Kalshi Trading Pod
Three Claude agents research, backtest and paper-trade Kalshi markets
What it is
A self-contained multi-agent pod for Kalshi prediction markets. A Researcher agent proposes strategies from live markets and news, a Developer agent backtests them on real Kalshi history, and ephemeral Trader processes paper-trade the approved ones with Kelly sizing and risk gates. The agents talk only by moving markdown files between pipeline folders, and a human approves each transition in a React dashboard. The folders hold 58 rejected ideas and 151 logged transitions from real runs.
The problem
Finding and validating prediction-market strategies by hand is slow, and an unsupervised trading agent is dangerous without gates.
How it works
- The pipeline is the filesystem: research_ideas, actively_researched, validated, live and archive, with an append-only moves.jsonl audit log.
- A Developer daemon backtests each idea against a Parquet dataset of settled Kalshi markets (weather, Fed, CPI, payrolls, GDP) and appends a verdict.
- Each Trader loops on the orderbook, calls Claude, sizes with a deterministic quarter-Kelly and a five-gate risk check, and orders on Kalshi's demo sandbox by default.
- A FastAPI backend with WebSocket events feeds a React dashboard showing live agent status; retired strategies get a final report the Researcher learns from.
Stack
Disciplines
For your business
A supervised research-to-execution agent pipeline where AI proposes and tests ideas and a human promotes each one through explicit gates.