L0 Specialists (The Big 11)
Relevant source files
The following files were used as context for generating this wiki page:
L0 Specialists, collectively known as "The Big 11," are domain-specific signal generators within the ZINC-Fusion-V15 architecture. Unlike the L0 Core models which provide multi-horizon forecasts, specialists are horizon-agnostic engines designed to quantify pressure and regime shifts within specific market verticals. They act as expert quantitative analysts, translating complex raw data into refined signals used as high-alpha input features for the core ensemble.
Sources: AGENTS.md:41-43, README.md:21-23
Architecture and Taxonomy
The Specialist layer consists of 11 distinct buckets, each utilizing a custom-built model architecture tailored to its specific domain. These range from statistical models like GARCH and VAR to machine learning frameworks like Gradient Boosting and Random Forests.
The Big 11 Specialist Buckets
| Bucket | Model Type | Domain Focus | Primary Signal Output |
|---|---|---|---|
| Crush | gbm | Soybean complex fundamentals | Crush margin z-score + momentum |
| China | gbm | Chinese import demand | Demand outlook + Brazil competition |
| Substitutes | rf | Competing vegetable oils | Substitution pressure + richness |
| FX | ardl | Currency dynamics | FX pressure index + carry |
| Fed | ridge | Monetary policy | Rates regime + change |
| Tariff | tree | Trade policy risk | Tariff risk + EPU spike |
| Energy | var | Petroleum complex | Energy spillover + momentum |
| Biofuel | nlp_ema | Renewable mandates | Policy pressure + trend |
| Palm | ecm_ridge | Palm oil complex | Cointegration + mean reversion |
| Volatility | garch | Market stress/fear | Conditional variance z-score + regime |
| Trump Effect | event_study | Political event risk | Intensity + volatility impact |
Sources: AGENTS.md:52-64, README.md:105-116, scripts/generate_specialist_features.py:65-330
Data Flow Logic
The specialist pipeline is distinct from the Core "all data" policy. Specialists consume hand-picked features (typically 20-50 per domain) to ensure expert curation and prevent noise contamination.
The diagram represents the specialized ingestion path where domain experts filter relevant features before signal generation.
Sources: scripts/generate_specialist_features.py:12-25, README.md:300-310
Signal Contracts and Output
Every specialist adheres to a strict output contract. They do not produce price forecasts; instead, they generate pre-quantified intelligence written to the training.specialist_signals_1d table.
Output Components
- Signal 1 (Required): The primary directional or magnitude indicator for the domain.
- Signal 2 (Optional): Secondary context or regime indicator.
- Confidence (Optional): A value between 0.0 and 1.0 representing the model's conviction based on evidence quality.
Sources: AGENTS.md:52, README.md:101-103, src/fusion/pulse/contracts/prompts/system.txt:65-70
Quantitative Methodology
Specialists utilize institutional-grade quantitative tools to derive these signals:
- Regression Analysis: OLS, Ridge, and Quantile regression for factor extraction.
- Cointegration: Johansen and Engle-Granger tests for mean-reversion (specifically in the Palm bucket).
- Volatility Modeling: GARCH family models for regime detection.
- Event Study: Methodology used for political and announcement effects (specifically Trump Effect).
Sources: src/fusion/pulse/contracts/prompts/system.txt:30-45, AGENTS.md:52-64
Feature Curation Policy
The "Surgical Feature" policy ensures specialists remain experts in their respective fields without being overwhelmed by unrelated macro noise.
Domain-Specific Feature Sets
- Crush: Focuses on soybean complex fundamentals including
board_crush(ZL11 + ZM22 - ZS*50),oil_share, and NOPA utilization. - Energy: Consumes petroleum complex data like
cl_close(WTI),ho_close(Heating Oil), and theboho_spread. - Volatility: Monitors stress indices (
VIXCLS,OVXCLS), realized volatility, and the volume of put/call skew. - Trump Effect: Uses binary regime flags (
trump_in_office), executive action counts from the White House, and Trade Policy Uncertainty (TPU) indices.
Sources: scripts/generate_specialist_features.py:68-80, scripts/generate_specialist_features.py:125-140, scripts/generate_specialist_features.py:280-300
Integration and Monitoring
Specialist signals are integrated into the broader ensemble as high-alpha covariates for the L0 Core models and L1 Meta-learner.
Monitoring Infrastructure
Monitoring is handled via generated Grafana dashboards that track:
- Prediction Coverage: Percentage of actual outcomes falling within the specialist's projected bands.
- MAE (Mean Absolute Error): Accuracy tracking for specific horizons (e.g., 21d MAE).
- Dissent Index: A metric calculating the standard deviation between the Big 11 signals to identify periods of specialist disagreement.
- Data Freshness: Tracking the age of input sources like FRED, CFTC, and BMD.
Sources: grafana/scripts/generate_specialist_dashboards.py:180-250, scripts/generate_synthesis.py:230-250
The sequence shows how specialist signals propagate through the hierarchical ensemble to inform the final weighted prediction.
Sources: AGENTS.md:45-50, README.md:118-125
Conclusion
The Big 11 Specialists represent the "intelligence layer" of ZINC-Fusion-V15. By isolating domain expertise into specialized models, the system avoids the overfitting risks of massive monolithic models while providing the Core predictors with refined, high-alpha signals. This architecture allows the system to identify regime shifts and non-linear interactions—such as the "Trump Effect" or "Crush Squeezes"—before they manifest in aggregate price movements.
Sources: README.md:55-60, src/fusion/pulse/contracts/prompts/system.txt:10-20