Market Eyes Live

How MELANY is tested

The complete statistical write-up behind this page, with full tables, confidence intervals, and limitations, is published as Working Paper No. 01: Validating a Systematic Equity Rating Engine.

1. Stress-tested against 19 years of market history

MELANY's risk and portfolio rules are validated on historical daily data reaching back to 2007, a window that spans the 2008 financial crisis, the COVID crash of 2020, and the 2022 bear market. The deepest test bed holds 541 large U.S. names with daily bars since 2007, sampled across roughly 714,000 reference days. A broad companion bed adds 4,390 names including small caps.

Survivorship bias, the classic way backtests flatter themselves, is addressed directly: portfolio construction was tested point-in-time on an eligibility universe that includes 1,789 delisted companies, so the names that went to zero are in the test, not quietly missing from it. Where a data set is survivor-only, our findings say so instead of hiding it.

What this does and does not claim: it is the engine's rules (risk bands, entry behavior, portfolio construction, drawdown control) that carry 19 years of testing. The published ratings themselves are younger, and they are held to a harsher standard: the live graded record below.

2. The ideas our own testing killed

A validation program that only ever confirms its own ideas is a marketing department. Ours rejects things, and we think the rejections are worth publishing:

• "Wait for the pullback" was reversed. A backtest of 12,805 events found no evidence that waiting for a moving-average pullback improved results under any statistical weighting, and the advised limit orders only ever filled about 37% of the time while the names that never pulled back kept running. The live advice was reversed the same day.
• Sector-rotation and seasonality entry gates were rejected. Plausible-sounding defensive filters failed holdout validation, and several signals pointed the wrong way entirely. Nothing shipped.
• A downgrade-on-dip hypothesis was ruled inconclusive. In-sample support did not hold up in the holdout period, so nothing was deployed and the finding was recorded as unresolved.

Every rejection above exists internally as a committed script, a raw results file, and a findings document, so each can be re-run and re-checked at any time.

3. Forward-tested daily: the live graded record

Since August 6, 2026, every rating MELANY publishes is captured at the moment of publication and graded automatically against subsequent market moves at fixed horizons. Captures are immutable and grading is mechanical, so the record cannot be edited, trimmed, or cherry-picked after the fact. It went public before any results existed, which is the point: a record you start publishing only after it looks good is not a record.

Graded summaries (hit rate versus benchmark by conviction tier and horizon) appear in the feed automatically once a cell reaches 20 graded outcomes spread across at least 8 distinct tickers; below either floor a number is noise and is not quoted, so a dramatic figure can never rest on a handful of names.

One company, two products

The Market Eyes Live app (iOS and web) and the free public Market Eyes Live MCP server for AI assistants are the same company and the same MELANY engine. The MCP serves the free public data tier; the app carries the full research depth (entry zones, alert levels, portfolio tools, Lock Radar).

Try Market Eyes Live free Read the raw record feed