How a Session Is Built

The engine rebuilds its tradeable universe before anything is scored. US large caps, megacap technology and liquid ETFs must clear median-spread, depth and borrow floors sized against the position the published idea would actually need. Anything that fails is removed for the session, however attractive its factor scores would have been. A published entry you cannot get filled at is not research — it is decoration.

Every survivor is scored across the eight weighted inputs and collapsed into one composite z-score. Scores are cross-sectional, not absolute, so the model is always asking which names look mispriced relative to the rest of the universe today, not whether the market as a whole is cheap.

The composite is mapped through a logistic calibration fit on out-of-sample sessions only. This is the stage most vendors skip: a raw signal tells you direction, a calibrated probability tells you how often that direction has historically resolved. The slope is deliberately shallow because honest directional equity accuracy lives in the mid-50s, not the 90s.

The calibrated probability is compared against the probability already embedded in the tape for the same outcome over the same horizon. Only a material gap survives. A beautiful composite score with no gap against market pricing is discarded without appeal — the whole business is the difference between the two numbers, not the size of either one.

Survivors receive a fractional-Kelly weight capped well below full Kelly, plus an explicit invalidation level set at the expected-move distance divided by 1.6. That ratio is what makes a 38.5% break-even the honest bar for the desk, and it is why a losing idea costs materially less than a winning one pays.

Thirteen ideas post with the full reasoning attached: six conviction calls, four swing setups and three volatility reads. Nothing is edited after the fact. If an idea loses, it stays on the record exactly as it was published, which is the only way a hit rate means anything at all.