GEO — The Kalshi Weather Bot

Live demo: apeabody007.github.io/GEO-The-Kalshi-Weather-Bot-DEMO

A live algorithmic trading system for Kalshi weather prediction markets — a CFTC-regulated event-prediction exchange. GEO scans 1°-wide bracket and tail-cutoff contracts on daily-high temperatures across 20 U.S. cities, prices each market against a multi-model probabilistic forecast, and trades mispricings with real capital.

---

What you're looking at

This repository hosts a public, read-only preview of GEO's live operations dashboard, rendered with fully synthetic data. No real positions, no real bankroll, no live API calls. Every section that would reveal the bot's competitive setup (per-model diagnostics, calibration curve, structural-arb detector, resting maker orders, project roadmap) is locked and intentionally absent from the public preview. Source code is private.

The live system runs on the operator's machine against real Kalshi markets, served as a Tailscale-private PWA.

What's here

index.html              the rendered dashboard (single self-contained file)
manifest.json           PWA manifest
icon-*.png              app icons

That's the entirety of the public surface. The dashboard's renderer, data feed, server, strategy code, calibration models, and trade ledger are all private and not part of this repository.

Engineering at a glance

• Stack: Python 3.11, signed REST client against the exchange's API, SQLite trade ledger, server-rendered vanilla-JS dashboard, Tailscale-served PWA.
• Forecasting: multi-model probabilistic ensemble with per-station bias correction calibrated against historical climate observations.
• Trading layer: fractional-Kelly position sizing under multiple risk caps; maker-only pricing to sidestep taker fees.
• Reliability: single-instance file lock prevents order collisions, circuit breaker around upstream APIs throttles rate-limits, year-round local-time handling avoids DST settlement misalignment, ~110 pytest cases gate every release.
• Observability: server-rendered HTML on every page load (no caching between operator and the truth), per-cell drift detection, full P&L attribution by multiple decision dimensions, calibration reliability curve, automated reconciliation against the exchange's official settlement source.

See also

Quant Toolkit — a venue-agnostic Cowork plugin distilled from operating GEO. Nine skills covering the full loop from idea to live trading: Kelly sizing, calibration audit, EMOS bias correction, backtest harness, maker pricing, P&L attribution, pre-flight checklist, market scanner, drawdown monitor. Bring your own model and exchange.

Built by

Aaron Peabody — BA Economics (Wisconsin), BS Psychology (UCF). For collaboration or hiring inquiries, reach me on LinkedIn.