A composable operating model for XAU/USD research

RM Intelligence treats a gold research workflow as four independent services that talk over internal APIs: data, alpha, risk and execution. Any one of them can be replaced without rewriting the others. This page documents what each service does, what it reads, and where its limits are.

Last reviewed 26 September 2026

Why composable

A monolithic terminal ties data, signals, risk and execution to one screen. A composable design separates them. The pricing model does not care which vendor supplied a real-yield series, and the risk check does not care which model proposed a trade. Three practical benefits follow:

Vendor independence: a data source or execution venue can be swapped behind a stable interface.
Point-in-time honesty: every input carries its own observation time, and stale inputs are flagged rather than silently reused.
Fail-closed behaviour: if a critical input is missing, downstream services refuse to act instead of guessing.

Pillar A — Data ingestion

All inputs are pulled programmatically and normalised with an observation time, a retrieval time, a source, and a freshness state (current, delayed, stale or unavailable).

Input	Source	Role
XAU/USD reference price	gold-api.com live spot, with RM MT5 broker terminals as automatic backup	Live price, risk marks
Gold price history	COMEX GC=F futures candles, scaled to spot by a constant basis	Returns, trend, volatility
US 10-year real yield	U.S. Treasury daily par real-yield curve	Opportunity cost of holding gold
US Dollar Index	ICE DXY daily closes	Pricing-currency effect
Gold implied volatility	CBOE GVZ	Risk premium
Speculative positioning	CFTC Disaggregated COT, COMEX gold, futures only	Context only: weekly and not OTC flow
Inflation	U.S. BLS CPI	Macro backdrop

OTC spot gold has no consolidated global volume tape. The platform never presents COMEX open interest, futures volume or broker tick counts as “global gold volume”.

Pillar B — Alpha engine

The core model is a multi-factor regression on daily changes, estimated over a rolling window of about 250 sessions:

Δln(XAU) = α + β₁·ΔRealYield10Y + β₂·Δln(USD) + β₃·ΔGVZ + ε

Theory predicts β₁ < 0 (higher real yields weigh on gold), β₂ < 0 (a stronger dollar weighs on gold) and β₃ > 0 (more fear, more demand). The model reports each coefficient with its t-statistic and does not force the expected sign. When a relationship weakens, as the real-yield link has at times since 2022, the output shows it.

The signal is the residual. Twenty sessions of unexplained return are summed and standardised into a z-score:

z ≥ +1.5 — rich: gold has outrun what its macro drivers explain; the research bias is mean reversion lower.
z ≤ −1.5 — cheap: gold has lagged its drivers; the bias is catch-up higher.
Otherwise — neutral: no residual edge.

A model fair value is derived by applying the explained 20-session return to the starting price. The daily XAU/USD macro report publishes the current reading.

Pillar C — Risk and compliance gate

Risk runs as its own service. Every proposed ticket (side and size in troy ounces) is scored before it can move on:

One-day parametric value at risk at 95% and 99%, from 60-session realised volatility of daily gold returns.
Notional exposure and an estimated margin requirement.
Net book exposure after the ticket.
Data freshness: a stale live quote or an unavailable model blocks the ticket.
Signal alignment: an advisory flag when the ticket opposes the model's bias, for the review committee to justify.

Hard limits block; they are never overridden by a single click.

Pillar D — Paper execution and review

By policy, RM Intelligence does not route orders. A ticket that clears the risk gate enters a paper book with its signal and risk snapshot, where an owner approves or rejects it. Paper entries and exits are marked against the live reference price, so the journal records how research ideas would have performed without anyone being exposed to them.

The research copilot

Members can question the data through a research copilot built on Anthropic's Claude (currently Claude Opus 5.5), running server-side. The copilot does not rely on memory for anything market-dependent. During a conversation it calls three read-only tools:

Tool	Returns
`fetch_live_xau_usd`	Live XAU/USD reference price, timestamp, freshness and source; optionally the daily change
`get_gold_technicals`	SMA20/50, RSI14, ATR14 and range on 1h, 4h or daily candles, computed at call time
`get_model_signal`	The Pillar B coefficients, rich/cheap z-score and fair value

Answers must cite a source tag for every figure, state when each input was observed, and default to “wait” when evidence is mixed or stale. The copilot cannot place, modify or cancel orders, and it never asks for credentials. A minimal open-source version of the live-quote tool pattern is described below. For the full design (tool routing, the live context pipeline and model choice), see the copilot architecture paper and the copilot capability page.

Governance and audit trail

RM Club's internal trading desk decides its own trades by committee vote. Every proposal, ballot and desk command is sealed into a tamper-evident SHA-256 hash chain. Each entry's hash covers the previous one, legitimate changes are appended as new versions rather than overwritten, and the chain head is published in every daily report, so history cannot be silently rewritten. The chain is public and can be verified with a short script, while record contents stay private. Live health of every component is shown on the system status page. RM Intelligence membership itself is research and decision support: it does not route or execute orders for members.

Public research output follows the same discipline. Every XAU/USD call in the Sealed Journal is frozen, canonicalised and SHA-256 hashed into the chain before its outcome is known, and the verdict is appended when the horizon closes — so a track record cannot be quietly improved after the fact. The public feed exposes each frozen document, its hash and the running scoreboard for independent verification.

Open-source fragment

We publish a small Python package, rm-gold-quote, that fetches a live XAU/USD reference price and exposes it as a Claude tool definition. It is the same pattern the copilot uses, reduced to one tool, for developers who want to build their own. The full, integrated workstation remains part of RM Intelligence membership.

Source code, MIT licence: github.com/srjhoney39-ui/rm-gold-quote

Known limits

Gold history uses COMEX futures scaled to spot, so small basis differences remain.
Public minute data can lag by several minutes; live prices come from the reference feed.
Regression relationships change over time. A factor that mattered last year may not matter now.
Nothing here is investment advice. The platform supports human decisions and does not make them.