Agentic Multi-Strategy Trading System

I built a multi-strategy US equities research system to test a narrow thesis: LLMs are useful in the creative research loop, but the money path should remain deterministic, replayable, and pitfall-aware. The system separates proposal and critique from execution, risk, and verification.

The final result is intentionally honest. In the 2013-2015 bull window, the system did not beat either benchmark on a risk-adjusted basis. What it demonstrates is a disciplined agentic workflow with gates that can say no, point-in-time data handling, and a zero-LLM deterministic path for portfolio construction, allocation, and PnL.

Locked OOS result

Annual walk-forward, 2013-2015

The system

-0.083

Sharpe

Total return: -2.01%
MaxDD: -9.66%

Equal-weight universe

0.879

Sharpe

Total return: 41.65%
MaxDD: -15.29%

Cap-weighted universe

1.012

Sharpe

Total return: 43.63%
MaxDD: -12.43%

The spread is the point: over this window, broad-market beta won. The system is lower-beta and includes market-neutral exposure, so the verification result reads as beta versus alpha rather than a victory lap.

Equity curve

The current image artifact is synced from the results directory. The final sync before deploy should regenerate this chart from the locked result bundle.

Open verification details

Equity curve for the system against the equal-weight benchmark over the 2013-2015 out-of-sample window

See the substrate

Architecture

Two-layer agentic design

LLMs research and critique candidates; deterministic code owns validation, portfolio construction, allocation, and PnL.

Results

Annual walk-forward verification

The locked 2013-2015 OOS result is shown against equal-weight and cap-weighted universe benchmarks.

Signals

Rejected creative candidates

SignalResearcher proposed six economically motivated variants. The pre-registered gates rejected all six.

Writeup

Guardrails and lessons

Correctness guardrails, agent-first communication, PIT discipline, and what the system demonstrates.

Policy population

as of 2012

PM	2006	2007	2008	2009	2010	2011	2012	Mandate
long_only	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	50L/0S / vol n/a
market_neutral	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	50L/50S / vol n/a
low_vol	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	ACTIVE	50L/0S / vol 8%

Market-neutral entered probation after the structural decay decision; the other PM mandates remained active.