Introduction to ZINC Fusion V15
Relevant source files
The following files were used as context for generating this wiki page:
ZINC Fusion V15 is an institutional-grade quantitative forecasting and decision-support system designed for bulk Soybean Oil (ZL) procurement. Developed for US Oil Solutions, the platform utilizes a multi-layer ensemble machine learning pipeline to predict CBOT Soybean Oil futures prices across four specific time horizons: 1 week (5d), 1 month (21d), 3 months (63d), and 6 months (126d).
Sources: README.md:3-8, AGENTS.md:3-8
The system is designed to provide "intelligence only," specifically focusing on probabilistic forecasting to support strategic timing for hedging and purchasing decisions rather than executing trades. It integrates institutional market intelligence, including crush margins, biofuel mandates, and global supply/demand dynamics, to achieve institutional-grade accuracy with a targeted Mean Absolute Percentage Error (MAPE) of less than 3%.
Sources: README.md:12-16, AGENTS.md:5-8, Docs/zinc-fusion-v15-preview.html:776-785
System Architecture
The ZINC Fusion V15 architecture follows a "v3" hierarchical structure comprising 19 distinct models organized into four functional layers. This design ensures that global market context is captured by core predictors while domain-specific nuances are handled by specialized "Specialist" models.
Sources: README.md:21-25, AGENTS.md:22-26
Hierarchical Model Layers
| Layer | Component | Description |
|---|---|---|
| L0 Core | 4 AutoGluon Models | One per horizon (5d, 21d, 63d, 126d) using the full Model Zoo to provide probabilistic multi-step forecasts. |
| L0 Specialists | 11 Signal Generators | Domain-specific models (XGB, GBM, GARCH, etc.) producing horizon-agnostic signals like crush economics, China demand, and volatility. |
| L1 Meta | 4 Stacked Ensembles | Combines L0 Core out-of-fold (OOF) predictions with Specialist signals to refine forecasts per horizon. |
| L2 Calibration | Conformalized Quantile Regression | Produces a truthful outer risk envelope (P10/P90) adjusted for heteroskedasticity. |
| L3 Risk | Monte Carlo Engine | Simulates 5,000+ paths to calculate Value at Risk (VaR), Conditional VaR (CVaR), and barrier probabilities. |
| Sources: README.md:21-42, AGENTS.md:22-30, cubic.yaml:141-160 |
Technical Implementation Flow
The data flow moves from raw data ingestion into engineered features, which then feed the various modeling layers.
The system architecture emphasizes out-of-fold (OOF) integrity to prevent data leakage during stacking.
Sources: README.md:85-95, AGENTS.md:65-75
The Specialist Ensemble (Big-11)
The system relies on "The Specialist Ensemble," a group of 11 domain-specific models trained in isolation to capture specific market drivers. These models act as signal generators rather than direct price predictors.
Sources: README.md:61-75, AGENTS.md:52-64
Specialist Buckets
| Bucket | Focus Area | Key Data/Indicators |
|---|---|---|
| Crush | Processing Economics | ZL/ZS/ZM spreads, NOPA crush margins. |
| China | Global Demand | GACC import flows, MOFCOM policy, CNY rates. |
| Energy | Biofuel Feedstock | Crude oil correlations, heating oil, BOHO ratio. |
| Biofuel | Policy/Mandates | RFS volumes, RIN pricing, 45Z tax credits. |
| Volatility | Risk Regime | VIX levels, CVOL agricultural implied vol. |
| Trump Effect | Political Risk | Executive orders, trade rhetoric intensity. |
| Palm | Substitution | MPOB inventory, Indonesian export levies. |
| FX | Currency Impact | USD/BRL cross-rates, DXY index. |
| Sources: README.md:61-75, AGENTS.md:52-64, Docs/zinc-fusion-v15-preview.html:842-950 |
Data Pipeline and Operations
ZINC Fusion V15 utilizes a Prisma-managed PostgreSQL warehouse across 12 distinct schemas. Data operations are governed by a "2+2=22" doctrine, prioritizing the discovery of intrinsic connections and lead/lag correlations over simple linear relationships.
Sources: README.md:44-55, AGENTS.md:15-20
Operational Rules and Constraints
- Database Isolation: Prisma manages the schema only; runtime queries must use the
pgPool orpsycopg2. - Time-Series Integrity: Every landing table is append-only. Standalone
UPDATEorDELETEstatements on data tables are strictly prohibited to preserve historical audit trails for ML training. - Inngest Orchestration: Serverless functions manage daily ingestion (Yahoo EOD, FRED, Databento) with a strict concurrency limit of 5 per environment.
Sources: AGENTS.md:20-21, cubic.yaml:25-50, cubic.yaml:123-140
Codebase Organization
Sources: README.md:135-145, AGENTS.md:10-14
AI Compute Layer (PULSE)
Beyond traditional ML, the system includes an "AI Compute Layer" utilizing large language models (Claude) to provide soft intelligence. This layer performs sentiment analysis on news and policy events, correcting statistical biases in traditional sentiment models (e.g., recognizing that "production fell" is bullish for commodities, unlike standard FinBERT interpretations).
Sources: src/fusion/ai_compute/agent_pool.py:30-50, frontend/src/lib/ai-intelligence.ts:15-30
The SentimentScorerAgent provides institutional-grade reasoning for the dashboard, mapping events to specific specialist buckets and providing narrative overlays for forecasts.
Sources: src/fusion/ai_compute/agent_pool.py:155-180, frontend/src/lib/ai-intelligence.ts:110-130
Summary
ZINC Fusion V15 serves as a decision-support engine that translates complex commodity data into actionable procurement postures (e.g., BUY, WAIT, HOLD). By combining deep time-series forecasting with domain-specific specialist signals and AI-driven sentiment analysis, the platform enables institutional users to navigate the high-volatility Soybean Oil market with calibrated risk metrics and forward-looking scenario analysis.
Sources: README.md:158-160, Docs/zinc-fusion-v15-preview.html:1034-1045