Skip to content

Transparency

Track Record and Changelog

Last revised: November 2026

On this page

This page documents how the index is built and how it has behaved. It is an analytical indicator based on news flow, not a prediction. The formula itself is explained on the Methodology page.

1. Methodology versions

Live version: v3 since 10 Nov 2026, 10:13 UTC.

  • v3 · 10 November 2026: Articles grouped into events and scored once, with a corroboration factor and a severity cap for single-source reports; actor-aware weights; an independent second opinion for high-impact items; a 5-year GDELT baseline instead of 12 months; a data quality label.
  • v2 · 6 November 2026: Computed UCDP structural component (38%) with a floor; a 14-day component next to the 72-hour one; de-escalation capped at 40%; percentile calibration against 12 months of GDELT event history.
  • v1 · Date not recorded: A 72-hour news sum with saturation, plus a 20% constant baseline.

Stored switch records

  • 10 Nov 2026: 1 switch recorded. First to v3 at 10:13 UTC (manual).

2. Reaction to reference events

From the stored backtest report of 10 Nov 2026, 10:19 UTC (not all checks passed), covering 10 Nov 2021 to 10 Nov 2026. The backtest measures the GDELT news-flow score of the live formula (0–100), because news articles are only kept for 15 days. Change is the largest move within 72 hours of the event (a fall for de-escalation); the percentile ranks it against every other day.

DateEvent72h changePercentile
2022-02-24Russia launches a full-scale invasion of Ukraine+60.997.8
2023-10-07Hamas-led attack on Israel and start of the Gaza war+63.698.3
2024-04-13Iran launches missiles and drones at Israel+62.598.1
2024-10-01Iran launches a missile attack on Israel-2.026.5
2025-05-07India and Pakistan exchange strikes+9.248.4
2025-06-13Israel strikes Iran+14.257.2
2025-06-22US strikes Iranian nuclear sitesn/an/a
2025-06-24Israel-Iran ceasefire announced (de-escalation)n/an/a

n/a: no backtest value for this date (GDELT has no data for 15 June to 1 July 2025).

3. Full index history

202220232024202520260255075100

Daily values from 9 May 2021 to 10 Nov 2026, methodology v3. Reconstructed from GDELT before 9 Nov 2026; stored live index from then on (solid vertical line). Dashed lines mark the reference events (orange escalation, green de-escalation).

GDELT has no data for 11 November 2022.

GDELT has no data for 22 March 2023.

GDELT has no data for 15 June to 1 July 2025.

4. Data sources

News outlets

Headlines and links belong to their publishers. We store only the headline, link, date and our own short classification.

RSS feeds

Datasets and services

5. Known limitations

  • Relative scale: News flow is ranked against recent years, so bands are relative to recent history, not absolute measures of danger.
  • Reconstructed history: Days before the first stored live day are reconstructed from GDELT with the live formula in its GDELT-only mode, with the structural component from the previous year's UCDP count. They do not include our article classification.
  • Daily noise: Older GDELT history uses one 15-minute sample per day, so reconstructed daily values are noisy.
  • Missing GDELT days: GDELT has no data for 11 November 2022, 22 March 2023 and 15 June to 1 July 2025. These days are left empty, never interpolated.
  • Backtest scope: The backtest covers only the GDELT-based component, because news articles are kept for 15 days.
  • Coverage and language bias: Major English-language outlets cover some regions far more than others; under-reported crises appear smaller.
  • Machine errors: GDELT coding and AI classification can be wrong. Every item links to its original source.

The full list is in the Methodology page. To report a mistake, use the contact page.

6. Corrections and methodology changes

  • 10 Nov 2026 · Methodology change

    Methodology v3 is live (previously v2)

    Index methodology updated to v3: event clustering, actor-aware weights and a second AI opinion for high-severity items. Values before and after this date are not directly comparable.

    Affected page