momentum-signals

Build evidence-only momentum signals from component scores and weights, and emit them as journal notes that never clear a ticket by themselves.

SKILL.md

GitHub

Momentum Signals

L1 - Summary

SIGNAL produces evidence, not clearance. Component scores (price action, trade flow, holder change) combine through pinned weights into a signal with its reasons, emitted as a journal.append note with payload note_kind=signal. A signal never buys, never clears risk, and never overrides a halt.

components scored -> weights applied -> signal with reasons -> journal note -> RISK and CHIEF decide

L2 - Method

  1. S1 - Score the components: read market.ohlcv for price structure, market.trades for flow, and market.holders_snapshot for concentration change. Each component reports its score plus the window and source it used.
  2. S2 - Apply the pinned weights: combine components with the reviewed weight table only. Weight changes go through replay-and-backtest first; per-signal weight tuning is forbidden.
  3. S3 - Emit with reasons: append the signal via journal.append kind=journal_note with payload note_kind=signal, carrying the components, weights version, and window. A bare direction without reasons is not a signal.
  4. S4 - Stay evidence-only: the signal proposes; RISK verdicts and human approvals dispose. Never phrase a signal as clearance, and never attach urgency to bypass review.
  5. S5 - Expire honestly: signals carry their window; past the window they are history, not standing orders. Re-score on fresh data instead of reusing a stale signal.

Failure modes

Symptom Cause Action
Signal without reasons Truncated emission Re-emit with components and weights version; reasons are the signal
Weights tuned per signal Overfitting to a favorite Revert to the pinned table; tune only with replay evidence
Stale signal reused Window expired but direction recycled Re-score on fresh reads; expired signals never trigger
Signal treated as clearance Misread authority Re-state: SIGNAL advises, RISK plus human decide

References

  • Tools: market.ohlcv, market.trades, market.holders_snapshot (signal inputs)
  • Fixture: tests/fixtures/http/holders.json (recorded holder response shape)
  • Gate G-2: signals feed approval-gated flow as evidence
  • Companion skills: replay-and-backtest (weight validation), desk-strategy-lab (rule experiments)