How Our AI Scores Financial News

The pipeline, end to end

1. Ingest: hundreds of headlines and summaries arrive daily from ~20 public RSS feeds (Reuters, MarketWatch, Bank of Korea, Maeil Business, TokenPost, and others), polled every few minutes.
2. Score: a large language model reads each story and decides three things — ① direction (bullish / bearish / mixed), ② impact (0–100), ③ the key evidence behind the call, plus English/Korean summaries.
3. Deduplicate: stories about the same event are clustered; one representative article is shown, and the event counts once in every index and accuracy calculation.
4. Aggregate: calls roll up into the Bull-Bear Index (BBI guide).
5. Verify: calls with impact 60+ are checked against the real price 24 and 72 hours later (verification guide).

What the AI looks at

• Direction: rate cuts, ETF approvals, earnings beats → bullish. Rate hikes, regulatory crackdowns, exchange hacks, earnings misses → bearish. Genuinely ambiguous or offsetting impact → mixed.
• Impact: could this move the whole market? Macro events like Fed decisions or CPI land in the 90s; single-stock news lands in the 40s–70s depending on size; market-irrelevant stories lower still.
• Balance rules: the model is explicitly instructed against known biases — "do not default crypto stories to bullish", "mixed is not an escape hatch; use it only when impact is genuinely ambiguous".

Known limitations (honestly)

• The AI gets things wrong. It can miss context, misread irony, or score news the market already priced in. That is exactly why every call goes into the public verification ledger.
• Headline-based: it analyzes headlines + summaries, not full articles, so deep-in-the-body nuance can be missed.
• News skews negative: media publishes more bad news by nature. A higher count of bearish stories doesn't automatically mean markets fall — impact weighting partially corrects for this.
• Speed over depth: an automated pipeline is not a substitute for a human analyst's depth. BullBear is a tool for quickly seeing what came out and which way it points.

Where humans are involved

Day-to-day scoring is fully automated. The operator monitors pipeline health (ingest volume, analysis rate, accuracy trend) and improves the source list and scoring rules. Individual story scores are never hand-edited, and sponsors cannot buy a score (Editorial Policy).

This guide is informational only, not investment advice.