Position Downside Risk Model

Generated 2026-07-20 00:38:46.718185+00:00 | version 20260719T195305Z

production_model_collecting_prospective_paper_evidence

Matured rows462499
Symbols2319
Holdout rows150698
RMSE lift3.57%
MAE lift9.69%
Rank correlation0.3841579951667029
80% envelope coverage82.77%
Historical shadow predictions53052
Matured historical shadow6889
Strict prospective predictions0
Matured strict prospective0
Paper integrationproduction_model_position_monitoring_active_no_entry_candidates
Orders submitted0

Promotion Decision

production_alias_preserved_paper_position_risk_output_only

The production alias supplies a required freshness gate and calibrated loss-budget cap to the guarded paper execution manager. The model itself cannot size a position; the separate execution policy performs paper-only sizing. The model cannot submit or modify an order, enable live trading, or claim broker-fill evidence.

Paper Execution Integration

production_model_position_monitoring_active_no_entry_candidates

Latest execution run alpaca_paper_execution_manager_20260720T003145Z; current model-scored intents 3; model-sized entries 0; model-blocked entries 0; paper orders submitted 0; live trading False.

Current Paper Positions

SymbolQuantityPredicted MAE80% risk envelopeStatus
STNG14.4590803934.81%12.69%paper_risk_score_current
TNK15.4010472815.24%13.11%paper_risk_score_current
ASA19.076497526.08%13.96%paper_risk_score_current

Point-in-Time Contract

Inputs use the split-adjusted close and trailing price/volume history at each cutoff. Targets use the lowest split-adjusted price over the next 20 sessions, excluding the anchor session. Training anchors are spaced 20 sessions apart. Scores at or before the model assignment date are historical shadow evidence; only later market cutoffs count as strict prospective evidence.

Training 2010-01-04 to 2020-09-14; validation 2020-10-13 to 2023-02-15; historical holdout 2023-03-17 to 2026-06-12.

A bounded 600-symbol architecture prototype observed aggregate performance over an overlapping calendar block before the full-universe fit. Treat this as a historical holdout rather than a pristine final test; prospective paper outcomes remain mandatory.

Promotion Gates

GateResult
adaptive_envelope_coveragepassed
adaptive_envelope_sharpnesspassed
mae_lift_vs_constantpassed
minimum_historical_holdout_datespassed
minimum_historical_holdout_rowspassed
minimum_training_datespassed
minimum_training_rowspassed
minimum_validation_datespassed
minimum_validation_rowspassed
purged_temporal_orderpassed
rank_correlationpassed
risk_quintile_separationpassed
rmse_lift_vs_constantpassed

Known Boundary

This is an end-of-day risk model, not an intraday stop executor. The adaptive envelope is a calibrated research estimate, not a guaranteed loss cap. Prospective paper outcomes and broker execution evidence remain separate gates.