.) and integrates them into the ML prediction pipeline as real-time features. Updates model inputs when key players are ruled out, downgraded, or return from injury. May use NLP to parse unstructured injury reports and extract player status (out, questionable, probable, day-to-day). Triggers re-prediction when material roster changes occur.

Maintains historical snapshots of odds across sportsbooks and computes line movement metrics: point spreads moved by X points, totals moved by Y points, moneyline odds shifted by Z percentage points. Identifies directional movement patterns (sharp money moving one direction, public money moving another) by correlating line movement with betting volume. Generates visualizations showing line history and movement velocity to help users understand betting pressure and identify late-breaking information.