Skip to main content

​
Purpose

The Physical Precious-Metals Divergence Index measures stress between Shanghai Gold Exchange physical benchmark prices and the related COMEX futures snapshots. It is a market-structure indicator, not trading advice. The method uses the daily gold and silver premiums from GET /api/market/v1/get-physical-premiums. The public derived response is available from GET /api/market/v1/get-physical-divergence-index and MCP get_market_data.

​
History

The daily seeder stores one list for each metal:
  • market:physical-premium-history:v1:gold
  • market:physical-premium-history:v1:silver
Each point is keyed by the physical print date. A repeated print date replaces the earlier point. The write, deduplication, and trim run as one Redis operation. Each list keeps at most 750 points, or about three trading years. The classifier uses the newest 250 points and needs at least 60 valid points. Every history point and response declares methodologyVersion: physical-divergence-v2.

​
Robust normalization

For the current premium x, the robust z-score is:
When MAD is zero, the score is 0 if the current value equals the median. Otherwise, the score is unavailable. This rule prevents division by zero and does not invent scale. The percentile rank is the share of window values less than or equal to the current premium. It is reported on a 0 to 100 scale.

​
Hybrid regimes

The classifier assigns one absolute regime and one relative regime. The higher regime wins. Relative thresholds apply only once the premium clears a magnitude floor of half the metal’s elevated floor — 0.5% for gold, 2.5% for silver. A sign test is not enough: the current print is part of its own trailing window and the percentile is inclusive, so any new window high scores 100 regardless of size. Without the magnitude floor a trivially small positive premium reads as extreme purely for topping a calm window. The floor is set below the elevated threshold rather than at it so that the 80th-percentile rung stays reachable. Gating at the full elevated floor would not disable relative escalation altogether — the 95th and 99th rungs could still lift an absolutely elevated premium to stressed or extreme — but it would make the 80th rung unreachable, because any premium clearing that gate is already elevated on absolute size alone. Half is the chosen point in that range, not the only one that satisfies the constraint. For example, a gold premium of 3% is at least stressed even when the trailing window also contains higher observations. The per-metal stress index is on a 0 to 100 scale. Absolute magnitude interpolates across compressed band tops of 45 (elevated), 70 (stressed), and just under 90 (approaching extreme; the stressed span stops short of 90 so two-decimal rounding cannot reach the extreme floor). Clearing the absolute extreme premium floor (gold 5%, silver 20%) publishes 100. When the relative ladder outruns absolute magnitude, index still means stress magnitude with a regime-ordered floor, not a second copy of the percentile: index = max(absoluteStressIndex, regimeFloor). A relative-only extreme therefore floors at 90 rather than saturating at 100. That reserves the top of the scale for absolute extreme premiums and keeps the index monotonic in regime rank: no stressed reading can report a higher index than any extreme reading. regime still carries the hybrid classification (absolute or relative, whichever is higher). Percentile remains a separate published field. This follows the ECB CISS lesson that a single-indicator percentile must not pin the published index at its maximum. Option chosen over dropping the floors entirely (which would let regime: extreme ship with a low magnitude index) and over widening the proto with a second index field. The composite is 70% gold + 30% silver. It is published only when both metals are in the ok state. The gold weight reflects its larger and more liquid benchmark role. The response always includes both weights so consumers can reproduce the result. The 5-day and 20-day changes are the current premium minus the premium 5 or 20 observations earlier.
  • A change greater than 0.01 percentage points is widening.
  • A change below -0.01 percentage points is narrowing.
  • Other changes are stable.
These are observation counts, not calendar-day offsets.

​
Explicit data states

Every per-metal result has one state. The checks run in this order:
  1. missing_input: no current physical-premium input exists.
  2. stale_input: the physical print date is more than 12 calendar days old, the daily COMEX cohort is more than 36 hours old, or the FX snapshot is more than 60 hours old. The physical threshold tolerates long scheduled Chinese market closures. The paper threshold matches the daily 08:00 UTC publisher with deployment and schedule jitter, while the FX threshold matches its source-health budget.
  3. insufficient_history: fewer than 60 valid history points exist.
  4. ok: the input is current enough and history is sufficient.
A 9-day-old carried-forward print is accepted for a market closure. A 13-day-old print is stale. Non-ok results omit the numerical index. The composite also fails closed when either metal is not ok.

​
Transition signals

The cross-source signal stream emits a physical-premium regime transition only when:
  • the previous and current states are both ok;
  • the regime changed; and
  • the same metal has not emitted a transition in the last 48 hours.
The 48-hour cooldown is exactly two times the daily seed cadence. Missing, stale, or warming data never creates a transition signal. Downward regime transitions are retained because normalization can also be material.

​
Provenance and operations

The response carries the physical benchmark source, symbol, physical print date, COMEX snapshot time, FX snapshot time, history key, sample count, window size, and methodology version. /api/health monitors the derived snapshot under market:physical-divergence:v1 with its own seed metadata and activation marker.

​
Version history