openInvest

A self-hosted investment decision engine built for modern AI agents. Multi-agent information isolation and cross-challenge protocol, providing an auditable decision trail (Audit Trail).

📚 Full Architecture Wiki · 🇨🇳 中文版

---

What is OpenInvest?

OpenInvest is a self-hosted investment decision engine built for modern AI agents.

It provides a verifiable investment committee, evidence-based reasoning, long-horizon backtesting, and auditable decision records. Instead of replacing Claude Code, Codex, Hermes, or OpenClaw, OpenInvest is designed to power them.

---

Live Performance & PnL

Data feed is automatically updated every 2 hours using jobs/pnl_snapshot and pushed to the pnl-data branch

Upper half: 30-day net asset value trend · Lower half: Net asset value comparison against 8 benchmark assets (transparent disclosure, not an alpha claim—the committee's proven value is discipline and transparency, not excess return, see ADR-023)

📌 Note: The current chart shows the author's live production portfolio. After self-hosting, the system will automatically render your own equity curve based on the holdings defined in your memory/ directory.

• Benchmark Portfolio: The system introduces 8 standard control benchmarks across 4 quadrants (AI advisors / Mutual funds / Wealth management / Broad market index). For details on the comparison methodology and data cleaning logic, see docs/wiki/README.md.

---

Research & Falsification

System Self-Disclosure: This system is an auditing tool to eliminate human investment cognitive biases and enforce reasoning transparency, not a return-amplifying black box. Latest automated audit (docs/verdict_accuracy.md): Directional verdicts (excluding HOLD) have a true hit rate of 42.2% (n=56, below random); HOLD accounts for 56% of all decisions. The system's value lies in transparency and discipline (mostly staying inactive, low turnover), not directional prediction. Detailed log stream can be found in docs/verdict_accuracy.md.

This project systematically attempts to falsify its own edge and publishes negative results as-is. The deterministic features the committee reads, and the timing signals around them, were tested against pre-registered statistical gates — none survived as tradable alpha.

Test	Result	Verdict
Q1 cross-sectional stock picking	6 features, mean-IC 0.025–0.067, Holm-corrected p=0.397	No significant stock-picking signal
M1 multivariate GBM (out-of-sample)	mean OOS IC +0.003, p=0.925	Feature combination doesn't help either — no signal
Q2 gold MA200 trend	p_holm=0.016, significant — but trend_dca shows it is beta, not tradable alpha: timing terminal value 3.07 vs 15.10 buy-and-hold, Sharpe +0.36 vs +0.68, max drawdown deeper (−57% vs −44%)	Statistically significant, economically untradable
Per-asset multi-signal families	3 assets × 4 signal families × parameter grid = 24 variants per asset; after costs + DSR deflation, none passes DSR > 0.95	No tradable signal in any family
Positive control	A cheating perfect-foresight timing signal scores DSR = 1.00	The harness can detect a real signal