Methodology

Last updated: 2026-11-06

This page explains exactly how VolFlash produces what you read. If a claim below does not match what you see on the site, tell us — that is a bug.

1. Where the data comes from

• Global event stream — GDELT, queried continuously for conflict, sanctions, and policy events.
• Financial newswires — public RSS from CNBC, MarketWatch, WSJ, FT, BBC and similar.
• Chinese-language newswires — public feeds, for coverage the English wires miss.
• Economic calendar — scheduled releases with consensus and previous values.
• Market quotes — index, commodity, FX and crypto snapshots for the ticker.

Every event keeps its source name and the original link. We do not republish full articles; we quote the headline and summary and send you to the source.

2. How the severity score is calculated

The 0–100 score is a heuristic, not a market model. It is computed in four steps:

1. Keyword matching. Each event's title and summary are matched against category-specific keyword rules. Matches accumulate weights within their category.
2. Category base + saturation. The dominant category sets a base score. Additional matches in the same category add up to 28 more points, with diminishing returns — so one speech that hits ten keywords does not automatically reach 95.
3. Cross-category and noise adjustment. Matching several categories adds up to 8 points; known low-value patterns (routine listings, recap pieces, promotions) subtract a penalty.
4. Source weighting. Wire services and major outlets carry a small positive adjustment over unattributed blogs.

The result is clamped to 0–100. Thresholds you will see: ≥70 high impact, ≥85 critical. Scores below 30 are treated as thin content and excluded from search indexing.

What the score does not mean: it is not a probability, not an expected move size, and not a measure of how much money is at stake. It measures how much of a market-moving signal the text contains, according to our rules.

3. How the analysis is generated

For each event above a severity floor, a language model drafts the commentary from the headline, summary, category and score. The prompt imposes hard constraints:

• No numbers that are not in the input — no invented price levels, thresholds, or percentage moves.
• No buy/sell instructions, no price targets, no "will definitely" claims.
• Speculation must be phrased conditionally ("if X, then Y").
• Exactly four key points, in a fixed order: fact, transmission path, what to watch, what would invalidate the call.
• The direction label must agree with the conclusion; a contradiction is treated as a defect.

Output is structured JSON, produced in Chinese and English separately — the English version is written as English, not machine-translated word for word. When the model is unavailable, the system falls back to a rule-based analysis and labels it as such.

4. Reading the labels

• Direction — risk-off / risk-on / mixed / neutral. Which way the event pushes risk appetite, not which way a specific ticker will move.
• Magnitude (1–5) — how large a reaction the event could plausibly cause, independent of how important the news is.
• Confidence (0–100) — the model's own confidence in the read, lower when the source is thin or the event is ambiguous.
• Horizon — intraday, days, or weeks. Most geopolitical shocks are intraday; policy shifts run for weeks.
• Asset bias — up / down / volatile / neutral per instrument, with a one-line rationale.
• Watchlist — the indicators to check to see whether the read is playing out.

5. Deduplication

Ten wires report the same headline. We normalise each title (lowercase, strip punctuation) and compare it against the last 400 titles; anything above 0.82 similarity is treated as the same event and dropped. This is why a story that appears on five sites shows up once here.

6. Timing

Collection runs roughly every 180 seconds. New events are analysed within about 40 seconds of ingestion. The free tier sees events after a 120-minute delay; Pro sees them immediately. Chinese titles and summaries are machine-translated for the English site, and untranslated events are withheld from English pages rather than shown in Chinese.

7. Known limitations

• Recall. If no source in our list covers an event, it is not here. Quiet markets and regional stories are under-represented.
• Precision. Keyword scoring produces false positives — a headline that mentions "war" in a metaphor still scores. Severity is a prioritisation tool, not truth.
• Generation. The analysis can be wrong, generic, or miss the real driver. It is written from the headline and summary, not from the full article.
• Translation. Chinese-to-English titles and summaries are machine-translated. Nuance can be lost; the original source link is always present.
• No price validation. We do not verify that the affected assets actually moved as described.

8. Corrections

If you find an error, email yeluyide@gmail.com with the event link. We correct the record or remove the event, and where the error is systematic we change the rules that produced it.

9. What we do not do

No personalised advice, no portfolio management, no performance claims, no paid placement in the feed. Paid members get more of the same product — faster and in full — not a different, more promising one.