Evidence: does World Tension Meter work?

The test. A signal is a three-day rise of at least one standard deviation of that index’s own daily movement. A hit is the market series it should lead rising by one standard deviation of its own movement, inside the window. Both thresholds come from the data itself, so nothing is tuned to flatter the result — and misses are printed, not deleted.

The number that matters. A hit rate alone proves nothing — oil rises often anyway. So every row also shows the base rate: how often that market made the same move in any window of the same length. The gap between them is the edge. Across the tested indexes: 38% of signals were followed, against a base rate of 49% — an edge of -11 points.

Read it index by index. The average hides the spread, and the spread is the honest part: some indexes lead their market clearly, others show no edge yet. Both are printed, with the misses.

And read it with the sample size in mind. Each index has produced only a handful of signals since the recalibration — far too few to be statistically meaningful. One event can swing a row from 100% to 50%. We publish it anyway, from day one, precisely so the record cannot be started later after a good run. The first real verdict is due March 2027.

What happened. The first engine ranked every signal against seed ranges rather than against its own history. During a period of real war news that pushed the World Tension score into the 90s and kept it there: 120 of 121 days before 23 July 2026 closed above 77.

What changed. From 23 July 2026 the engine compares every signal with its own observed history (a rolling percentile), so “extreme” means extreme for this world, not for a guessed range. Since then: 0 of 63 days above 77. Average moved from 91.3 to 59.4.

Why the old numbers are still there. We do not rewrite history. Every day is kept as it was published, with the engine version attached, and the monthly stories from that period carry a calibration note. If we quietly recalculated the past, nothing else on this site would be worth trusting.

What it means for you. Compare inside a period, not across the join. Use the record since 23 July 2026 for anything where the level matters. The track record above is tested only on that period for the same reason.

No — and here is the split. Today, 48% of the World Tension score comes from measured market prices (oil, gas, gold, yields, the dollar, freight, grain), 47% from counting our own news corpus, and 5% from public attention data. The split moves with the world: in a quiet week prices carry more of it; during a war, coverage does.

Why headlines at all. Markets price what traders already know. Coverage moves first on wars, sanctions and blockades — that is exactly the lead we are trying to measure. The safeguard is that a headline count is never used raw: it is ranked against our own history, so a loud news day only counts if it is loud for us.

And the words never write the number. A language model classifies headlines and writes plain-English summaries. Every score is arithmetic on counts and prices, reproducible from the published weights.

Four inputs. How many headlines the cycle read (50), how many different sources they came from (3), whether any market field is estimated (1), and how fresh the data is (0.0 hours). Weighted 30 / 30 / 25 / 15 into a 0–100 reading shown on every index page.

What a low number means. Fewer signals stand behind today’s score — a holiday weekend, a feed outage, a thin news day. The score is still computed the same way; you just have less evidence under it. We would rather show that than hide it.

All of them, named. The engine reads these feeds twice a day. No paid wire, no private data, nothing behind a login — anyone can fetch the same articles and reproduce our counts.

Where they publish from

Western outlets dominate English-language world news, and we do not pretend otherwise. Two things reduce the bias: conflict share is measured over world feeds only, so business coverage cannot dilute it; and every count is ranked against our own history, so a persistent editorial tilt affects the baseline rather than the daily move. Regional gaps we know about: Latin America and Central Asia are thin. Suggest a feed: [email protected].

Built for: watching pressure build across the world in one place; early warning weeks before official inflation and freight numbers catch up; research and journalism on an open, citable dataset; teaching how geopolitical risk reaches prices.

Not built for: trading signals, insurance pricing, security decisions about a specific place or person, or any call where being wrong is expensive. We publish a measurement of pressure, not a forecast of events.

What we claim. That the numbers are computed the same way every day, from sources anyone can check, and never revised. That direction over 7 and 30 days carries more information than any single day. That the track record above is the honest result of a fixed test.

What we do not claim. That six months is long enough to prove anything. That a high score means an event will happen. That expert judgement is unnecessary — a good analyst reading the same news will see things this engine cannot.

Known limits. English-language sources dominate. Media attention is not the world: a war with no cameras scores lower than it should. Prices are delayed or estimated when a feed is down, and always flagged. The AI Risk Index and country scores are young; their pages say so.

Every month the engine commits to numeric ranges for the month ahead, timestamped and append-only, then scores itself in public. See the record →

Method: methodology · Data: track_record.json, predictions.json, portal · Independence: funding statement · Changes: changelog. Informational only, not investment, legal or security advice.