How company announcements are analysed and scored
Each NSE and BSE filing is read by an AI model, given a 1-10 relevance score against a published rubric, and checked against the company's own numbers.
Listed companies file thousands of announcements a week with NSE and BSE, and most are routine. Live Market Tracking reads each one and says what it is, how much it matters to the company, and what to watch next.
1. Text and pre-screen
- The filing's PDF is downloaded and its text extracted (the first 15,000 characters are analysed).
- A statistical pre-screen (a logistic regression over the filing's subject line, the exchange's summary and the first 6,000 characters) skips filings that are very likely routine. In testing it skipped about 39% of NSE filings while missing 2 of the 548 that went on to score 8 or more. It never skips filings about dividends, results, corporate actions, fundraising, mergers, orders or capital spending, and a random share of the filings it would skip are analysed anyway, so its miss rate keeps being measured.
- An exact re-filing of an earlier document reuses the earlier analysis.
- Procedural filings (trading-window notices, takeover-code disclosures, employee stock option allotments, newspaper copies, compliance certificates) get a shorter analysis of the first 3,000 characters, still scored with the full rubric.
2. The analysis
An AI model reads the filing together with a summary of the company's own numbers: price, market value, P/E, trailing revenue and profit, eight quarters and five years of results, promoter holding and peers. That lets it state how large the news is relative to the business, for example an order worth about 52% of trailing revenue. The model is told to use only the document and those figures, to write "not disclosed" for anything in neither, and to give no buy, sell or hold call. It runs at a low temperature, so the same filing gets nearly the same answer every time.
3. The relevance score (1-10)
| Score | What earns it |
|---|---|
| 9-10 | Fraud, mergers and acquisitions, an order or capital spending above about 25% of revenue, a rating change, a large fundraise, major regulatory action |
| 7-8 | A meaningful order win, a capacity expansion, an earnings surprise, a dividend change, a key management change |
| 5-6 | Routine but notable: in-line results, small contracts, scheduled dividends |
| 1-4 | Procedural: intimations, board-meeting dates, duplicates, trading-window notices, newspaper copies |
Each filing also gets an impact label: positive, negative, neutral, or "watch" for material news whose outcome is genuinely uncertain (a dispute, a pending regulatory step, a non-binding agreement). The analysis lists the key figures, risk flags and a confidence level (high, medium or low).
4. Where the score is used
- Live Market Tracking shows filings scoring 5 or more by default; lower scores are one click away.
- An order win or capacity expansion scoring 8 or more goes on to Multibagger AI for a deeper brief (see its methodology).
Limits
The score is an AI model's judgement against a fixed rubric, not a measured fact, and the model can misread a document. Every analysis links to the original filing, which is the authority. The score describes how material the news is to the company; it is not a view on the share price.
