LLM-Guided Portfolio Allocation

A quant core sets the baseline; the LLM proposes tilts; constraints and human review bound the result before execution

LLM-Guided Portfolio Allocation A quant core sets the baseline; the LLM proposes tilts; constraints and human review bound the result before execution 01 / Quantitative Core 02 / LLM Advisory 03 / Risk Constraints 04 / Human Review 05 / Execution EX / Rejected -> Reset Baseline + propose Constrain + review Execute Market Data · prices + covariance · Quantitative Core › Baseline + propose Market Data prices + covariance Baseline Weights · MVO / risk parity · Quantitative Core › Baseline + propose Baseline Weights MVO / risk parity Context · news / filings · LLM Advisory › Baseline + propose Context news / filings Propose Tilts · views + rationale · LLM Advisory › Baseline + propose Propose Tilts views + rationale Risk Constraints · limits / turnover · Risk Constraints › Constrain + review Risk Constraints limits / turnover Human Review · sign-off · Human Review › Constrain + review Human Review sign-off Execute · OMS / broker · Execution › Execute Execute OMS / broker Reset · to baseline · Rejected -> Reset › Constrain + review Reset to baseline context approved breach fall back Legend User UI Agent logic Policy Tool action Context / trace

Quant Sets the Baseline

  • • Optimizer produces defensible starting weights
  • • The LLM never sees an empty portfolio to fill
  • • Every tilt is a delta from a known anchor

LLM Stays Advisory

  • • Proposes tilts and views from qualitative context
  • • Must attach a rationale to every proposed change
  • • Never places an order or writes final weights

Bounds Before Execution

  • • Position limits and turnover caps are hard checks
  • • A human signs off on anything consequential
  • • Any breach or reject falls back to the baseline

Workflow diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • [/] views • P play story • T theme • E export