Tradability Analysis: Entry, Hold, Exit

Market Synthesis is validated on detection accuracy. Whether a fund manager could profit from these signals is a second, much harder question, and this page now documents why the answer is not yet demonstrated.

This page describes the trading algorithm derived from backtesting 303 signals (2014-2026) against real market data via yfinance, using an adversarial LLM-based exit system. The algorithm is real and runs in paper trading. Its edge is not established.

Correction (2026-08-25): the exit system does not beat a random exit

This page previously led with “Key Finding: Adversarial Exit Beats Mechanical Rules” — mean P&L +7.98% against +4.86% for a fixed hold, and an estimated ~22% p.a. That comparison was not a control, and the claim is withdrawn.

Beating a fixed hold is the wrong test. A debate that exits early to cut a loss beats a fixed hold partly by having a stop, which says nothing about whether it picked a good day. The control that answers the actual question is:

given that a trade was exited on day D, is D better than a random day drawn from the same distribution of realised exit days?

That arm was run on 2026-08-25 over 116 trades, recomputed from the price cache:

Edge: −1.23pp. Only 2 of 5 categories with n≥8 beat their own random-exit control. Across the full available range of exit days, the debate captured 34% of the span between the best and worst achievable exit; a random exit captured 37%.

The debate’s exit timing is not adding information. The per-category “alpha vs hold” table that stood here is a comparison against the weaker alternative and has been removed rather than annotated.

Six other execution-layer claims have failed their own controls (base-rate and date-shift placebo tests), including a withdrawn 24.5% CAGR portfolio simulation, the commodity strangle (39.3% win rate against a 50.6% base rate), and the central-bank strangle (0 wins in 16). The detection results elsewhere on this site are unaffected — they are measured against labelled outcomes, not P&L.

What this does not change: detection precision holds ~0.75 out of sample across 12 years and five macro regimes with no look-ahead, and commodity detection reads 0.974 precision on 39 signals. The signal is real. Converting it into return is the open problem.

Entry Timing Algorithm

Entry timing is category-specific. Fast-repricing events need immediate entry; slow-building events benefit from confirmation.

Category-specific entry timing — derived from backtest of 4 strategies on each category

Entry Rules by Category

What Does NOT Help for Entry Timing

Tested and rejected:

• Volume spikes: Only 33% correlation with optimal entry. Volume peaks on day 1 (headline splash), not at the drawdown trough.
• VIX reversal: Marginal improvement for sovereign_debt only. Not worth the complexity.
• RSI / oversold signals: Insufficient data, unreliable on macro events.

Position Management Algorithm

Position management: time-based exit with headline-driven partial profit-taking

Category Parameters

Adversarial Exit System

The exit system is an LLM-based weekly check that asks: “is the narrative-reality gap that justified this trade still open?”

Adversarial exit: LLM weekly check replaces mechanical time-based exits

How it works

Each week, the exit detector receives:

1. The original signal (thesis, category, direction, gap summary)
2. Current news headlines from that week
3. Current price data for the instrument

It then decides whether the gap is still exploitable. This produces two key behaviors:

• Holds winners longer — if a commodity supply disruption is still unresolved, the system keeps the position open past the mechanical hold period
• Cuts losers earlier — if a central bank successfully defends, the system exits before the stop-loss is hit

Exit verdict distribution (303 signals)

The rationale for adversarial exit — and what testing it showed

The design argument was that mechanical exits (fixed hold + stop-loss) fail in two ways:

1. Exit too early on winners where the thesis is still playing out (commodity supply disruptions can last months)
2. Exit too late on losers where the gap has already closed (central bank successfully defended)

An LLM exit check should solve both, because it reasons about whether the thesis is still valid rather than only whether price moved.

Measured, it does not. Against a random exit drawn from the same distribution of realised exit days, the debate exit is 1.23pp worse on median return (see the correction at the top of this page). The reasoning is plausible and the mechanism does fire — 45 of 303 signals exited early — but the days it chooses are not better than chance. The argument above is why the system was built this way, not evidence that it works.

Instrument Selection

The system uses a 3-layer instrument selector:

1. Disk cache — previously validated ticker assignments (living map, updated as results come in)
2. Static registry — category × region → default instruments (e.g., sovereign_debt + Europe → EUFN, TLT)
3. LLM fallback — Sonnet adjudicates the best tradeable expression for novel signals, with regression testing before cache update

Category instruments

The per-category “avg P&L” column that stood here (commodity +32.8%, currency peg +16.8%, and so on) is the same uncontrolled exit-vs-hold measurement corrected at the top of this page, so it has been removed rather than restated. Each of those figures is an average against no null; where a base rate has since been computed, the edge disappeared — the commodity strangle reads 39.3% win rate against a 50.6% base rate.

Implementation

The adversarial exit backtest is implemented in tests/backtest_adversarial_exit.py:

# Run the full backtest
python -u tests/backtest_adversarial_exit.py --workers 5 --label v10

# Key components:
# - Adversarial exit detector: detectors/adversarial_exit_detector.py
# - Instrument selector: shared/instrument_selector.py (3-layer with cache)
# - Signal store: shared/signal_store.py (635 signals, append-only)

The entry timing algorithm is implemented in shared/entry_timing.py:

from shared.entry_timing import get_entry_recommendation

rec = get_entry_recommendation(
    category="sovereign_debt",
    score=85,
    direction="short",
    pre_move_pct=0.0,
)

# Returns: EntryRecommendation(
#   entry_type="momentum_confirm",
#   momentum_days=2,
#   stop_loss_pct=-6.0,
#   position_size=0.75,
#   hold_days=23,
#   peak_day=25,
# )

All metrics derived from feed-driven backtest on real news data (2014-2026), validated against yfinance market prices. 303 signals backtested with adversarial LLM exit system. No hand-crafted test events.