0012. Factor composition and long-only portfolio construction¶
Status: proposed (2026-09-28)
Context¶
SignalQuarry factors produce cross-sectional scores, while ordinary strategies
already return long-only portfolio weights through Decision.target. Factor
outputs need a deterministic path into that existing strategy, backtest, and
paper-planning contract. The factor diagnostics available today remain
descriptive and unverified; they cannot by themselves support evidence claims.
Decision¶
Each factor's available scores are standardized cross-sectionally with the SDK
population z-score. The default composite is an equal-weight mean of the
standardized scores available for each symbol. A caller may instead supply a
dated weight schedule. Every schedule row records known_at and
effective_from, requires known_at < effective_from, and becomes usable only
on or after effective_from. The caller owns the schedule and its statistical
provenance; this helper does not estimate IC weights or label them verified.
Missing scores are not imputed. A symbol with no positive-weight standardized score is ineligible. The target set is the top requested quantile, with a rank-based exit buffer for current holdings. Score ties use symbol order for deterministic selection. Selected symbols receive equal weights up to the declared per-symbol cap; unallocated exposure remains cash.
The result is an ordinary Decision.target. Backtest and paper continue to use
the shared order planner, configured transaction costs, minimum order notional,
validation gates, and paper arming controls. No factor-specific execution or
paper authority is added.
Consequences¶
An agent can compose reusable @factor functions inside a normal strategy and
run the resulting target through existing parity and evidence workflows. A
dated weight schedule makes the information cutoff explicit but does not prove
that its inputs were correctly generated. Factor diagnostics remain
unverified; provenance, trial accounting, holdout controls, and ordinary
strategy gates remain separate requirements.