mostly-autoagent

AutoAgent for prediction market strategy discovery. Can an AI agent, given only raw weather data and market prices, autonomously discover a profitable Kalshi trading strategy?

Forked from kevinrgu/autoagent.

What This Is

A meta-agent loop that iterates on a prediction market trading agent harness overnight. The meta-agent modifies the system prompt, tools, and strategy logic in agent.py, runs backtests, checks the score (PnL), and keeps or discards changes.

The trading agent gets access to:

• Weather observations (Therminal API / local parquets) — temperature, humidity, wind, pressure
• Kalshi OHLCV market data — 1-minute candles with real bracket definitions
• Backtester — simulates limit order fills, settles against real brackets, computes PnL after fees
• Python execution — for custom analysis

It does NOT get pre-built models, features, or existing strategies. It discovers edges from raw data.

Data Dependency

This repo requires data from mostly-harmless-v2:

~/Documents/GitHub/mostly-harmless-v2/data/
├── raw/ohlcv/           # Kalshi OHLCV candles per city series
│   ├── kxhighny/        # NYC: KXHIGHNY-26MAR19_1m.parquet, ...
│   ├── kxhighchi/       # Chicago
│   └── ...              # 20 cities total
├── raw/observations_therminal/  # Hourly weather observations
│   ├── nyc_metro/       # NYC_2024.parquet, NYC_2025.parquet, ...
│   ├── chicago_metro/
│   └── ...
└── raw/cli_ncei/        # Settlement ground truth (NWS CLI daily high)

Set the path in .env:

cp .env.example .env
# Edit MH_DATA_ROOT to point to your data directory

Files

agent.py           — trading strategy agent harness (meta-agent edits this)
backtester.py      — clean Kalshi settlement engine (ground truth, don't modify)
test_backtester.py — settlement logic tests
program.md         — meta-agent directive (human edits this)
tasks/             — evaluation tasks (NYC, Chicago, Miami, multi-city)
.env.example       — required environment variables

Backtester Details

The backtester handles the tricky parts of Kalshi weather markets correctly:

• Real bracket definitions from each OHLCV file (not hardcoded alignment)
• Even AND odd alignment — bracket floors vary per event (e.g., [66,67] vs [67,68])
• Settlement: round(actual_temp) checked against [floor, cap] inclusive
• Tail brackets: lower ≤ cap, upper ≥ floor
• Fill simulation: scans post-prediction candles for limit order fills
• Kalshi fees: taker ceil(7 * N * p * (1-p)) / 100, maker ceil(1.75 * ...)

Quick Start

# 1. Clone and set up
git clone git@github.com:helloiamvu/mostly-autoagent.git
cd mostly-autoagent
cp .env.example .env
# Edit .env with your data path and API key

# 2. Install deps
uv sync

# 3. Run the meta-agent
# Point Claude Code at the repo and say:
# "Read program.md and let's kick off a new experiment!"

Running Backtests Directly

from backtester import Backtester

bt = Backtester(ohlcv_dir="path/to/ohlcv/kxhighny")
signals = [
    {"date": "2026-03-19", "market": "KXHIGHNY-26MAR19-B47.5",
     "side": "YES", "limit_price": 0.40, "actual_temp": 48.0},
]
trades, summary = bt.run(signals)
print(summary)  # {'score': ..., 'win_rate': ..., 'n_trades': ...}

License

MIT (our additions) + LGPL-3.0 (upstream AutoAgent/NautilusTrader portions)