Interpretation Boundary
This entry documents a modeling workflow. It should not be read as a live trading claim.
HERMES-IC+ feature and evaluation workflow
Workflow record for OHLCV-only causal feature generation, score-cell-aligned expert modeling, purged out-of-fold fusion, and information-coefficient evaluation.
Reproducibility
Repository status
Complete reproduction materials are available but not yet public. A GitHub repository will be linked here when it is ready.
Environment
Language: Python Package: pip / conda OS: Linux recommended for full training Hardware: CPU for feature generation; GPU optional for heavier model training
Reproduction Steps
- Use only historical OHLCV inputs for inference-time feature generation.
- Generate the 463-dimensional causal feature matrix under the feature-count cap.
- Train expert models using purged expanding out-of-fold splits to reduce leakage risk.
- Fit event-aware fusion only from out-of-fold predictions and documented training labels.
- Evaluate by dataset, horizon, and normal/extreme score cells using information coefficient summaries.
Expected Outputs
- Finite factor and signal matrices with no future-information inputs at inference time.
- Score-cell-level IC tables and negative-cell counts.
- An evidence summary with leakage-control, held-out behavior, and robustness boundaries.
Next steps