AI & Mathematical Methodology
Published: July 2026 | Sentinel Framework Whitepaper
Sentinel bridges the gap between unstructured news flow and quantitative market execution by coordinating top-down macroeconomic surprise scalars with bottom-up NLP sentiment analysis.
1. Macro Surprise Scalar (St)
Macroeconomic surprises are normalized into standardized scalars measuring deviations from consensus expectations relative to historical standard deviations.
- Formula: S_t = w_static * |(Actual_t - Consensus_t) / sigma_historical|
- Static Event Weight (w): Red = 1.0, Orange = 0.5, Yellow = 0.2.
- Rolling Volatility (sigma): Standard deviation of release surprises over historical 5-year rolling windows.
2. Shock-Adjusted & Confidence-Weighted Sentiment
Combines raw confidence-weighted sentiment scores (s_i in [-1, 1], c_i in [0, 1]) with asset-specific macro beta coefficients (beta_j).
- Formula: Effective Sentiment_{j,t} = S_{raw, j, t} * (1 + beta_j * S_t)
- S_raw: Weighted average compound score generated by fine-tuned Llama-3-Finance models.
- Beta Sensitivity: Asset-specific exposure to macroeconomic release categories.
3. The CARE Prompting Standard & Intent Governance
To guarantee deterministic output, Sentinel enforces the CARE framework (Context, Ask, Rules, Examples). All LLM inference agents must output JSON populated with a mandatory 'reasoning_summary' key before generating numerical scores.
All calculations are validated by the IntentCore Compliance Gateway enforcing a strict 15% portfolio drift limit.
Multi-Agent Governance Architecture
Ingests economic releases and computes normalized macro shock tensors (St).
Aggregates news headline sentiment, corporate QA tension, and 10-K textual inertia.
Stateless gateway node enforcing a strict 15% turnover circuit breaker limit.