L1 Meta Ensembling

Relevant source files

The following files were used as context for generating this wiki page:

• README.md
• AGENTS.md
• scripts/generate_synthesis.py
• cubic.yaml
• src/fusion/config.py

L1 Meta Ensembling represents the second layer of the ZINC-Fusion V15 hierarchical ensemble architecture. Its primary objective is to act as a "meta-learner" that determines the optimal weighting and trust levels for various base models across different time horizons. By combining the probabilistic outputs of the L0 Core models with the domain-specific intelligence of L0 Specialists, the L1 layer produces a unified, high-conviction forecast that serves as the foundation for downstream calibration and risk analysis.

The system utilizes a stacked ensembling approach, where four distinct meta-learners are trained—one for each target horizon (5d, 21d, 63d, and 126d). This ensures that the ensembling logic is specifically tuned to the unique market dynamics and signal decay associated with different forecasting windows.

Sources: README.md:16-24, README.md:120-121

Architecture and Components

The L1 Meta layer sits between the base signal generation (L0) and the conformal calibration (L2). It is designed to mitigate individual model biases and leverage the strengths of specialized predictors in varying market regimes.

Hierarchical Position

In the v3 architecture, the pipeline follows a strict dependency order:

1. L0 Core: Generates base probabilistic multi-step forecasts.
2. L0 Specialists: Generate horizon-agnostic domain signals (e.g., Crush, China, FX).
3. L1 Meta: Stacks Core Out-of-Fold (OOF) predictions and Specialist signals.

Sources: README.md:120-140, README.md:204-215

Data Flow Overview

The L1 Meta learner consumes data from multiple schemas within the Prisma Postgres database to form its training and inference matrices.

This diagram illustrates the convergence of L0 outputs into the L1 training matrix.
Sources: README.md:144-149, README.md:204-215

Input Requirements and Integrity

The L1 Meta learner relies on three primary categories of input data to make informed weighting decisions.

Input Matrix Composition

The training.meta_inputs_1d table is the authoritative source for L1 training. It includes:

• Core OOF Predictions: Out-of-fold predictions (p30, p50, p70) from the L0 Core models to prevent data leakage during stacking.
• Specialist Signals: Signal values and confidence scores from the "Big-11" specialist buckets.
• Regime/Calendar Features: Temporal context such as seasonality and market volatility regimes.

Sources: README.md:146-150, README.md:205-207

Stacking Discipline (Hard Rules)

To ensure the statistical validity of the ensemble, the following rules are enforced:

Rule	Description
OOF Integrity	Meta must train ONLY on out-of-fold Core predictions to avoid overfitting/leakage.
Monotonicity	Quantile uncertainty ranges must widen as horizons increase; narrowing ranges indicate model failure.
Regime-Block CV	Validation folds use regime blocks (e.g., macro eras) with embargo windows around boundaries.

Sources: README.md:152-153, README.md:225-231, cubic.yaml:148-155

Technical Implementation

The L1 Meta layer is implemented as a series of Python-based stacked ensembles managed by the root ML pipeline.

Model Catalog and Storage

L1 Meta models are trained per horizon and stored as artifacts. The resulting predictions are persisted into production tables for dashboard consumption.

• Training Script: src/fusion/core_training/train_models.py (referenced for the zoo of models).
• Output Tables: forecasts.production_{H}d_1d (where H ∈ {5, 21, 63, 126}).
• Registry: Models are registered in model.model_registry with metrics and versioning.

Sources: README.md:210-213, README.md:267-270, AGENTS.md:68-71

Quantile Policy

The L1 Meta layer operates on a specific quantile contract to support probabilistic forecasting:

• Stacking Quantiles: [0.30, 0.50, 0.70].
• Constraint: L1 must enforce a "quantile crossing guardrail" where $p30 \leq p50 \leq p70$.

Sources: README.md:126-128, README.md:162-164

Downstream Consumption

The outputs of the L1 Meta learner (from the model.meta_ensemble table) are consumed by the L5-C Synthesis Engine to generate natural language market posture summaries.

# Example of loading L1 Meta outputs for synthesis
# Source: scripts/generate_synthesis.py:82-95
def load_forecast_data(conn, horizon: int) -> Dict:
    """Load latest meta-ensemble forecast for horizon."""
    with conn.cursor() as cur:
        cur.execute(
            """
            SELECT as_of_date, p10, p50, p90
            FROM "model"."meta_ensemble"
            WHERE horizon = %s
            ORDER BY as_of_date DESC
            LIMIT 1
        """,
            (horizon,),
        )
        row = cur.fetchone()
    # ... logic to return p10, p50, p90

Sources: scripts/generate_synthesis.py:82-105

Specialist Integration

The meta-learner's power comes from its ability to integrate the 11 specialists. The L1 layer learns to adjust weights based on the confidence scores provided by these specialists.

Specialist Bucket	Input to L1
Crush / China	Demand outlook and competition z-scores.
Tariff / Trump Effect	Event intensity and policy risk scores.
Biofuel / Energy	Policy pressure and energy spillover momentum.
Volatility / FX	Regime conditioning and currency pressure indices.

Sources: AGENTS.md:88-100, README.md:58-62

Summary

L1 Meta Ensembling is the intelligence-layer of the ZINC-Fusion pipeline that synthesizes base predictions and domain-specific signals. By enforcing strict OOF discipline and quantile monotonicity, it ensures that the resulting forecasts are not only statistically robust but also contextually aware of market regimes. The outputs of this layer directly inform the calibration and risk engines, ultimately providing the data-driven basis for institutional procurement decisions.