How we score a post: 13 factors, two gates, one 24-hour window

Every analysis on this site answers one question: if this post (or event) is the only new information, which US-listed companies move, in which direction, and by how much over the next 24 hours of regular trading? This page documents exactly how that number is produced.

1. Two gates decide whether a name can move at all

Before any weighting, two gate factors are scored for each candidate company and multiplied, not added. If either gate is zero, the name is dropped from the board.

• Association — how tightly this post ties to that specific company (named party, supplier, customer, regulator, counterparty). A company reached only through "sentiment" scores zero.
• Surprise — how much the news deviates from what the market already expected. A rumour already priced in scores low even if the headline is large.
• Base capacity (institutional) — float and liquidity. If institutions cannot build a position, the institutional read is zero.
• Tradability (retail) — whether the crowd can actually act on it: coverage, borrow, liquidity.

2. Nine weighted factors produce two propensities

Institutional propensity and retail propensity are computed separately, because the same headline does not move the same money the same way. Weights are fixed in code, not chosen per article.

Institutional: specificity 30%, value magnitude 30%, liquidity delta 25%, position room 15%.

Retail: emotional charge 25%, virality 20%, comprehensibility 20%, price appeal 20%, familiarity 15%.

The scorecard printed on each analysis shows the raw factor scores and both composites. The article text explains the reasoning; the numbers come from the platform, never from the language model.

3. The percentage is absolute, not relative

A printed move such as +1.2% means we expect that stock to move +1.2% from the price at the moment of analysis over the 24-hour window. It is not a relative call against the market, and it is not a target price.

The no-event forecast (what the stock would have done anyway, read from the last sessions) is added to the event move; the net decides whether a name lands on the bullish or bearish side. That is why a small positive event call on a bleeding stock can still print as bearish.

4. How a call is settled

Each call creates a settlement row: direction, predicted move, window length, and the price at the moment of analysis (the anchor). Only regular-session minutes count toward the 24-hour window — nights, weekends and market holidays are excluded.

When the window closes we read the price again and compute the actual move from the anchor. A hit means the realized move had the sign we called. Misses stay on the page and in the public hit rate.

Published hit rates are segmented by how many transmission steps separate the company from the news: direct (named in the post), one or two hops (supplier, customer, counterparty), and three or more (sector mood).

5. What this model does not do

• It does not read order flow, positioning, or options data.
• It does not predict prices beyond the stated window, and it makes no long-term recommendation.
• It cannot see news published after the analysis, including corrections to the original post.
• Nothing here is investment advice.

To report an error or question a call: [email protected]

Keep reading

• How we score a post
• Data sources
• Editorial policy
• All analyzed posts