| name | run-pipeline |
| description | Run ML training pipeline steps (specialists, core, evaluation, forecasts, monte-carlo). Use when user asks to run or re-run any part of the prediction pipeline. |
| disable-model-invocation | true |
Run the ML pipeline step specified in $ARGUMENTS.
Current State
- Branch: !
git branch --show-current
- Venv: !
which python
Available Steps
| Step | Command | Description |
|---|
specialists | python scripts/generate_specialist_signals.py | Generate Big-11 specialist signals |
features | python scripts/generate_specialist_features.py | Build specialist feature columns |
core | python -m fusion.core_training.run_pipeline | Train 4 AutoGluon ensembles (5d/21d/63d/126d) |
evaluate | python scripts/evaluate_oof.py | Evaluate OOF predictions vs actuals |
forecasts | python scripts/generate_production_forecasts.py | Generate production forecast outputs |
monte-carlo | python scripts/autorun_monte_carlo.py | Run L3 Monte Carlo simulation (10,000 runs) |
all | Run all steps above in order | Full pipeline end-to-end |
Pre-flight (MANDATORY before any step)
- Confirm venv is active:
source .venv/bin/activate
- Set PYTHONPATH:
export PYTHONPATH=src:$PYTHONPATH
- Run
make check — abort if it fails
Rules (from AGENTS.md)
- NEVER modify
src/fusion/core_training/config.py — it is FROZEN (2026-02-19)
- Core trains 4 independent TimeSeriesPredictor ensembles (one per horizon)
- Target = price level (
close.shift(-horizon)), named target_price_{h}d
- Core metric = MAE (point forecast accuracy)
- 11 specialists (NEVER 10) — trump_effect is the 11th
- MODEL_ZOO_FROZEN has 19 active models — do not add/remove without approval
Post-run
- Report exit code and any errors
- If
evaluate, show the OOF summary table
- If
core, confirm all 4 horizons completed with WeightedEnsemble
