The Self-Learning Macro Engine: How DEPTH4 Calibrates Conviction, Verifies Room, and Evolves (2026)
Most macro AI summarizes the news and stops. DEPTH4 now learns from its own resolved theses — calibrating conviction by mechanism, verifying room with market math, retrieving past wins and misses at generation time, and evolving its writer playbook under human review.
Published 2026-07-04. Last updated 2026-07-04. · 10 min read
Why static macro AI stops being useful
Macro markets punish the same failure mode repeatedly: analysis that sounds informed but never updates when the engine is wrong. A headline summary does not know it called three similar setups incorrectly last quarter. A generic "geopolitical risk elevated" note does not down-weight conviction when real-yield trades have stopped working for that mechanism.
DEPTH4's July 2026 machine-learning stack treats the engine as a system that should know its own edge, verify room with external data, learn from structural examples, and evolve its playbook — without auto-publishing untested prompt changes. The product promise on the welcome page — Headlines → Math → Signal → Evolve — is now backed by four production loops described below.
See the live product framing on the welcome page and the headline vs thesis quality filter — this article covers the engine upgrades behind that positioning.
Four learning loops now running in production
Each loop answers a different question traders implicitly ask before sizing a macro position: Has this setup worked before? Is room real? What did a good version look like? What rule should we add so we stop repeating the same mistake?
1. Mechanism calibration — it knows its own edge
Every AI thesis is tagged at generation with a mechanism slug (for example: central_bank_pivot, safe_haven_bid, cta_trend_unwind, supply_disruption). When a new thesis is written, the engine looks up how similar mechanisms performed over the last 90 days on resolved trades. Weak historical win rates trigger a conviction clamp — the model may still draft bullish language, but published conviction cannot exceed what the track record supports. This is enforcement, not a suggestion in the prompt.
2. Quantitative room — real math, not adjectives
Blind spot / room is how much repricing the market has not yet absorbed. Where options data exists, DEPTH4 compares scenario-implied move budgets against current pricing instead of trusting LLM guesses. For FX and commodities without liquid options chains, realized-volatility proxies provide a floor on quantitative sizing. Room and conviction are orthogonal: a thesis can be high-conviction but mostly priced in, or moderate-conviction with wide room if the market is still behind the cascade.
3. Few-shot RAG — learns from resolved examples
Aggregate win rates are not enough. Before writing a new thesis, the engine retrieves one resolved success and one resolved failure for the same mechanism slug — depth-matched to where the trade edge lives (D2, D3, or D4). The writer sees structural shape: how causality was stated, how timing was framed, what invalidated the loser. Rules forbid copying tickers or paraphrasing headlines; the goal is mechanism discipline, not template plagiarism.
4. APO — evolves the playbook under human gate
Automated Prompt Optimization closes the loop at the prompt layer. A weekly metaprompter reviews recent invalidated and stopped-out AI theses, proposes exactly one new heuristic for the Step 4b writer, and stores it in an admin approval queue. Nothing auto-merges to production. Approved rules inject at runtime; golden regression fixtures must pass before a rule ships. RAG teaches instance memory; APO teaches durable guardrails.
How the loops fit the engine pipeline
The operational map is unchanged at the product layer: Read headlines → trace the cascade → surface SIGNALS with direction, blind spot, and thesis backing. What changed is what happens inside Stage 3 thesis intelligence before anything user-visible publishes.
Step 4a still structures the causal chain and assigns mechanism slug + primary edge depth. Step 4b now receives three memory layers: aggregate mechanism performance, few-shot resolved examples, and any admin-approved APO heuristics. After the draft, scenario clamps and mechanism calibration clamps run in series — numeric gates, not vibes. Room repricing runs through the depth book and quantitative tiers where market data exists.
Downstream, SIGNALS conglomeration, tradability audit, and profitability audit are unchanged — they remain algorithmic quality gates between model output and the fleet card you see on /theses/signals.
For cascade definitions, see D1–D4 cascade help and why most tools stop at D2.
What traders see — and what they should not expect
On the SIGNALS fleet, direction, blind spot, conviction, and contributing theses update as evidence arrives. Early-call examples on the welcome page (Gold flip, Silver pressure, NZDUSD funding link) illustrate timing — surfacing a read before the chart crowd catches up, not claiming omniscience.
The engine will still publish MIXED and Quiet rows when thesis distribution is split or thin. Learning systems should get sharper; they should not get louder. When mechanism history is sparse, calibration fail-opens — the engine does not fabricate win rates to sound authoritative.
DEPTH4 remains complementary to Bloomberg, TradingView, and execution platforms. It is the macro intelligence layer that names why a chart may move before the move finishes printing — not a broker, not personalized advice.
Track live SIGNALS after signup, or browse resolved theses that feed calibration and RAG pools. Compare terminal and charting layers in best macro signal tracking tools (2026).
Further reading
- Macro headline vs macro thesis
Five fields that separate a tradable thesis from a headline rewrite.
- Why most market intelligence stops at D2
Where news aggregators and AI summaries run out of cascade depth.
- What is macro intelligence? (2026)
Definition, platforms, and where DEPTH4 sits in the stack.
- Macro intelligence reference guide
D1–D4 framework, room, and SIGNALS field definitions.
- Help: D1–D4 cascade
Product documentation for cascade depth and timing windows.
- Resolved theses
Public archive of closed theses that feed calibration and RAG pools.
Related: How macro insights improve investment decisions · Macro signal tracking for retail traders · How geopolitical risk becomes a trade
Frequently asked questions
- What does "self-learning" mean for DEPTH4?
- It means the engine uses its own resolved thesis history and market data to calibrate new outputs — mechanism win rates clamp conviction, quantitative tiers verify room, few-shot examples shape new drafts, and approved prompt heuristics accumulate over time. It does not mean unsupervised auto-trading or auto-editing prompts without human review.
- How is mechanism calibration different from few-shot RAG?
- Calibration is aggregate and numeric: "this mechanism hit target 42% of the time in 90 days" → conviction cap. RAG is instance-level and structural: "here is what a successful safe_haven_bid D3 thesis looked like vs a failed one." Both feed Step 4b; calibration enforces after the draft.
- Does DEPTH4 auto-rewrite its prompts when trades fail?
- No. APO proposes one rule per week from recent failures. Admins approve or reject in the proposal queue. Approved rules inject at runtime and must pass golden regression replay before production use. The engine learns; humans retain veto.
- What is blind spot / room in plain terms?
- Room is how much of the expected move the market has not priced yet. A 31% blind spot means roughly a third of the scenario-implied repricing may still be ahead, subject to data tier and liquidity. It is an engine estimate for research, not a profit guarantee.
- Will calibration block thesis generation when history is thin?
- No. When sample size is below threshold or slug is unclassified, calibration fail-opens — generation continues without fabricated statistics. Trust-preserving behavior: empty or thin history does not become fake confidence.
DEPTH4 is a macro analysis and information tool, not personalized investment advice. It is not a broker and not a registered investment adviser. All signals, theses, and room estimates are research outputs for informational purposes only.
Try DEPTH4
Live D1–D4 cascades, calibrated conviction, quantitative room, and SIGNALS that evolve as evidence arrives.