🗄️ 歴史的文書(アーカイブ) — この文書は過去の研究フェーズの記録であり、現在の結論・手法を反映していません。現在の研究状況は解説セクションを参照してください。

FreqAI-Inspired Roadmap for ATC

This document maps FreqAI concepts to this repository and defines concrete implementation steps.

1. Concept Mapping

FreqAI concept	ATC equivalent (current)	Gap
DataKitchen (feature filtering, scaling, split, outlier handling)	atc.research.ml_linear, atc.research.image_alpha, FeatureEngine	Centralized training/inference preprocessing contract is still fragmented
DataDrawer (model metadata, persistence, purge)	data/derived/models/*.json, ad hoc loaders in strategies	No unified model registry/retention policy
do_predict / outlier gating	Risk constraints + spread/flow guards in strategies	Need unified outlier confidence flag at inference
Sliding retrain windows	walk-forward + explicit training CLI	Needs unified scheduler + reproducible training manifests

2. Immediate Upgrades (done / in-flight)

1. Robust model loading fallback for logistic strategies.
2. Session-specific strategy variants for regime-aware execution.
3. Outlier guard (z-score based) in ml_direction_logistic_swing family.

3. Next Upgrades (priority order)

1. ATC DataKitchen module

• Single preprocessing contract for tabular/image models.
• Explicit feature schema version + strict column matching.
• Reusable normalization objects for train/infer parity.

2. ATC Model Registry (DataDrawer-like)

• Canonical model manifest:
  • model_type, feature_names, train_range, valid_metrics, created_at, git_commit.
• Retention policy (keep_latest_n) and cleanup command.

3. Unified confidence flags

• Emit do_predict-style fields:
  • predict_confidence
  • outlier_score
  • outlier_blocked
• Use in strategy and report summaries.

4. Producer/Consumer training separation

• Optional async trainer process writing model artifacts.
• Live strategy process hot-reloads newest validated artifact.

4. Evaluation Rules (fixed)

1. Benchmark is always fixed hold: benchmark_hold_eth=10.0.
2. Business-day boundary is JST 06:00.
3. Hard flat by 05:59 with fallback policy.
4. Gate promotion by:

• alpha_total_pnl_jpy > 0
• deadline_breach_days == 0
• win-rate and drawdown checks.