kalshi-bias-engine

Automated trading system exploiting behavioral mispricing on Kalshi by comparing quoted contract prices against an independently-modeled fair value, executing entries and exits when calibrated edge clears fees and risk checks.

V1 trades crypto contracts only. Architecture is domain-agnostic from day one — sports, cross-venue event arb, and other categories get added later as plug-in modules, not rewrites.

Layout

• engine/ — Python trading engine (long-running process, Docker linux/arm64).
• dashboard/ — Next.js dashboard on Vercel, reads Neon Postgres.
• ops/ — deploy notes and infra scripts.

Topology

• Engine — Phase 1: local. Phase 2: Oracle Cloud Always Free ARM VM.
• Neon Postgres — source of truth AND control plane (kill-switch, phase gates, heartbeat). Vercel-Neon integration.
• Dashboard — read-mostly, only writes control-plane rows. Never touches the Kalshi order API.

Build phase

Phase 1 (paper). The execution/live module is import-gated and intentionally inert. Order placement unlocks per-domain when out-of-sample calibration passes the phase-gate checks in the calibration store — not by flipping a config flag.

Engine runtime loops

The engine process runs the following concurrently. All writes flow through the local SQLite spool, so Neon transience never stalls the trading loop.

• Trading loop (_discover_and_trade_once) — every loop_interval_sec discovers open Kalshi markets, upserts contracts, snapshots quotes, generates signals via oracle + bias model, and asks the PaperExecutionPolicy to decide. On a FILLED paper open an attached_fill piggybacks on the paper_order write so order + fill land in the same Neon transaction.
• Kraken ingest (WebSocket) — streams trades into the multi-horizon realized-vol store the oracle reads.
• Settlement ingest (settlement_poll_interval_sec, default 60s) — finds contracts past settlement_time without a settlements row and spools settlement writes from Kalshi's per-market endpoint.
• Refit loop (calibration_refit_interval_sec, default 15m) — pulls settled records, rebuilds the per-band per-domain calibration report, refits each bias feature with a 70/30 time-based in-sample/OOS split, spools calibration_snapshot + bias_params, and hot-loads the fresh params into the in-memory bias model. evidence_ok (min OOS sample AND min Brier improvement) is what actually toggles a feature on at runtime.
• Heartbeat (heartbeat_interval_sec) — dashboard reads staleness for the dead-man's-switch.