P1 · The landscape — pair × timescale matrix
Rows = the 15 pairs (full name + premise-class chip). Columns = the canonical
window set in two groups: window direction (90m·4h·1d — Pearson-free "both up / both
down" read; too few bars for a meaningful correlation) and rolling corr n≥24
(1w·1m·1q·1y·5y). Cell colour = premise ALIGNMENT for pairs that HAVE a premise:
blue = moving PER the pair's premise,
red = COUNTER-premise; intensity =
correlation strength / joint-move rarity. NONE / REGIME pairs have no premise, so their
cells instead show the RAW correlation SIGN on a neutral ramp
(teal = positive /
brown = negative) — NOT premise alignment.
Colour scale note: this ramp is a FIXED semantic (with-premise vs counter-premise / corr
sign) that means the SAME across every timescale column, so it is deliberately NOT re-scaled
per column — a counter-premise cell reads counter at 90m and at 5y alike. (Per-column magnitude
heat is used where it is a magnitude claim — see P4 Extremes and P4b Co-move degree, each column
hottest→coldest on its own; every column here is still click-sortable via the header row.)
Cell text = correlation (corr cols) or the two legs' window return "±x%/±x%"
(direction cols). Superscripts: Nd = days since a corr decouple onset · ☉ =
co-extreme (both legs stretched the same way) · ▲ = extreme correlation ·
r = ramping. Grey ╱ hatch = HK/US sessions don't overlap
(structurally n/a); amber "?" hatch = feed unavailable. Hover any cell for the dates + numbers.
Click a cell to drive the pair detail below.
sort:
building matrix…
P2 · All-pair ratios + selected-pair premise
Every pair drawn as ONE line = leg A ÷ leg B, rebased to the window start
(unit-free — works across price/price, price/index, price/yield%): a FLAT line = the two
move together (converging); a TRENDING / steep line = decoupling. Toggle indexed-ratio ↔
z-score-of-ratio. Click a line (or a matrix cell) to SOLO that pair and show its premise, class,
flag state, onset date and live elapsed timer in the card. Nothing premise/onset is lost — it's
reorganized around the ratio view.
click a ratio line or a matrix cell above…
loading pairs…
Select allDeselect all
P3 · Seven-instrument overlay + stretch tape
The overlay (precise, chaotic): all 7 instruments normalized vs NAS100 (SPX is a
member line). 4 view modes — z-score · force full (fit) · % from beginning ·
individual % — to eyeball any pair's divergence (gold vs oil, oil vs spx, 30Y vs gold…).
The tape below (glanceable): every instrument row vs its OWN history on one diverging ramp — 6
coloring lenses. Both perspectives are mandatory together (arch law §1.10).
P4 · Extremes board — single-instrument reads
Per instrument × scale: 20-bar Bollinger z (how many σ from its 20-bar mean),
17-EMA deviation %, RSI-14, and own-range level percentile. All computed reads with the sample
depth n; no bare word-flags. Below, the CO-EXTREME rail lists pairs whose BOTH legs are
stretched the same way at the same scale — the "gold AND stocks both above the top band" tell.
loading…
Co-extreme rail
…
P4b · Co-move degree board — recorded magnitudes
Each instrument's signed move (% · yields in bp) over 12 dated windows,
plus how many moved the SAME way per window (breadth) and three named baskets. This states
WHAT moved, HOW MUCH, WHICH tickers, over which dated windows — arithmetic on real bars.
It is a MAGNITUDE display, not a rarity claim: whether a broad joint move is unusual needs
a banked co-move baseline that does not exist yet (see the ramping chip). NAS100 (^NDX) is shown
as a reference row but never counted in any breadth or basket (it overlaps the S&P). Hover any
cell for its dated window + source; click a breadth window to focus the grid on that window's
majority; pick a window for the basket lines below.
computing co-move degree…
Breadth — same-direction count per window (all 8 instruments)
building grid…
Basket rail — pre-registered co-move descriptors
window:
P5 · Flag feed
A dense grid of compact flag tiles — EVERYTHING-RALLY (≥3 instruments
stretched the same way) pinned + flashing on top, then decouple (premise violation), extreme
correlation, co-extreme — freshest onset first, each one tight line: type · pair/instruments ·
scale · number · onset · timer · n. Fresh co-moves flash. "All holding" when nothing fires.
…
P6 · Methodology & honest limits
What this isa counter-trend / decoupling detector on a fixed 7-instrument universe (incl. WTI oil); the unit is the PAIR (21) across 8 timescales, premise printed, deviation dated, persistence timed.
Premise-firsteach pair carries a hardcoded, curated prior (the anti-data-mining spine). The signal is the DEVIATION from it — not a fitted pattern.
Two readsdirection (Pearson-free "both up/down" for short windows) and rolling correlation (only where n≥24 return bars).
Everything-rally≥3 instruments simultaneously stretched the same way at a scale raises one flashing banner (named legs + since-date) — a decomposable read over the per-instrument extremes, no hidden composite.
No aggregatesthe matrix is per-pair by construction; every cell decomposes to two named legs on one click. No composite line anywhere.
Honest gapssession-n/a ≠ quiet ≠ ramping ≠ unavailable — each is shown distinctly.