An intraday statistical-signals service for US equities: data pipeline, model, API, billing, and brand. Here's the system, the decisions, and the call to wind it down.
Statistical models scanned hundreds of US equities every market hour and surfaced the ones with a measurable short-horizon edge. Each signal shipped with a direction, a win rate, and an entry/exit reference — delivered through a tiered, API-first SaaS with a Python SDK. The interesting part wasn't the signals; it was taking the whole thing — pipeline to billing — from zero to shipped, and being honest about where it worked and where it didn't.
A staged feature pipeline and a LightGBM ranker behind an AUC deploy gate — where every signal declared how trustworthy its own win rate was.
Tiering mapped to real cost and conviction, API keys, rate limits, billing, instrumentation wired to answer the question that mattered.
The engineering held; the edge and the business case didn't hold up. It was decommissioned deliberately rather than left to drift.
per_symbolIllustrative example of the API's signal payload — not a live or current signal. The case study explains how the win rate and expected return were actually derived.
How the pipeline worked, how the win rate was computed, the product and pricing decisions, the honest limitations, and why the service was wound down.
IrrationalSignals is archived and no longer operates as a live service. Nothing here is investment advice. Signals were statistical observations, not recommendations.