More records about the same source story
Group exact source URL + encoded occurrence date. Record count stays visible, but it is not multiplied into review priority. Ten coded mentions are not ten closures.
MODEL NOTES / SCREENING 2.0
A quieter alert is not
a smaller disruption.
Blackdeck separates source observations, inference and unknown quantities. Every score is a review aid, not a claim of economic loss.
First, use the correct unit of analysis.
RECORD
One coded observation
A GDELT row with an action, actors and location. Many rows can come from the same news article.
STORY GROUP
One source URL + event date
These records are grouped for display. A story can contain several actions and therefore several taxonomy labels.
REAL INCIDENT
Still needs verification
Same URL does not guarantee one real incident; different URLs may describe the same incident. Cross-publisher semantic deduplication is not yet complete.
Review priority, with inspectable components.
Review priority = S × E × N × R
- S · coded type severity
max(0, −Goldstein) × 10, capped at 100. An event-type reference value, not observed magnitude. A small and a large protest can have the same code value.- E · evidence weight
0without an eligible pathway;0.30for URL-only mechanism + asset clues;0.45for a coded mechanism + URL asset clue;0.65for the specific coded economic-restriction pathway. These are design weights, not calibrated probabilities.- N · new-information weight
- Discounts similar prior coverage, then restores attention when coded severity, an asset clue or a mechanism changes. See below.
- R · recency
2^(−age_hours / 72), relative to the source snapshot cutoff. Old source data is dated visibly; the website does not present it as a live measurement.
We removed the old “Global Risk Index” from the landing page. Averaging coded severity is not a defensible measurement of worldwide business loss. Likewise, correlation between a score and the mentions/source counts already inside that score is not an independent impact study.
REPETITION MODEL
Three cases, three different treatments.
Similar coverage, without a material new signal
Compare the country, actor pair and set of root event types with earlier story groups. The comparison uses earlier reporting dates within seven days. It is a similarity heuristic, not proof that the actual incident is identical.
N₀ = max(0.20, 1 / [1 + 0.5 × ln(1 + n)])
Here, n is the number of similar prior story groups. For an already seen source URL, the base new-information weight is capped at 0.25.
Prior similar groups0149Base novelty retained100%74%55%46%
Illustrative formula outputs, not measured reactions from markets or supply chains.
A worsening or materially different situation
Restore attention if the coded type severity exceeds the earlier maximum, or a newly matched asset or transmission mechanism appears.
N = N₀ + (1 − N₀) × material_change
material_change is the largest of: severity increase / 25 (capped at 1), 0.70 for a new asset clue, and 0.50 for a new mechanism clue. A large coded escalation can restore the full novelty weight. Evidence still requires review.
Keep persistence separate from novelty.
Blackdeck also measures how persistently a similar subject appears across reporting dates in the last 14 days. Repeated records on the same date count only once in this persistence measure.
A = Σ 2^(−days_ago / 7)
Reporting persistence = 100 × (1 − exp[−A / 5])
This is reporting persistence, not outage duration. Several days of reporting cannot establish that a port remained closed throughout those days. The value is not used as a fabricated cumulative loss.
Physical accumulation needs different inputs. Verified lost output can be accumulated from a measured production shortfall over time. Delivery risk requires actual delay, route alternatives and inventory buffers. Those measurements are not present in the current event feed, so actual barrels, tonnes, days and monetary loss remain “Not verified”.
How the model avoids false precision.
Multiple tags are not additive losses
A port strike may carry protest, labor, access and shipping labels. It still counts as one story group, and each impact edge keeps its own supporting evidence.
Unknown is not zero
No estimated production loss or price return is inserted just to complete a dashboard. Unmapped events may still be important.
Source coding is not truth
The map uses machine-coded geographic centroids. Captions derived from URLs are explicitly identified. Neither proves the location of an affected facility.
Coefficients need validation
The novelty and evidence parameters are a transparent initial design. They are not trained or validated business-impact coefficients. Future calibration requires labeled incidents and observed outcomes.
References and scope.
GDELT: Goldstein describes event types, not the scale of a particular event ↗
Tetlock (2011): investor responses to stale news ↗
The study concerns financial-news responses. It does not validate these coefficients, or imply that repeated conflict always causes less physical damage. The present formulas are Blackdeck design choices.