---
name: mlp-compare
description: Compare MLP input feature sensitivities across all trained assets and timeframes to identify universal vs asset-specific signals.
---

```bash
source .venv/bin/activate && python3 tools/mlp_compare_weights.py $SKILL_ARGS
```

After printing the table, analyse:
- **High-variance features** (top rows): signals where the model disagrees across assets — asset-specific behaviour
- **Low-variance features** (bottom rows): signals that are consistently positive or negative everywhere — universal signals
- **Near-zero features** (≈0 everywhere): candidate for removal from FEATURE_COLS to reduce dimensionality
- **Sign flips**: a feature that's strongly positive for BTC but negative for ETH is a meaningful cross-asset insight
- Note that values are a linear approximation (W3·W2·W1); nonlinearity means this understates the true sensitivity for large-weight features

Useful flags:
- `--top 15` — focus on the 15 most discriminating features
- `--tf 4H 1D` — compare specific timeframes only
- `--csv` — emit CSV for spreadsheet analysis