Multi-factor scoring and regime-aware projections, with the grading machinery pointed back at itself. Every forecast is scored against what happened. Every strategy is measured against random entries with identical mechanics. When the answer is no edge, that gets written up at the same length as a win.
It ran live for two months and looked like it worked. Then we replayed the frozen rules over 4.7 years and ran them against a control that drew entries at random using the same slots and exits.
Our rule landed on the 50th percentile of forty control runs. Against simply holding the same ten stocks, it turned $1,000 into $1,492 while doing nothing made $2,653.
Read the post-mortem →The scoring is ordinary quantitative work. The part worth paying attention to is what checks it.
Projections are scored against what actually happened, on a fixed horizon, whether or not the result is flattering. Calibration error is published, not summarised.
A long-only strategy in a rising market earns from exposure alone. Every result is measured against random entries using identical slots, exits and cadence.
Dashboards have reported page-length caps as totals. A verdict job once sent an error string as its verdict. Those are written up in the log, not quietly patched.
The intraday strategy was retired in September after out-of-sample testing put it on the 50th percentile against random. That post is the first one in the log.
Billing is off. Paid tiers switch on when the weekly ship gate passes — sustained calibration, positive net expectancy on every active lane, and a clean verdict, ten weeks running. It has not passed. The dashboard and the research log are free and stay free.