How The Silk works

The methodology behind every daily brief: signal detection, probability scoring, cascade analysis, exit rules, and continuous calibration.

The Silk is a daily forecasting and invention journal that uses statistical signals, multi-domain probability scoring, and continuous calibration to identify market opportunities. This page describes the methodology in enough detail to evaluate the daily briefs, not just consume them.

Signal detection

The Sentinel scans an asset's recent price history for z-score breakouts — how many standard deviations the current move sits from its rolling mean. The scoring vocabulary:

• WATCH (1.5σ – 2.0σ): weak momentum signal.
• ALERT (2.0σ – 3.0σ): moderate momentum signal.
• CRITICAL (> 3.0σ): extreme overextension.

Z-score is the primary screen. Subsequent heuristics — domain weights, cascade analysis, confidence cones — refine the signal's interpretation rather than replace it.

Trade direction

WATCH and ALERT signals trade in the direction of the breakout (momentum continuation). CRITICAL signals are excluded by default in ensemble mode: backtest analysis of 2022 – 2025 showed CRITICAL reversal trades are catastrophic in trending bear markets (−126% gross, −132% maximum drawdown). Removing them was the single highest-impact change to live trading discipline.

Domain-weighted probability scoring

Each opportunity's probability is a weighted average across five expert perspectives:

Weights were optimized via Differential Evolution on 473 backtest trades (2026-02-04 calibration). They affect probability assignments only — not signal detection (z-score driven) or price outcomes.

Cascade analysis

Each signal triggers analysis of three cascade layers:

1. Primary — direct price arbitrage (e.g. spot-futures basis trade).
2. Secondary — correlated ripple effects (e.g. downstream assets).
3. Tertiary — derivative innovations (new contracts, structured products, hedging tools).

The cascade is The Silk's structural bet: that interconnected markets propagate signals along discoverable paths.

Confidence cones

Each daily brief carries a confidence label:

• Tight — narrow forecast range; high conviction.
• Medium — moderate uncertainty.
• Wide — broad possibilities; low conviction.

Confidence widens automatically when data sources are degraded (e.g. a fetcher failure surfaced in the brief footer as a Data note) or when the system is in a cold-streak cycle state.

Exit rules

The Silk's equity rule (calibrated 2026-02-25, validated cross-regime): exit the position when any of three triggers fire first.

1. Profit target — 2.5σ favorable move (≈ 2.5 × ATR(9) × √4 from entry).
2. Time stop — close after 4 trading days regardless of P&L.
3. Stop loss — ATR(9)-scaled stop placed at entry. Widened on WATCH (1.25σ vs the original 0.75σ) after stop-rate analysis showed 60% of WATCH trades were stopping out and 31% of those were whipsaws.

Commodities use a tighter uniform 1.5 × ATR(9) stop with no profit target — the strategy's raw signal performance is the test, not the shape of its exit curve.

Continuous calibration

Three automated feedback loops keep the system honest:

• Brier feedback — each forecast's probability is compared to the realized outcome and a rolling Brier score is computed. If rolling Brier drifts above 0.05, weight recalibration is triggered.
• Bayesian drip — yesterday's probability becomes today's prior. The LLM sees the prior on every reforecast, and 7-day slope plus reversals greater than 15 percentage points are flagged.
• Adversarial premortem — every forecast is paired with a premortem that assumes the forecast was wrong and writes a postmortem grounded in out-of-sample win rates with explicit sample sizes ([n=X]).

Heuristic algebra

LLM reasoning across the system is grounded in three operators drawn from the project's heuristic submodule:

• ⊕ Combination (Union) — define interdisciplinary fields by requiring all axioms to hold.
• ¬ Transformation (Negation) — create new models by changing a core assumption.
• ∼ Equivalence (Translation) — map insights across domains by structural analogy.

What this page does not cover

Implementation specifics live in the repository at github.com/hcarstens/The-FIIJ. Historical hit rates and calibration metrics will live on a separate track-record page (forthcoming). For terminology, see glossary. For what markets are covered and at what cadence, see coverage.