Inside ASCEND-EQ: How the AI Stock Engine Prices a Session
Every published equity idea carries a live reference price, a target, an invalidation level and a position weight — the four numbers required to grade it later.
Universe first, ideas second
Liquidity and spread floors remove anything where a published entry and invalidation could not realistically be filled. US large caps, megacap technology and liquid ETFs clear that bar; thin names do not, regardless of how attractive their factor scores look.
One idea per ticker per session. Duplicate exposure dressed as diversification is one of the easiest ways to make a record look better than the risk taken to earn it.
Eight weighted factors, published in full
The composite score combines trend quality, earnings-revision drift, flow, breadth, volatility term structure and the remaining published factors. Weights sum to one and are refit quarterly against out-of-sample sessions — never against the window being reported.
The composite is mapped to a calibrated probability, then compared against what the tape already prices. Ideas without a material gap are discarded at that stage.
The reference price comes off the tape
Entry, target and invalidation are derived from the live quote at publication and recent realised volatility, so a quiet mega-cap and a high-beta name are not handed the same distance by default. Each card states whether its price came from the live tape or from the model's fallback band.
Position weight is fractional Kelly with an explicit cap, and every idea states the horizon it is expected to resolve inside.
Settlement is mechanical, and open ideas stay visible
Published ideas are logged with their real entry price, then checked against subsequent real prices. Target first is a win, invalidation first is a loss, and anything still open at the end of its horizon closes at the prevailing price and is recorded as it landed.
Because settlement is mechanical, an idea cannot be quietly abandoned. The accuracy panel stays empty until real settled rows exist rather than showing a projection dressed as a result. Directional equity models live in the mid-50s; anything advertising far above that is describing a curve fit.