Uncertainty and validation

We recomputed the whole five-year history 300 times, each time redrawing the pillar weights, the event-type severities (±20%) and the half-lives (±30%). Band agreement is the share of runs in which the reading landed in the same band as the actual reading. The 90% interval covers 90% of the runs.

The real-time series uses on each day only the information that was public at the time. It is the headline because it shows what the index would have said that day. The revised series dates events to when they actually happened and uses later, corrected information; it is meant for analysis after the fact.

The two series differ by 0.15 points on average and by 11.8 points at most.

The model contains tests that run before every release. They check, for example, that ten airspace violations cannot push the reading above 49, that a verified armed attack lifts it to at least 70 and the warning level to 4, that unverified information cannot raise it by more than 15 points a day, and that a state of defence lifts it to at least 90.

Latest run: 24/24 tests passed.

The 7-day outlook in Premium briefings gives two probability estimates on a fixed scale: will the warning level rise, and will the band change within the next week. Every estimate is stored, and their accuracy (Brier score) is published here monthly. The first score is published once a full month of outlooks has accumulated.

• Public sources lag. Some events only become public days or weeks later.
• The weights are expert judgements. They were not fitted to history, which avoids overfitting but means they are choices. The sensitivity analysis shows how much those choices matter.
• Official assessments are cautious. Public statements by the authorities may be deliberately reassuring.
• The model cannot see secret information. Intelligence that is not public does not show in the reading.
• New phenomena. Anything that does not fit the codebook is recorded as unclassified and handled in the next model version.