PROJECT — 2026

An automated prediction-market trading system with a two-stage LLM architecture and fractional-Kelly risk management.

Two models share the desk: a scout proposes probabilities, and a judge vetoes overconfident trades. Position sizing is fractional-Kelly with per-market exposure limits, stop-loss, and take-profit rules.

A market classifier routes each question to external data — FRED, Open-Meteo, news, sports odds — and a probability-calibration layer shrinks estimates toward market-implied prices. An offline backtesting harness evaluates strategy calibration against a 1,200-market dataset, reporting per-category Brier scores.

Python · LLMs