# Additional Implementation Notes


Stage 2 uses horizons {1, 3, 5, 10, 15} to cover immediate, within-week, and slower multi-week reactions while keeping multi-horizon supervision compact. We selected this set after comparing several five-horizon configurations on validation performance; it offered a strong balance between accuracy and economic interpretability, while retaining both the 1-day reaction window and delayed monitoring windows up to 15 trading days.

The Stage 1 event schema contains 20 event types across macroeconomic, company, market-knowledge, and geopolitical families, plus a none class. The schema-constrained LLM prompt emits one JSON record with event type, affected tickers, polarity, signed magnitude, scope, credibility, surprise, novelty, and confidence. We also use a manual silver-label audit protocol that checks event type, affected ticker, polarity, and magnitude direction on a stratified sample; this audit is used only as quality control for the observation layer, not as a new extraction benchmark.

Finally, CausalLagStock should be used as a decision-support and monitoring framework rather than an autonomous trading strategy. Its graph summarizes temporally ordered predictive associations from observational news and price data, and should not be interpreted as evidence of intervention-level mechanisms.