A fleet of autonomous agents that reads all 503 S&P 500 stocks after each U.S. close — scoring momentum, fundamentals and sentiment against the whole market, and publishing its highest-conviction signals with the reasoning attached. Built and operated by one engineer. Research only; it trades no money.
Updated automatically after each nightly run. Every signal shows its factor breakdown — why the system ranked it, not just that it did. These are research outputs, not recommendations.
MERIDIAN is one agent inside a larger autonomous system. A central orchestrator dispatches specialised agents on schedule: MERIDIAN generates the signals, LEDGER manages a simulated portfolio against them, a quantitative advisor reviews the mathematics weekly, and an infrastructure agent handles backups, monitoring and data-integrity guards.
Every number the system reports must cite its source query and sample size — a standing rule adopted after live operation showed how easily a confident-looking figure can be computed on broken data. The system is designed to distrust itself.
Each stock is scored not in isolation but against all 500 — a cross-sectional z-score answers “how unusual is this name tonight, relative to everything else?” The composite blends four views of that question.
Trend strength and persistence — RSI, MACD, moving-average structure, volume confirmation.
Business quality — return on equity, revenue growth, margins, leverage and cash generation.
News tone over recent coverage, scored for each name and normalised across the universe.
Market regime context — volatility, rates and policy backdrop. Under redesign as a regime modulator following validation findings.
Stocks inside an earnings blackout window are excluded before scoring — the system does not take positions into binary event risk. The top quartile of the composite becomes Tier 1; the middle half is held as a watchlist.
MERIDIAN completed a 60-day live paper validation against criteria locked before day one. The system ran unattended, generated signals nightly, and a simulated portfolio traded them under fixed rules — stops, targets, sector caps and capital limits — with no human intervention in the trades.
| Closed trades in window | 23 |
| Simulated result | positive, modest |
| Average win vs average loss | 1.49 : 1 |
| Sample required for statistical confidence | ~100 trades |
| Verdict on edge | cannot yet be confirmed |
The honest reading: at 23 trades, the confidence interval around any win rate spans roughly ±20 points — far too wide to distinguish genuine edge from chance. A positive result at this sample is encouraging and statistically silent at the same time. The validation also surfaced real findings — data-freshness gaps, a structurally inert macro factor — now driving the v2 redesign.
The system continues in extended paper validation toward a statistically meaningful sample, across at least one full earnings season. No real capital is deployed, and none will be unless the evidence earns it.