What problem does it solve?
Algorithmic trading systems that touch retail capital must produce interpretable, auditable explanations of model predictions to satisfy EU AI Act and SEC Reg-AI guidance. This Skill ranks the features that drove an LSTM or Transformer trading signal and persists the result as a cryptographically signed, verifiable artifact.
Core Features & Use Cases
- Feature Attribution via PageRank: Builds a feature-contribution graph from model output and runs single-entry seeded PageRank to rank the top-K features that drove a prediction.
- Signed Audit Artifacts: Signs each attribution result with Ed25519 and stores it in the long-lived
trading-analysis namespace for regulator-facing verification.
- Deterministic Fallbacks: Falls back to a local seeded power-iteration kernel when the MCP PageRank tool is unavailable, and to a z-score heuristic when the model lacks an
--explain flag, tagging degraded artifacts accordingly.
- Use Case: A compliance engineer runs
trader-explain SIGNAL-42 --top-k 10 --seed 42 to produce a reproducible, signed ranking of the features behind a model's buy signal, then verifies the artifact before including it in a regulatory filing.
Quick Start
Ask the agent to explain trading signal SIGNAL-42 with trader-explain using top-k 10 and seed 42, then review the returned ranked feature table and signed artifact.
