L0 Core Models (AutoGluon)
Relevant source files
The following files were used as context for generating this wiki page:
Introduction
The L0 Core Models represent the foundational forecasting layer of the ZINC-Fusion-V15 system. This layer consists of four distinct AutoGluon TimeSeriesPredictor ensemble models, each dedicated to a specific multi-step forecasting horizon: 5 days (1W), 21 days (1M), 63 days (3M), and 126 days (6M). The primary objective of these models is to provide probabilistic price forecasts for Soybean Oil (ZL) futures to support strategic procurement and hedging decisions.
Within the v3 hierarchical architecture, the L0 Core models act as primary estimators that ingest a comprehensive feature matrix. This matrix includes traditional market data (OHLCV), macroeconomic indicators, and high-level signals generated by the 11 domain-specific "Specialist" models. The outputs of these Core models—specifically out-of-fold (OOF) predictions—serve as critical inputs for the L1 Meta-learner layer.
Sources: AGENTS.md:33-40, README.md:14-22, README.md:154-165
Architecture and Ensemble Composition
The L0 Core layer utilizes a "Model Zoo" approach, training an ensemble of 25 models per horizon. The system is restricted to CPU-only training and explicitly avoids presets or time limits in favor of an explicit allowlist of algorithms. AutoGluon automatically selects and weights these models to create a WeightedEnsemble.
Model Zoo Categories
The models are categorized into distinct mathematical and architectural families:
| Category | Description | Included Models |
|---|---|---|
| Baselines | Simple statistical benchmarks | Naive, SeasonalNaive, Average, SeasonalAverage, Zero |
| Statistical | Exponential smoothing and ARIMA variants | ETS, AutoETS, AutoARIMA, AutoCES, Theta, DynamicOptimizedTheta, NPTS, ADIDA, Croston, IMAPA |
| Deep/ML | Deep learning and machine learning architectures | DeepAR, TemporalFusionTransformer, DLinear, PatchTST, SimpleFeedForward |
| Neural | Modern neural forecasting models | TiDE, WaveNet |
| Tabular TS | Tabular-based time series approaches | DirectTabular, PerStepTabular, RecursiveTabular |
| Pretrained | Foundation models (disabled on macOS ARM) | Chronos2, Chronos, Toto |
Sources: AGENTS.md:43-60, README.md:158-163
Technical Configuration
The models are configured with specific parameters to ensure institutional-grade probabilistic outputs:
- Metric: Weighted Quantile Loss (WQL).
- Quantiles: The system focuses on [0.3, 0.5, 0.7] for training and stacking.
- Frequency: Business day (
B). - Validation: 4 expanding windows to ensure temporal integrity.
- Covariates: Uses observed covariates only (no known future values).
Sources: AGENTS.md:62-67, README.md:79-88
Data Flow and Integration
The L0 Core models sit at the center of the data pipeline, consuming engineered features and specialist signals to produce out-of-fold predictions.
This diagram illustrates the flow from the initial training matrix and specialist signals through the horizon-specific AutoGluon ensembles into the OOF prediction tables and model artifact storage.
Sources: AGENTS.md:72-84, README.md:154-165
Training Window Policies
The system applies differentiated training windows based on the horizon to balance signal purity with regime learning:
- Short Horizons (5d/21d): Trained on data from 2020 onwards to prioritize recent signal purity.
- Long Horizons (63d/126d): Trained on data from 2000 onwards to capture diverse market regimes.
Sources: README.md:94-97
Data Model and Registry
The L0 Core models are strictly registered and monitored within the PostgreSQL database using specific schemas.
Schema Ownership
training: Contains the input matrix (training.matrix_1d) and the output out-of-fold predictions (training.oof_core_1d).model: Houses themodel_registry, which tracks model IDs, versions, and performance metrics (e.g., MASE, WQL).forecasts: Stores production forecasts for the various horizons (e.g.,forecasts.production_5d_1d).
Sources: src/fusion/config.py:100-110, grafana/model_registry_queries.sql:19-50
Output Contract (OOF Predictions)
All Core models write to a single out-of-fold prediction table with the following structure:
| Field | Description |
|---|---|
trade_date | The date the prediction was made for. |
horizon_days | The specific horizon (5, 21, 63, or 126). |
p30 / p50 / p70 | Probabilistic quantile predictions. |
target_value | The actual observed value (used for validation). |
trained_at | Timestamp of the model training session. |
run_hash | Unique identifier for the training run. |
Sources: README.md:167-175, grafana/model_registry_queries.sql:102-115
Dependency Management
The L0 Core models rely on specific Python libraries managed via pyproject.toml.
[project.optional-dependencies]
autogluon = [
"autogluon.timeseries>=1.2",
"autogluon.tabular>=1.2"
]
The system also requires scikit-learn, statsmodels, and various deep learning frameworks (implied by AutoGluon's Deep/ML zoo) to execute the ensemble training.
Sources: pyproject.toml:35-36, pyproject.toml:13-28
Conclusion
The L0 Core models provide the probabilistic backbone of ZINC-Fusion-V15. By leveraging AutoGluon's automated ensemble capabilities across four strategic horizons, the system captures both short-term momentum and long-term structural regimes. These models are not standalone; they are integrated into a multi-layer ensemble where their out-of-fold predictions are further refined by L1 Meta-learners to ensure optimal procurement decision support.
Sources: AGENTS.md:10-15, README.md:18-22