Measured fact, confirmed event, interpretation, and unknown are different objects.
Temporal proximity does not prove a cause. Detector severity does not equal explanation confidence. A primary source can strongly confirm an event while the event’s market impact remains uncertain.
Directly recomputed from normalized stored data.
Supported by a primary source or independent high-quality evidence.
A qualified inference linked to evidence and alternatives.
No responsible unique cause is established.
Versioned detectors create candidates.
We use log returns, normalize against an asset’s own volatility and liquidity, reject stale inputs, and retain detector features for replay.
| Detector | Core input | Output |
|---|---|---|
| Price anomaly | Log return vs. rolling median/MAD | Abnormal move severity |
| Volume anomaly | Log volume vs. rolling baseline | Participation confirmation |
| Volatility break | Realized volatility ratio | Regime change |
| Market breadth | Advancing share and trend participation | Broad vs. narrow move |
| Correlation break | Return correlation vs. prior equal window | Relationship change |
| Scheduled catalyst | Approved calendar/primary source | Reminder and post-event review |
The first score is deliberately inspectable.
materiality =.25 × move abnormality+ .15 × volume confirmation+ .15 × liquidity quality+ .15 × breadth context+ .15 × evidence quality+ .10 × novelty+ .05 × source velocityPersonal relevance is calculated separately from direct watchlist exposure, optional manual quantity, related assets, saved topics, repetition, and explicit feedback.
A bounded packet is the only explanation input.
- Primary official sourceProtocol, regulator, court, exchange status, or company publication.
- Verifiable recordExplorer, signed governance record, or other direct technical evidence.
- Recognized reportingWire or specialist source with attribution and editorial controls.
- Secondary corroborationMultiple independent sources rank above a single unverified source.
Unverified social posts may start an investigation but cannot independently support a confirmed explanation.
Two scores, never one.
These are editorial product defaults, not calibrated probabilities. Source trust tiers impose ceilings on event confidence.
Structured output must survive validation.
- JSON Schema and required fields
- Every factual claim’s evidence IDs
- Numeric and time-window consistency
- Allowed assets, entities, dates, and URLs
- Source-tier confidence ceilings
- Prohibited advisory and guarantee language
- Similarity and duplicate active explanations
After one bounded retry, a failed generation moves to human review or becomes an observation-only card.
Figures are generated from data, not generative imagery.
Correlation views use price log returns and Pearson coefficients on overlapping observations. Adjacent equal-length windows expose changes. Exports contain derived values only where provider terms permit redistribution.
Monitoring software cannot see every cause.
- Approved sources are intentionally narrower than the full information environment.
- Market prices can incorporate private, diffuse, reflexive, or unobservable information.
- Historical normalization can fail during structural breaks.
- Entity mapping and provider data can be wrong; quality flags and corrections remain necessary.
- Pulse is research and monitoring, not investment advice, custody, execution, or a prediction guarantee.