Catalyst Radar: Real-Time Market-Event Detection Engine

The production build behind the Catalyst write-ups: a single-process Python engine that scans ~163 perpetual markets (crypto, equity, commodity, forex) every five minutes, ranks unusual movers by a vol-normalized composite score, routes each to the right news sources, classifies the catalyst with Claude Haiku, and runs survivors through a five-rule suppression chain before pushing a deterministic alert. Three cooperating asyncio loops in one process.

Code

View catalyst-radar on GitHub →

The full arc

Everything in this project, in the order it happened — the thinking is as much the artifact as the code:

1. Multi-Timeframe Execution Engine: A Complete Architecture → (initial research) — early architecture work on reconciling conflicting signals across timeframes into a single net position. One of the first specs I worked through before building anything.
2. Catalyst Radar v1: Architecture Optimization Brief → — tore apart the v1 spec’s three structural flaws (uniform polling cadence, one LLM call per ticker, naive ranking) and redesigned the loop, the two-stage prefilter pipeline, and the ranking formula.
3. Catalyst Detection Engine v2: Peer Review & Spec → — an adversarial peer review of my own architecture: where it leaves IR on the table, the LightGBM meta-label gate, CryptoBERT classification, and five free-tier data assumptions that died in 2026.
4. A Quant Backtest of 482 Live Breakout Alerts → (validation) — an adversarial study of 482 live alerts with the raw CSVs attached: the engine has no edge at baseline, one two-variable filter survives every robustness check, the biggest lever is a time-based exit, and the LLM direction call adds nothing measurable. The finding that settled it — this is infrastructure and research, not a trading desk.