TirraMind collected 34 free public data sources and detected statistical anomalies across them. Then I ran the experiment almost nobody runs: I tested whether my own product actually predicted anything.
Every figure reproducible from the repository below.
Zero of fifty-one hypotheses survived Benjamini-Hochberg correction. The study honoured the CFTC's publication lag, used the unconditional return as its null rather than zero, and reported its own statistical power — about 10%, which is stated up front because overclaiming a null is the same error as overclaiming an edge.
It replicates the published literature exactly. Re-running the same correction on a well-cited 2009 paper's own p-values gives the same answer: nothing survives.
Along the way the pipeline lied to me thirteen documented times — while reporting green every time. A ranking I was about to sell was 90% Unix timestamp. A quarter of my anomalies were artifacts of a four-month baseline where three years of history sat unused. Every one passed the tests I'd written.
So I wrote it all down, open-sourced the collector, and published the null with the power calculation attached. The interesting artifact was never the product.