NEDI

What is Measured

For every year, NEDI reads the Events section of the Wikipedia page (for example “1943”) and asks of each dated entry: is this a negative event, and how sure are we? Share of record — the paper's volatility index — is negative entries divided by all entries for the year. Incidents is the raw count; est. deaths sums tolls quoted in the text; deaths per 100,000 divides those by approximate world population.

How Entries are Classified

About 150 weighted patterns map entries to nine categories. Unambiguous terms (massacre, earthquake, coup) carry high weight; ambiguous ones (war, crisis, fire, collapse) carry little and need corroboration — a casualty word, a toll or a second term. Evidence combines as 1 − Π(1 − w) into a confidence score, and the confidence setting chooses the cut-off: strict ≥ 0.75, balanced ≥ 0.55, inclusive ≥ 0.35.

Sub-entries inherit their parent bullet (“World War II: …”) at reduced weight. Known traps — heart attack, Dead Sea, post-war, Cold War, Operation Typhoon, Star Wars, italicised titles — are neutralised before matching.

What is Set Aside

Mentioning a crisis does not make an entry a negative event. These rules exclude an entry or reduce its score; every set-aside entry is listed on the front page with its reason, and may be restored.

Death Tolls

Tolls are parsed from phrasing such as “killing at least 944” or “sinks with all 98 crew”; ranges take the lower bound. A running total (“the death toll passes 20,000”) adds only its rise since the same story's last total, so one war's toll is never set against another's. Totals spanning several years are shown but not added to any single year. A story's deaths are the sum of its reports; for a story that began that year or the year before, its largest running total counts instead when it is higher. “Killed or wounded” and “left homeless” are never read as deaths. Many entries give no figure, so tolls are a lower bound of what is written, not of what happened.

The Story of the Year

Each entry's topic is read from its label, its parent bullet or its leading link (“World War II: …”). Topics with two or more counted entries are stories, and every story is scored on one fixed index so that editions compare: its share of the edition's entries (40 per cent), the months it runs through (25), its share of the deaths quoted, on a logarithmic scale (20), and the number of languages Wikipedia covers it in (15).

A story leads the front page when . The next story then takes the second slot, if it has three entries or more and scores 0.20 or better. A bare place name is a dateline, not a story, and an era, anything that ran for more than 35 years such as the Cold War, never heads the page.

Gathering the Strands

Wikipedia records a war through its battles. The paper asks Wikidata what each topic is part of, or which conflict it belongs to, up to three steps: the Battle of Kursk is part of the Eastern Front, which is part of the Second World War. Topics that share such a parent are read as one story, filed under the parent that gathers the most entries. A parent qualifies only if it is a dated event under way that year, lasting no more than 35 years, and not an aftermath, history, list or relations page. The chain is printed beside each strand.

How the Front Page is Made Up

When no story leads, the lead is the counted entry with the highest product of confidence, reported toll (on a logarithmic scale, so that magnitude matters without drowning all else) and a category weight, and the year's top story runs beside it. Readers may choose a different lead. The further reports follow the same ranking; all else appears in the Summary of the News, by section, in the order Wikipedia lists it. Where one column runs longer than the others, reports are carried over beneath the plate and brief items set in the short columns, so that no column ends in blank space.

Readership

From 2016 the paper also reports what readers read: Wikimedia's monthly page views of English Wikipedia, human readers only, for the 25 articles behind each edition's highest-ranked reports. Views under an article's former titles are added in, since the figures do not follow a page when it is renamed. A story's figure is the median of the twelve months from its first report, so that a single day's spike does not decide it and a December story compares fairly with a January one. The figures arrive after the page is set, and it is reprinted once as the Late Edition. Readers may lead the page by attention; readership never enters the index, for the figures begin in July 2015 and would break the series there.

Of Bias

Recency: modern pages are far longer, so raw counts rise regardless of how violent the world was; share of record divides this out. Geography: English Wikipedia over-represents Europe and North America; the World pages compare coverage with population. Missing years: early years that redirect to a decade page are marked as having no edition. Severity: a massacre and a skirmish each count once; use the toll measures for magnitude.

The Trend Test

The Trends page uses a Mann–Kendall test with a Theil–Sen slope, both robust to single outlier years, and reports a direction only when p < 0.05.

Data, Caching and Citation

Pages are fetched live from the MediaWiki API in batches, with shared back-off when Wikipedia throttles, and kept in this browser — 24 hours for current years, 7 days for the last decade, 30 days otherwise. Shift+click Run analysis to refresh. Each edition names the exact Wikipedia revision from which it was compiled. Wikidata figures and readership are kept for 30 days (readership for the current year, a day). The CSV and JSON exports carry each edition's stories, their scores and lead, its readership where measured, and each entry's story and chain.

Citation: Gostt, J. A. (2026). Negative Event Density Index (NEDI), parser .