For Funds: What this project does

Market Synthesis is a research prototype that uses structured AI debate to identify potential macro regime shifts before they are fully priced.

In plain terms: it is an “early warning + reasoning” layer for portfolio managers and risk teams.

Soros-style market logic

The project follows a Soros-style lens: look for moments where market belief and structural reality diverge.

Typical setup:

Market Synthesis flags the tension before credibility breaks

This is especially relevant in policy-sensitive markets such as FX, rates, sovereign risk, and macro-sensitive equities.

The problem it targets

Most big losses in macro portfolios come from regime breaks that are obvious only in hindsight:

• policy pivots
• liquidity shocks
• currency dislocations
• cross-asset narrative flips

Traditional models often detect these late because they focus on price history.
This system adds a narrative layer from news and macro context, then forces competing views to argue before producing a final score.

How it helps investment teams

1. Pre-mortem signal: highlights events that deserve immediate PM attention.
2. Gap characterization: describes what the market is getting wrong, affected assets, key tension, time horizon, and example actions — not specific trade instructions.
3. Auditability: shows why the model reached the conclusion, including the full Prosecutor/Defender/Expert debate.
4. Workflow fit: designed for weekly committee cadence, not high-frequency trading.

Why this approach is different

Instead of one model giving one opinion, the system uses 8 specialized agents with domain routing:

8-agent orchestration with domain-routed experts — every thesis faces a strong rebuttal before verdict

This structure reduces one-sided model bias and improves interpretability for humans.

Domain specialization

The system routes each event to 1 of 7 specialized experts:

Each expert can be tuned independently — a currency peg specialist knows about reserve adequacy and intervention capacity, while a market structure specialist knows about hidden leverage and margin mechanics.

Why not just use one agent?

A single agent tends to produce a coherent story, but not a rigorous internal challenge.
In markets, that is dangerous: one plausible narrative can hide the opposite trade.

The adversarial setup improves decision quality because:

1. every thesis is forced to face a strong rebuttal
2. domain expertise is matched to the event type
3. confidence is produced after contention, not before
4. output is easier to defend in investment committee discussions
5. failure modes become visible (timing uncertainty vs structural disagreement)

Why this matters in controversial macro episodes

In many high-profile market controversies, price is driven by a fight between:

• policy intent vs policy capacity
• official narrative vs private positioning
• short-term defense vs long-term sustainability

The orchestration is explicitly designed to model that conflict instead of hiding it behind one “final answer.”

Evidence: Feed-Driven Backtest with Adversarial Exit (Real News Data)

The latest validation uses only real news feeds — no hand-crafted events. The system discovers events from cached headlines the same way it would in production, then manages positions with an LLM-based exit system.

What these numbers mean for a PM:

• +7.98% average per trade, ~22% annualized. For each signal, the system enters the category-appropriate instrument and uses an adversarial LLM exit check (“is the gap still open?”) to decide when to close. These are real P&L numbers from yfinance historical prices — verifiable by re-running the backtest.
• Tail risk dramatically reduced. The adversarial exit cuts worst-case drawdown from -57.3% to -19.3% — it recognizes when a thesis has failed and exits before catastrophic loss.
• ~3 in 4 major crises caught. The system misses 7 of 29 hand-curated events (Swiss Franc 2015, Rusal 2018, Repo 2019, Archegos 2021, Terra-Luna 2022, Red Sea 2023, Yen Carry 2024). It’s a decision-support layer, not a replacement for manual coverage.
• Commodity and currency peg are the strongest categories — classic Soros-style trades with +32.8% and +16.8% mean returns respectively.
• Direction, instrument, and exit logic attached to every signal — ready for review, not just a score.

P&L by Category

The system validates against category-specific instruments (FX pairs, commodity futures, bond ETFs, country funds) — not a blanket proxy like VIX.

Practical User Guide for Fund Managers

What You Get

Each signal from Market Synthesis provides:

1. Event name and category — what the system detected
2. Direction — long or short
3. Confidence score (0-100) — how strongly the adversarial debate agreed
4. Entry recommendation — when to enter (immediate, wait, or confirm)
5. Position sizing — based on score and category R:R
6. Stop-loss — category-specific initial stop
7. Hold period — expected time to peak profit
8. Exit triggers — headline signals for partial profit-taking

How to Act on a Signal

Step 1: Receive signal (score >= 70)

Check the category and apply entry rule:

Step 2: Size the position

Reduce sizing for weak R:R categories (real_estate, currency_peg) by additional 40%.

Step 3: Set initial stop-loss

Sovereign: -6%, Commodity: -7%, Central Bank: -7%, Geopolitical: -5%, Real Estate: -8%, Currency Peg: -6%.

Step 4: Manage the position

• Day 5: If profitable, move stop to breakeven
• First headline exit signal (policy response, exhaustion, containment): Reduce 30%
• Signal cluster (2+ signals in 3 days): Reduce another 30%
• Category peak day: Close remaining

Step 5: Watch for resolution signals

If subsequent headlines indicate crisis resolution (bailout approved, intervention), this is a separate long opportunity — but requires full system conviction, not just a headline scan.

What Not to Do

• Do not enter against the system direction on a high-score signal
• Do not ignore stop-losses (max drawdown averages -6% but can reach -10%+)
• Do not hold past the category peak day without a strong reason
• Do not treat exit signals as full-exit triggers — they mean “reduce,” not “close”

Expected Use Inside a Fund

This is best used as a decision support module, not autonomous execution:

• morning/weekly macro risk review
• watchlist prioritization
• scenario discussion for portfolio hedges
• challenge function against house view

What a pilot can look like

Low-risk adoption path — prove value in shadow mode before capital impact

Commercialization path

A practical pilot model:

1. Integrate the signal feed into your existing research/risk workflow.
2. Run in shadow mode against your current process.
3. Measure uplift in hit rate, drawdown reduction, and decision speed.
4. Move to production only after governance sign-off.

Important note

This is research software and does not provide investment advice.
Any live deployment should include risk limits, human oversight, and model governance.