Weather-Alpha Trading Desk

An autonomous agent that finds mispriced weather prediction markets and trades them — monitor, decide, transact, publish, 24/7, no humans.

Comment

Inspiration Prediction markets price tomorrow's weather off public data, while physics-based AI weather models see the atmosphere differently. That gap is an information edge, and an agent can harvest it 24/7 with discipline no human trader has.

What it does Every 5 minutes, with no human in the loop, an autonomous desk pulls forecasts for 7 US cities, compares them against live Kalshi prices on ~84 temperature contracts, and trades wherever the market disagrees with its model by more than 8%, sized with fractional Kelly. Every trade triggers a cited research note and a Slack alert, and its signal feed is monetized with x402, requests return HTTP 402 demanding $0.01 USDC paid to the agent's own wallet. Live: https://weather-alpha-agent.onrender.com

How we built it Python/FastAPI on Render, with a Jua-integrated forecast layer — full EPT-2 API client, built and tested against their spec — running on NOAA's physics ensemble, plus live Kalshi production prices for signals and hand-built RSA-PSS request signing for execution. ClickHouse stores every signal, trade, and P&L mark; Composio executes the Slack publishing; the x402 paywall verifies through the x402.org facilitator on Base Sepolia. Decisions come from Gaussian bin probabilities and quarter-Kelly sizing, covered by 9 unit tests.

Challenges we ran into Kalshi's RSA-PSS signing was the hardest hour of the day, and the demo exchange turned out to have no liquidity, so we read real prices from production and execute paper trades on demo. Then the agent lost money: it was trading same-day markets where the crowd already has live thermometer readings. We diagnosed the observation leakage in ClickHouse and restricted it to next-day contracts, where forecast skill is the only edge.

Accomplishments that we're proud of A closed autonomous loop, monitor, decide, transact, publish, get paid, running live in production right now. An objective scoreboard measured in P&L rather than vibes. A real x402 paywall where the agent quotes its own price to its own wallet, across six load-bearing sponsor tools.

What we learned The hard part of agents acting on the web isn't the LLM, it's signed requests, market microstructure, and honest probability math. When the agent lost money, the fix wasn't a better prompt; it was understanding where information lives: markets know today, forecasts know tomorrow.

What's next for Weather-Alpha Trading Desk Live trading with real risk limits and full on-chain settlement for the signal feed. More markets, rain, snow, energy load, anywhere a physics model beats crowd priors, with the agent recalibrating itself from its own trade history in ClickHouse.

Sponsors used: Jua, ClickHouse, Composio, Render, Senso/cited.md, CDP/x402

Built With

  • base
  • clickhouse
  • coinbase
  • composio
  • fastapi
  • jua
  • kalshi
  • noaa
  • python
  • render
  • senso
  • usdc
  • x402
Share this project:

Updates