Mirage
A pre-registered engine that put ~50 trading-strategy families on trial — 40 reached a verdict, none produced a deployable edge, and the negative result is honestly documented. Built to not fool itself.
The problem
Most backtests quietly confirm what you hoped. The real question: does any retail-accessible trading edge survive a test designed to disprove it?
The approach
Every strategy was pre-registered against a fixed eight-step gate — friction realism, look-ahead audit, out-of-sample sign stability, matched-control permutation — with kill criteria written into each pre-registration. The numeric core is pinned by a unit-tested reference evaluator.
Highlights
- A fixed eight-step gate built to catch the recurring classes of false positive
- Every test carries a written pre-registration — hypothesis, cohort, windows, kill criteria — and the gate fails any run without one
- A unit-tested reference evaluator pinning the numeric core, with per-test gates aligned to it
Results
- 40 strategy families tested, 0 deployable edges
- 1,000+ behavioral tests
- A taxonomy of how fake edges fail
- Open-source, with a written paper draft
We build systems that refuse to fool themselves — valuable anywhere a number drives a decision.