Tested on every crisis since 2007.

1. Only data published at the time. Each day's reading uses only figures that had been published by that day, never later revisions or knowledge of what came next.
2. Every event listed. We test 31 stress events since 2007 and list every one, including the misses and those caught only after they began.
3. A check for hindsight. A method designed after the fact can flatter itself, so we also ran one with thresholds locked on older data (below).

About these results. They are hypothetical: no reports were sent on these dates. The method was designed in 2026, with knowledge of these events, and a method tuned on past crises tends to look better on them than it will on the next one. Past readings do not guarantee future ones, and no method catches every crisis.

Out-of-sample check. To test for hindsight, we set every threshold by one fixed rule using only 2007–2016 data, locked them, and ran the method on 2017–2026, which the rule never saw. It still showed Caution or Alert in the 30 days before 16 of the 19 events of that period. Caution is common under that rule, about a quarter of days, so a random day has one in the 30 days before about half the time; 16 of 19 by chance would happen less than 1% of the time. Alert is rare, about 2% of days, and 10 of the 19 events reached it. We chose the events, and nineteen is still a small sample.