Methodology

Evaluation-First Design

AIVEX was built with evaluation as the primary concern, not prediction. Every signal module is required to produce a confidence score alongside its output, and every composition step must be reproducible from the same inputs.

This means we prioritize falsifiability: a signal that cannot be evaluated against a ground truth is not a signal — it is speculation. Our backtesting harness runs continuously against historical data to validate that module outputs remain well-calibrated.

Signal Governance

Confidence Scoring

Each atomic signal module produces a scalar confidence value in [0, 1]. Confidence is derived from multiple factors: source reliability, signal agreement across independent feeds, recency of underlying data, and historical calibration error.

Signals below a configurable minimum confidence threshold (default: 0.35) are filtered before reaching the composition layer. This prevents low-quality inputs from polluting the composite signal.

Cooldown Logic

To prevent signal churn, each module enforces a per-symbol cooldown period after emitting a signal. During cooldown, new signals for the same symbol are suppressed unless their confidence exceeds the previous signal by a configurable margin (default: 0.15).

Cooldown periods are adaptive: higher-volatility periods automatically shorten cooldown windows to allow faster signal updates.

Governor Gating

The Governor is the final checkpoint before any signal reaches the API. It applies:

• Hard limits: Symbol-level exposure caps and sector concentration limits
• Kill switch: Global disable flag controllable via environment variable
• Human override: Manually blocked symbols are excluded for a configurable window
• Audit logging: Every gating decision — allow or deny — is logged with full context

Data Quality Principles

AIVEX enforces strict data quality checks at ingestion:

• All timestamps are normalized to UTC with microsecond precision
• Price data is validated against exchange-reported OHLCV bounds
• News sources are scored for historical reliability and deduplicated by content hash
• Stale data (age exceeding configurable threshold) is quarantined, not silently used

Research Disclaimers