NBA Spread Prediction Model

Project Overview

This project is an end-to-end NBA betting decision-support system. It fuses quantitative modeling with a large-language-model synthesis layer: statistical models compute projected margins and cover probabilities, and GPT-4.1-mini reasons over those signals to produce a final pick, confidence score, and risk flags for each game.

Key Features

• • Multi-model statistical edge engine combining four versions of the model, including L1 and L2 logistic regression, to compute projected margins, fair spreads, cover probabilities, and per-game expected value
• • GPT-4.1-mini synthesis layer with a structured system prompt and enforced JSON schema output, fed the quantitative signals to generate a final pick, confidence score, and risk flags
• • Historical game data pulled from a Neon PostgreSQL database and live odds from the NBA API
• • Tkinter GUI with live and demo prediction modes, progress tracking, and result export

Technical Approach

• • End-to-end pipeline from data ingestion through modeling to final recommendation output
• • L1 and L2 regularized logistic regression for edge estimation across multiple model versions
• • Enforced JSON schema on the LLM output for reliable, machine-readable predictions
• • Confidence scoring and risk flagging to communicate uncertainty on every pick

Results & Insights