Adversarial Exit System

Core principle: debate decides all exits

The exit system applies the same adversarial debate structure used for entry to the exit decision. There are no mechanical P&L gates, no time limits, and no arbitrary thresholds. The debate determines whether the narrative-reality gap that motivated the trade is still open.

The only exception is the volatility-based stop-loss (3-sigma of trailing 5-day returns), which acts as a mechanical downside bound regardless of debate verdict. This is the single concession to risk management that overrides the debate.

Why debate-driven exits work

Mechanical exit systems (time-based, P&L-based) suffer from a fundamental problem: they apply the same rule to structurally different situations. A commodity supply disruption may take 6 months to fully reprice. A currency peg break may complete in 2 weeks. A time limit that works for one destroys the other.

The adversarial exit system solves this by asking the right question each week: “Is the narrative-reality gap that motivated this trade still open?”

In the v38 backtest, this approach:

• Holds winners longer when the gap is widening (commodity trades averaging +23.5% mean)
• Cuts losers earlier when evidence turns against the thesis
• Bounds worst-case to -19.4% via the vol stop-loss

Exit debate structure

Exit system: weekly adversarial re-debate with continuous vol stop-loss as the only mechanical override

Exit debate roles (reversed from entry)

In the exit debate, the roles invert from entry:

This inversion means the same agent that was sceptical at entry becomes the advocate for exiting. The intellectual honesty of the debate is maintained because the defender-at-entry (who argued against the trade) now argues FOR closing it — they are consistent in their scepticism about the thesis.

What the exit debate considers

Each weekly exit review receives:

1. Recent headlines from the same news sources (Guardian, NYT, GDELT) — has the narrative shifted?
2. Price action since entry — is the instrument moving as predicted?
3. Market data — the same signals used by entry gates (TMV momentum, ZB=F returns, SPY drawdown) as context
4. Original thesis — the entry debate verdict and reasoning, so the exit debate can assess whether the original conditions still hold

The adjudicator must determine: “Given the original thesis and current evidence, is there still a meaningful gap between narrative and reality that has not been priced?”

Volatility-based stop-loss

The stop-loss is the only mechanical override in the system. It operates on a simple principle: if price moves against the position by more than 3 standard deviations of the trailing 5-day return distribution, exit immediately regardless of debate verdict.

This captures genuine regime breaks — moments when the market has new information the debate cannot incorporate in time. In the v38 backtest, the stop-loss bounded the worst trade to -19.4%.

Why 3-sigma and 5 days

• 5-day window captures recent volatility regime without being dominated by a single outlier day
• 3-sigma threshold is wide enough to avoid triggering on normal noise but tight enough to catch genuine adverse moves
• The combination means: “this price move is statistically abnormal given recent conditions — something has changed that the weekly debate cycle cannot address in time”

Examples of exit debate verdicts

Example 1: Commodity hold (oil supply disruption, 2022)

Entry thesis: OPEC supply cuts creating physical shortage not reflected in futures curve.

Week 4 exit debate:

• Prosecutor (hold): “Physical spreads still in deep backwardation. No SPR release announced. Demand unchanged.”
• Defender (exit): “Crude has rallied 15% since entry. Move may be complete.”
• Expert: “Backwardation structure shows physical market still tight. 15% move reprices spot but the 6-month curve still implies surplus that will not materialize.”
• Verdict: HOLD. Gap between curve pricing and physical reality persists.

Week 9 exit debate:

• Prosecutor (hold): “Curve still implies surplus beyond month-3.”
• Defender (exit): “SPR release announced. OPEC signaling willingness to increase quotas. Physical spreads narrowing.”
• Expert: “The structural driver (supply restriction) is reversing. Gap between narrative and reality is closing from the reality side — supply is actually increasing.”
• Verdict: EXIT. The structural impossibility that created the gap is resolving.

Result: +18.3% over 9 weeks.

Example 2: Geopolitical early exit (sanctions trade, 2018)

Entry thesis: Russia sanctions creating supply disruption in aluminium not reflected in base metal prices.

Week 2 exit debate:

• Prosecutor (hold): “Sanctions still in place. Supply disruption ongoing.”
• Defender (exit): “US Treasury announced 6-month wind-down period. Market interpreting as softening of regime.”
• Expert: “Wind-down period gives counterparties time to restructure. The immediate supply shock thesis is undermined — disruption will be gradual, not acute.”
• Verdict: EXIT. Policy response removed the forcing function.

Result: +7.1% in 2 weeks (captured the initial shock, exited before reversal).

Example 3: Stop-loss trigger (sovereign debt, 2019)

Entry thesis: Argentine peso under reserve pressure, CDS widening.

Week 1: Surprise primary election result causes 30% peso devaluation overnight — in the OPPOSITE direction of the trade instrument selected (the system was positioned for gradual depreciation via a different expression).

Vol stop-loss triggered: 5-day return exceeded 3-sigma. Position closed mechanically.

Result: -14.2% (bounded by stop, would have been -31% without it).

What the exit system does NOT do

To be explicit about design choices that were tested and rejected:

1. No trailing stops. Trailing stops exit winners prematurely in volatile categories. The debate is a better judge of whether the thesis is intact.

2. No time limits. A good commodity trade can run for months. A currency peg break can complete in days. Time limits destroy one category to help another.

3. No profit targets. Taking profit at +X% leaves money on the table when the gap is still wide. The debate holds through further repricing.

4. No drawdown gates. A -10% unrealized loss does not mean the thesis is wrong. The debate evaluates whether the gap has closed, not whether the current mark-to-market is uncomfortable.

5. No correlation-based exits. The system does not exit because “the market is going up” or “vol is low.” It exits when the specific narrative-reality gap that motivated the trade has closed.

Each of these was tested in iterations v11-v30 and removed because they degraded overall performance. The recurring pattern: mechanical overrides that help one category hurt another. The debate, by contrast, adapts to the specific situation.

Performance impact

Comparing adversarial exit to mechanical time-based exit (hold for fixed category-specific durations):

The adversarial exit system produces both better returns AND lower risk — it is not a tradeoff between the two. This is because it holds winners longer (when the gap is widening) and cuts losers earlier (when evidence turns against the thesis), rather than applying a single time rule to both situations.

For pipeline architecture, see Methodology. For full backtest results, see Findings.