FreqAI-Inspired Action Items (ATC)

This file maps useful FreqAI concepts into ATC's event-driven ETH_JPY stack.

Already implemented

• Outlier gating by feature z-score stats in logistic swing family.
• Session-aware label drops (cross-day and no-new-entry window).
• Rolling retrain-friendly model artifacts and metrics JSON.
• Shadow compare for tabular vs image alpha scores.

Added in this iteration

• hold10_oracle target mode for logistic training:
  • long by default unless downside threshold is exceeded.
• ml_direction_logistic_hold_overlay strategy:
  • base long (10 ETH) by default;
  • flatten/short only on strong downside probability.
• Outlier guard was also enabled for hold-overlay strategy.

Next integrations (short horizon)

1. Add explicit do_predict-style confidence flag series into backtest report JSON/CSV.
2. Add walk-forward artifact bundle:
   • train/valid split metadata
   • feature drift stats
   • threshold gate diagnostics
3. Add Optuna-based hyperparameter search entrypoint for hold-overlay policy.
4. Add model audit report:
   • feature contribution summary
   • threshold sensitivity table
   • fallback usage diagnostics near close deadline.

Guardrails

• Keep benchmark fixed at 10 ETH hold for all evaluations.
• Enforce JST session boundary and forced-flat policy.
• Prefer limit-first execution; market fallback only near deadline.