MarketMania is a forward-only record of what seven frontier language models predict about live crypto markets, and of what those markets then did. At every forecast slot each model receives the same disclosed payload and answers the same question on the same five instruments. Slots run every hour for the 1h horizon, every four hours for the 4h horizon (read on the 1h and 4h timeframes), and once a day for the 1d horizon (1d and 4h timeframes) and for the 1w and 1M horizons. The answer is stored before the horizon closes, and it is scored against the resolved outcome afterwards. Nothing is back-filled and nothing is re-scored.
Two things are measured, and they are kept apart. Prediction quality asks whether a model was right and whether its stated confidence matched how often it was right — directional accuracy, Brier score, calibration gap. Trading performance asks what a mechanical paper-trading rule would have made of the same forecasts on Binance USDT-M futures, net of fees — win rate, net PnL, max drawdown. The two are never merged into one score, because a model can be well calibrated and unprofitable, or profitable and badly calibrated, and one number would hide which.
The research reports are windows cut from that record and frozen: the numbers an issue prints are the numbers its window held at publication, and they are not restated when later data arrives. Each issue ships as a PDF for people and as a Markdown and JSON pair for machines, at URLs that do not move. This page is the description of the record itself; the reports are on Research, the live indices on Indices, and the arithmetic behind every index on Methodology.
Counters are read live from the public stats endpoint when this page is rendered. A figure the endpoint did not return reads — rather than a remembered value: this page publishes no number it has not just fetched.
The grid is fixed by design, so a reading from one week is comparable with a reading from another.
Three surfaces, one record: frozen report files, the live read API, and the daily snapshots the reports are cut from.
Versioned, published in full, and named on every issue that used it.
Payload transparency, the validation rules, the paper-trading simulation and every weighting formula behind a composite index are set out on Methodology, with the benchmark-side rules on Benchmarks methodology. A report states the methodology version it was produced under; amendments are numbered and ship with the issue that introduces them.
The 6 most recent issues. The full rail, including every earlier issue, is on Research.
Cite the issue you read, by its title, its publication date and the URL of the file — an issue is frozen, so a citation to one is stable.
MarketMania Research (2026). Config Watch #8. October 1, 2026. https://marketmania.ai/research/reports/config-watch-2026-09-23.pdf
To cite the record rather than one issue, name the dataset and this page: MarketMania Research (2026). Research reports (weekly, monthly), PDF/MD/JSON. https://marketmania.ai/research — dataset card: https://marketmania.ai/research/dataset. To cite a single live reading instead, name the index, the instrument and cell, and the date it was read: readings move, issues do not.
What may be done with the files, and where to write.
The reports and the public endpoints are free to read and carry no key. Use is governed by the Terms of Use. Questions about the dataset, corrections and access requests: contact@marketmania.ai.