Radar — the morning's markets

This morning

One row per instrument, ordered by its score — the number of the five flags up — and on a tie by R5's percentile, the last five days' moves against their own three years. A flag is ● up, ○ down or – undefined where its data is absent (the market's two exist only for the products whose settlements the box pulls). Rows scoring two or more are marked: on the two years phase 34 never read, a flagged day moved far 0.084 of the time at one flag, 0.224 at two and 0.373 at three or more, against 0.10 by construction. The strip is the last thirty mornings, oldest left: a hit is flagged and moved far, a miss moved far unflagged, a flagged day that stayed put, a calm one, and a day whose 24 bars have not closed yet. It says where to look, never which way.

Forward test

loading…

Every instrument-day whose 24 bars have closed, since the first morning the box wrote. A day moved far when its excursion — the furthest the high rose above or the low fell below the 06:00 UTC close over the next 24 bars — was above the 90th percentile of the instrument's own three years: one day in ten by construction. Precision is the share of flagged days (score one or more) that moved far, the pre-registered reading (§7.27 part 6: at least 0.15, phase 34's holdout 0.187); recall the share of the large moves that fell on a flagged day (holdout 0.802).

By score

Score	Days	Large	Share	Holdout

The share of each score's days that moved far on the box, beside phase 34's holdout (2024-09 to 2026-09, docs/models/radar.md in the repository).