News Event Impact Model
Generated 2026-07-20 05:39:59.937819+00:00 | version 20260720T053747Z
production_model_collecting_strict_prospective_paper_evidence
Promotion Decision
production_alias_preserved_advisory_paper_event_risk_only
A production alias identifies the immutable artifact used for advisory paper-event risk scoring. It is not permission to size a position, submit an order, or enable live trading. Those authorities remain false until separate prospective, execution, and operator gates clear.
What This Model Predicts
The target is the larger of the maximum upward and downward split-adjusted price excursions during the five sessions after the input cutoff, capped at 100%. The input cutoff is the first market session strictly after the UTC event date. This is a movement-risk model, not a directional-alpha claim.
Training 2016-07-18 to 2022-12-28; validation 2023-01-06 to 2024-07-01; historical test 2024-07-10 to 2026-07-10.
The historical holdout was observed while finalizing this first model specification. It is reproducible historical evidence, not a pristine final test; strictly post-assignment paper outcomes remain mandatory before the score may affect sizing or order decisions.
Governed News Contract
Only factual or mixed-context rows with source quality at least 0.60, entity relevance at least 0.70, confidence at least 0.55, and an approved publisher class enter the model. Each symbol/event-key/duplicate cluster is capped at one article. Opinion forecasts remain in their separate track-record model.
Promotion Gates
| Gate | Result |
|---|---|
| bounded_test_bias | passed |
| minimum_historical_holdout_rows | passed |
| minimum_temporal_depth | passed |
| minimum_training_rows | passed |
| minimum_validation_rows | passed |
| news_residual_activation_gate_obeyed | passed |
| purged_temporal_order | passed |
| risk_quintile_separation | passed |
| subperiod_news_lift | passed |
| untouched_test_lift_vs_constant | passed |
| untouched_test_news_incremental_lift | passed |
| untouched_test_rank_correlation | passed |
| validation_news_incremental_lift | passed |
Historical Subperiod Stability
| Block | Start | End | Rows | News RMSE lift |
|---|---|---|---|---|
| 1 | 2024-07-10 | 2025-01-07 | 10745 | 0.058% |
| 2 | 2025-01-08 | 2025-07-11 | 15707 | -0.113% |
| 3 | 2025-07-14 | 2026-01-08 | 13778 | 0.201% |
| 4 | 2026-01-09 | 2026-07-10 | 23578 | 0.173% |
Known Boundary
The news residual is deliberately compared with a strong trailing-price baseline and its incremental lift is small. Historical scores are shadow evidence. At least 500 matured strictly post-assignment predictions across 20 dates are required before review for any decision influence. Broker fills and intraday execution remain separate evidence.