Same model evaluated on 3 news-filtered subsets of the 2024 test set.
Backtest Results
Entry score + predicted hold-time vs actual trade outcomes from closed trades.
Feature Promotion Experiments
Systematic grid search — every candidate feature tested individually and in combination to find the best models before Shadow Mode.
Entry Quality AUC (higher is better — baseline: Loki 0.584)
Hold-Time MAE (lower is better — baseline: Loki 3.73d)
Magnitude Regression (predicts max recovery % — lower MAE is better)
Hourly Resolution (hourly bar aggregation vs 15-min — compare AUC / MAE)
Phase 2e — Model Architecture
Two LightGBM models for Loki, trained on 23 entry-time features from the historical dip dataset.
Entry Quality Model
Binary classifier: "Will this dip recover the trader's target% within the cap window?" Output: P(recovery) ∈ [0,1]. Used to filter low-quality entries.
Hold-Time Model
Regression: "How many trading days until this dip hits the target?" Output: predicted days. Informs adaptive exit timing. Censored observations weighted 0.5×.
Per-Trader Targets
Loki: 3% in 14 days Demeter: excluded (guardrails only)
News Filter Variants
Holdtime model evaluated on 3 news subsets: • All — all qualifying dips • News-Only — headlineCount > 0 • High-Severity — maxSeverity ≥ 3
23 Entry Features
Drop metrics, momentum, RSI, Bollinger, volume, SPY context, VIX, sector, news catalyst, severity, breadth, timing — all measured at dip entry date.