ZenInvest
An open experiment in where LLMs actually add value to investing.
ZenInvest is an open-source investing system with deterministic risk rules, broker-truth reconciliation, and a full audit trail. Its first design let a multi-vendor committee of Claude, GPT, and Gemini debate every trade. Six months of live paper-trading evidence showed that committee had no entry edge, and the results were published as a working paper. The redesign, ZEN-2, lets a deterministic full-universe ranking choose the shortlist, asks one LLM analyst for a falsifiable thesis memo per name, and measures the agent against an index arm and a rules-only arm. Autonomous buying is paused while ZEN-2 is built.
Read the R&D brief: what six months of evidence showedRead the working paper
featured
Product framing
Most AI investing claims are never tested against a baseline. ZenInvest publishes its own evidence, including the negative results, and treats every claim about LLM value as something to measure rather than assume.
Capability278 realised trades analysed
CapabilityPre-registered decision rule
CapabilityThree-arm evaluation (index, rules, agent)
Stack
Why ZenInvest
Measured against baselines, not assumed.
ZEN-2 keeps the model in an advisory role, makes every thesis falsifiable, and tests the agent against control arms on the same universe and costs.
One LLM analyst writes a falsifiable thesis memo: catalyst, horizon, entry, size, invalidation level, and review date
An optional skeptic can only veto; it cannot add or resize a trade
Deterministic risk rules resize or veto every final weight, and stops sit at the thesis invalidation level
An index arm and a rules-only arm run on the same universe and costs, so the agent's value is measured, not assumed
How it works
A daily loop from ranking to review.
Each cycle is staged so outcomes remain inspectable, governed, and testable.
Rank
A deterministic engine ranks the full universe and shortlists at most five new names a day, plus positions due for review.
Write the thesis
An LLM analyst researches each name with audited tools and writes a thesis memo with an explicit invalidation level and review date.
Decide
Deterministic sizing and risk rules set or veto the final weight. No model can override them.
Measure and learn
Exits follow the thesis, not the clock. Weekly post-mortems feed a lessons loop, and every result is compared with the index and rules-only arms.
Differentiation
Built as an inspectable agentic system.
The emphasis is on structured challenge, clear risk handling, and inspectable behavior rather than black-box promises.
Negative results published, not buried
The committee's six-month record (a 37.8% win rate against a ~57% break-even, and a loss while the index rose) is published with the reasons behind it.
Measured against control arms
ZEN-2 is only kept if it matches or beats a rules-only arm and stays within three points of the index after costs, six months after go-live. The rule is frozen before the first trade.
Deterministic rules keep the veto
Risk rules, sizing limits, and stops are deterministic Python that no model can override. Chat-initiated trades still require explicit human confirmation.
Research posture over return promises
A research and engineering experiment with educational value, not a machine for returns.
Architecture and ecosystem
A governed agent backed by a wider tooling ecosystem.
ZenInvest combines orchestration, market data, evaluation, and execution tooling across a broad set of APIs and platforms.
Core architecture
Deterministic ranking, audited research, falsifiable theses.
The system combines market intelligence, audited research tools, and deterministic rules that no model can override. The goal is not to replace judgment, but to find out, with evidence, where a model genuinely improves it.
Ecosystem
Use cases
Useful beyond investing.
The evaluation discipline transfers to any team asking where AI genuinely helps a decision.
A test bed for LLM value claims
For teams asking where language models genuinely help a decision process, and how to prove it against a baseline.
Governed agent design
For builders exploring how to keep models in an advisory role while deterministic code owns constraints and execution.
Honest evaluation practice
For anyone who wants a worked example of pre-registration, control arms, and publishing negative results.
Governance and operations
Safety rules and operator visibility are first-class, not bolted on.
The system enforces deterministic controls and maintains transparency across runs, decisions, and reporting.
Safety guardrails
Hard constraints remain above model output.
- Concentration limits for single names and sectors
- Drawdown state machine: ACTIVE, CAUTIOUS, and HALTED
- Cash floor and exposure constraints enforced by policy
- Risk veto cannot be overridden by LLM outputs
- A kill switch pauses autonomous buying while exits and stops keep running
Operational signals
- 1,400+ automated tests with fail-closed gates
- Docker Compose deployment on VPS
- Run history, alerting, and audit trails
- Daily and weekly reporting with cost and outcome visibility
Disclaimer
Public project, not financial advice.
The product can be ambitious while remaining careful about claims and live use.
ZenInvest is not financial advice. It is an educational, research, and product experimentation project running on a paper-trading account, with autonomous buying currently paused. Human oversight remains essential, and any live use should be approached cautiously.