Analyst Action Scanner

WHAT IT READS. The last 14 days of news about each company. A language model reads each article's headline and summary and keeps the analyst actions the article states - upgrades, downgrades, initiations, reiterations and price-target changes - with the firm, the ratings and the targets exactly as written. An action is kept only when the article's own sentence states it and every firm, rating and target in it appears in that sentence: the model can misread a sentence, but it cannot originate a firm or a figure. The same firm's same action reported by several outlets on one day counts once.

WHY A MODEL AND NOT A WORD LIST. 'Upgraded to Buy' and 'denies a downgrade' share vocabulary and mean different things, and which firm acted on which company is a question of reading the sentence. A rating change itself is a discrete, dated event; the model's job is only to find it in the text and copy it, and code checks the copy against the sentence.

COVERAGE IS THE LIMIT. An action reaches this analysis only when an outlet reports it in a headline or summary. A large company's rating changes are widely reported; a small company's may not be, so a company with no reading may have had quiet coverage rather than quiet analysts.

SEPARATING RATING CHANGES FROM TARGET-ONLY UPDATES is the right distinction and worth understanding. A rating change is a change of RECOMMENDATION, a considered and relatively rare act. A price-target adjustment is often mechanical housekeeping - the target follows the price, or a model input is refreshed - and analysts revise targets far more often than ratings. Treating them as equivalent would let routine maintenance outweigh genuine changes of view.

INITIATIONS AND REITERATIONS COUNT BY THEIR RATING. Starting coverage at Sell is not good news about a company. An initiation or a reiteration counts as positive at a buy-side rating, negative at a sell-side one, and not at all at hold or neutral.

THE ACADEMIC RESULT. Womack (1996) studied new BUY and SELL recommendations and found announcement returns followed by drift in the same direction, with the effect asymmetric - sell recommendations produced larger and more persistent negative reactions. It is a result about new recommendations in that sample rather than about every modern upgrade, downgrade, initiation, reiteration and target-only change. The asymmetry has a plausible explanation: sell-side coverage skews positive, so a sell is rarer and therefore more informative.

WHICH BRINGS IN THE BIAS. Sell-side recommendations have historically skewed heavily toward buy, with sells rare, a pattern documented over decades and attributed to incentives including banking relationships and maintaining access to management. Post-2000 US reforms expanded conflict disclosure and analyst certification, while the 2003 Global Research Analyst Settlement imposed structural separation of research from banking on the firms it covered rather than on every broker. The distribution improved without becoming symmetric. The practical consequence: because ratings have skewed positive, a HOLD has often represented a below-consensus view - though what it means depends on the broker's scale and the prevailing distribution.

SCALES ARE NOT COMPARABLE ACROSS BROKERS - three-point and five-point systems, different labels, different meanings for the middle - so netting actions into one score requires a mapping someone chose.

A STRUCTURAL POINT WORTH KEEPING. Every input here is other people's OPINIONS about a company, not measurements of it. That is genuinely useful for knowing what is already expected, and it is not evidence about the business.