The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management

Battle Count: 50

Highly relevant to institutional quantitative investing and strategic asset allocation. The paper directly addresses portfolio construction, CMA generation, risk management, and ensemble methods central to quantitative trading. However, it focuses on strategic (long-horizon, quarterly/semi-annual) allocation rather than tactical or high-frequency trading. The agentic architecture, multi-method ensemble, and IPS governance framework are directly applicable to quantitative investment management. The LLM-as-judge and multi-agent deliberation components represent novel approaches to the portfolio construction problem.

Extremely complex system requiring: (1) orchestration of ~50 specialized agents with distinct roles, descriptions, scripts, skills, and output contracts; (2) multi-step reasoning pipelines with structured deliberation (peer review, Borda-count voting, adversarial diversifier); (3) integration of 20+ portfolio construction methods spanning heuristics, mean-variance, risk parity, CVaR, TPA, and novel methods; (4) LLM-as-judge frameworks for CMA selection and CIO ensemble; (5) a self-modifying meta-agent that rewrites code and prompts; (6) IPS governance layer; (7) data infrastructure (Bloomberg, FMP, Finviz APIs, web search); (8) structured output contracts (JSON schemas + markdown reports); (9) backtesting infrastructure; (10) security considerations for autonomous tool-invoking agents. Requires significant engineering, LLM expertise, and financial domain knowledge.

The paper describes the architecture in detail with agent configurations, skill definitions, and workflow steps, but no code, data, or specific LLM model is disclosed. The pipeline runs on an internal system. Reproduction would require access to the same LLM infrastructure, data APIs (Bloomberg, FMP, Finviz), and the full agent configuration files. The March 2026 run results are illustrative but not independently verifiable.