Prescient · prediction-market intelligence
Prescient, a cross-platform market-error detection & execution system
I built Prescient, a cross-platform market-error detection and execution system for Kalshi and Polymarket. It ingests the full prediction-market tape alongside a crypto and macro firehose, estimates true event probabilities better than the displayed price, and promotes strategies gate → shadow → canary → live entirely on evidence.
22collectors
9data layers
28models + fusion
20/20integration tests
System map
Five layers, strictly ordered.
L1Ingestion spine
22 collectors: Kalshi REST + WebSocket + user-fill streams, Polymarket REST/WS, short-horizon binary book recorders, sharp-book odds (de-vigged Pinnacle fair values), Alpaca equities + news, FRED economics, CDE futures, 19-symbol crypto spot breadth, sports ratings. All fan into a Redpanda/Kafka spine.
L2Data platform (3-tier)
HOT: Redis · WARM: TimescaleDB compressed hypertables + continuous aggregates · COLD: zstd Parquet lake via DuckDB. Beneath it, the nine-layer data model: raw event lake → canonical contract ontology → structural graph → derived microstructure → cross-asset reference → settlement & oracle truth → trader/flow intelligence → external signals → feature→model→ledger spine.
L3Analytics & models
90+ module domains: an edge catalog (structural, flow, microstructure, cross-venue) · truth calibrator + structural-residual models · swarm fusion over 28 models · queue-accurate replay engine · walk-forward evaluation harness · counterfactual crypto engine (off-policy evaluation + Thompson-sampling bandit).
L4Execution & MM engine
Four-core market maker: live local order book → logit-space fair-value model → inventory- and toxicity-aware quoting (Avellaneda-Stoikov family) → event-factor hedge engine. 426 tests. RSA-PSS venue auth · capacity-aware sizing · canary harness with hard-gated order submission · verified venue-truth reconciliation.
L5Governance & ops
The self-proving loop: 9-point statistical gate → shadow daemon → gate-calibration harness → automated confirm/decay classifier → canary. Control tower: live operator boards. Self-healing: feed-freshness watchdog with Docker-level restart, disk-monitor auto-prune, Postgres-independent watchdog.
Full architecture, evidence record, and honest economics in the white paper. Operational specifics redacted for the public edition.