How Aveluro Classifies Vietnamese Stock News
A detailed look at the two-pass classification pipeline Aveluro uses to extract structured facts from Vietnamese financial media and score their market impact.
Published 2026-07-20
Why classification matters for Vietnamese stock news
Vietnam's financial press publishes hundreds of articles per day across outlets like CafeF, Vietstock, VnEconomy, and VnExpress. For foreign investors who do not read Vietnamese, the challenge is not just translation — it is knowing which articles matter. A routine regulatory filing and a surprise earnings beat look identical in a raw RSS feed. Classification separates the signal from the noise.
Aveluro's pipeline processes every article through a structured extraction step that identifies what happened, which companies are affected, and how significant the event is likely to be. The result is a queryable database of market events, not a wall of untranslated headlines.
Pass 1: LLM fact extraction
The first pass sends each article to a large language model with a constrained prompt. The model does not score or opine — it extracts structured facts. The output schema includes an event type drawn from a fixed set of 27 categories (earnings beat, dividend announcement, regulatory action, M&A, leadership change, and so on), a magnitude object containing numeric facts like revenue growth percentage or deal size, a list of affected tickers and sectors, sentiment, urgency, time horizon, and source credibility flags.
Constraining the LLM to a strict schema prevents hallucination and makes the output machine-readable. The model fills in fields or marks them unknown — it never invents numbers. This approach costs roughly 30 times less than general-purpose summarisation because the prompt is short and the output is structured JSON, not free text.
Pass 2: deterministic scoring
The second pass is pure Python — no LLM involved. A scoring engine multiplies a base impact score for the event type by modifiers for magnitude, source credibility, novelty, and whether the event is likely already priced in. The formula is transparent and reproducible: the same facts always produce the same score.
Base impact scores are calibrated from historical data. An earnings beat starts at 7 out of 10; a routine board appointment starts at 3. Magnitude modifiers scale the score up or down — a 40% revenue jump is more impactful than a 5% one. Source credibility discounts scores from blog-tier sources relative to Tier-1 press. Novelty detection checks whether the same event was already classified from an earlier article, preventing double-counting.
Feed credibility and source tiers
Not all Vietnamese financial media is equally reliable. Aveluro assigns each RSS feed source a tier: Tier 1 for established financial press with editorial oversight (CafeF, Vietstock, VnEconomy), Tier 2 for general news outlets that cover markets (VnExpress, Thanh Nien), and Tier 3 for blogs and aggregators. The tier multiplies directly into the credibility factor of the deterministic score.
This tiering means that a rumour from an unverified blog scores lower than the same claim reported by CafeF, even if the extracted facts are identical. The system does not suppress low-tier sources — it weights them appropriately so that investors can see the full picture while understanding the reliability of each source.
What comes next: realized impact calibration
The planned third pass will compare each classification's predicted impact against the actual next-trading-day price move of affected tickers. This realized impact signal will be used to calibrate the base impact table over time, making the scoring engine self-improving. If the model consistently overestimates the impact of dividend announcements, the base score for that event type will be adjusted downward.
This feedback loop is what separates a static rule engine from a learning system. The LLM extracts facts; the deterministic scorer produces a reproducible estimate; and the realized impact data tells us whether that estimate was useful. Each layer has a clear job, and none of them requires subjective human judgement to operate.