Agentic Multi-Strategy Trading System

A multi-strategy US equities research system built around a two-layer architecture: an LLM research and critique layer (per-strategy signal specialists, an adversarial critic, and a training-time signal proposer), and a deterministic execution layer (validation, portfolio construction, capital allocation, risk controls, and OOS replay). The primary question was how to keep LLM judgment bounded while still using it where it adds value.

Focus: point-in-time data discipline, agentic architecture with a zero-LLM money path, pre-registered primary aggregation (capital-weighted), and honest OOS reporting — including the concentration of returns in 2013 and two near-flat years after.

Deliberately out of scope: factor-model risk decomposition (build-or-buy decision deferred), live execution reconciliation, capacity and slippage modeling beyond a flat 10bps round-trip, and multi-year OOS validation. The 2013-2015 window is spent; any further parameter changes would require a new pre-registered OOS run.