Methodology

Data Sources

News is ingested every five minutes from Tier-1 Vietnamese financial RSS feeds: CafeF (securities, corporate, banking, macro), Vietstock, VnEconomy, VnExpress, and VietnamNet. Articles are deduplicated by URL. Price and volume data comes from HOSE, HNX, and UPCOM via vnstock. Financial statement ratios are sourced from company filings.

AI Classification (Pass 1)

Each article is sent to a large language model with a structured extraction prompt. The model returns an event type (one of 27 categories: earnings beat, M&A announcement, rate decision, regulation change, etc.), a magnitude dictionary of numeric facts (revenue growth %, deal size in USD, rate delta in bps), affected tickers, affected sectors, sentiment, urgency, and source credibility flags. The prompt explicitly prohibits opinion or speculative scoring — only extractable facts are returned.

Deterministic Impact Score (Pass 2)

A Python scoring engine converts AI-extracted facts into a deterministic_score between 1.0 and 10.0:

score = base_impact[event_type] × magnitude_factor × credibility_factor × novelty_factor × priced_in_factor

• base_impact — a lookup table per event type (e.g. rate_decision=9, earnings_beat=7, general_news=3)
• magnitude_factor — scales by numeric facts (e.g. profit growth >50% → 1.4×)
• credibility_factor — Tier-1 press 1.0×, general press 0.85×, blogs 0.65×; rumours discounted by 0.7×
• novelty_factor — cosine similarity against recent articles on the same event type; duplicate stories discounted
• priced_in_factor — if the affected ticker already moved >3% the day before publication, the news may be priced in

Page Generation

Articles with deterministic_score ≥ 4.0, a specific event type (not general news), and at least one identified ticker are eligible for a dedicated research page. The page body (~500 words) is generated by an LLM given the original Vietnamese article, extracted facts, ticker price context, and company metadata. High-scoring articles (score ≥ 7.0) use Claude Sonnet for higher-quality English prose; others use DeepSeek V3. Pages with fewer than 300 words are not published.

Composite Ticker Score

Each ticker's AI score shown on stock pages is a daily composite of recent article sentiment weighted by deterministic scores, technical momentum, and fundamental ratios (P/E, ROE, debt-to-equity). It is a relative ranking signal, not a buy or sell recommendation.