// the find
AgriQuantAI/AgriQuant-AI
AI-powered weather intelligence for agricultural commodity futures. Predicts price moves 48-72 hours early across OJ, Coffee, Cocoa, Sugar, Corn and Wheat using Claude Sonnet + satellite data
AgriQuant AI is a Python system that combines NOAA weather feeds, satellite imagery, and Claude Sonnet to generate price-direction signals for agricultural commodity futures 48-72 hours ahead. It targets quant traders and agri-commodity desks who want a weather-driven edge over traditional analyst timelines. The backtest numbers are the main draw.
The domain decomposition is solid — separate collectors per commodity (cocoa_westafrica_collector.py, coffee_brazil_collector.py, grains_collector.py) rather than one god-class pulling everything. The 'What Worked / What Didn't' section in the README is genuinely rare and useful: freeze warnings with 48hr lead time at 75%, hurricane divergence shorts at 85%, and honest callouts about trading preliminary model runs being too noisy. The signal logic is concrete and rules-based rather than hand-wavy, so you can reason about when it breaks. GARCH volatility forecasting alongside the ensemble predictor is the right call for commodities where vol clustering matters.
The backtest runs 2023-2025 and shows 23 trades — that is an extremely small sample for claiming 70% win rate and 1.8 Sharpe. The disclaimer buries 'hindsight signal construction', which is the actual problem: if thresholds were tuned on the same data used to report performance, these numbers are meaningless. There are files named crypto_fundraise.py and investor_distribution.py sitting in the root alongside the trading logic, which suggests this was partially built to raise money rather than to be used — a real red flag for how seriously to take the architecture. The entire system is a flat directory of 30+ Python files with no package structure, which will become painful to maintain or test. Claude Sonnet is used as the 'AI analysis pipeline' but the README never explains what it actually does in the signal chain — if it is summarizing NOAA text into a structured judgment, that is fine; if it is making the trade call, that is a different risk entirely.