Crypto market data without the plumbing.

Basis. smart_money_confirm, 4–24h, direction-adjusted move, live forward-generated, re-emissions collapsed to distinct signals.

Distinct calls. A live call is re-confirmed every few minutes for as long as it persists, so the raw feed holds thousands of rows. Each episode counts once here (a gap of more than 12h starts a new call), which is why the win rate cannot be inflated by how long a call stays live or how often a bot polls it.

Forward holdout. The out-of-sample tile counts only signals resolved after the scorer was frozen (—) — scoring cannot have been tuned on them. It reports accruing instead of a percentage until enough signals have resolved, and it never falls back to the in-sample figure.

Cost basis. The plotted curve is net of round-trip taker fees. Per-signal statistics (profit factor, average win/loss) are the server's gross, pre-cost numbers and are labelled as such on each tile.

Sizing. Compounded fixed-fraction sizing at 5% of bankroll per signal, both series starting at 100. Return and drawdown scale with size; the ratios (win rate, profit factor) do not.

The solid line replays the published per-signal returns at one fixed position size; the dashed line is BTC buy-and-hold over the identical window, so the two are directly comparable from the same start value of 100.

It is a record of signals we emitted, not a simulation of an account: it does not model funding or borrow cost, leverage, liquidation, margin calls, partial fills, or the fact that these signals overlap heavily — real margin could not have carried every one of them at once.

The full version of this record, every horizon and cost basis, and the leverage replay are on the performance page. The five studies that came back negative — the ones that found no durable edge — are written up on the methodology page.