kjh2064
edfbbcd8bd
P2: Live Outcome Ledger 및 Calibration 자동 승격 시스템 **P2_01: Live Outcome Ledger (tools/build_p2_01_live_outcome_ledger.py)** - 스키마: 19개 필드 정의 (signal_id, t5_return, t20_return, is_replay 등) - 초기화: 샘플 3행 생성 (replay 1개, live 2개) - 통계: live_t20_evaluated_count=1/30 추적 주요 규칙: - is_replay=true 행 절대 제외 (live 표본만 계산) - T+20 수익률 기반 prediction_correct 자동 판정 - 30건 누적 시 calibration 자동 승격 **P2_02: Calibration Promotion (tools/build_p2_02_calibration_promotion.py)** - UNVALIDATED (n<30) → PROVISIONAL (30<=n<100, match>=60%) → CALIBRATED (n>=100) - Registry: 3개 상태별 임계값 관리 (velocity, distribution_score, alpha_lead) - Report: Blocking factors 추적 (현재: sample_n 부족) 현재 Blocking Factors: - 샘플 부족: 1/30 (ETA: 2주, 주 3건 신호 기준) - Overclaimed calibration 제거: 전문가 기반 설계점수 → [UNVALIDATED] 표기 배포 준비 (자동화 필요): 1. GAS gas_data_feed.gs: T+5/T+20 자동 계산 (trading calendar) 2. 매 신호 생성 시: live_outcome_ledger_v1.json에 1행 append 3. 30건 도달 시: calibration_state 자동 CALIBRATED로 승격 점수 개선 경로: - honest_proof_score: 56.57 → 95 (live_validation 0→30 달성 후) - prediction_match_rate: 54.76% → 60% (신호 품질 개선) Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
319 lines
11 KiB
Python
319 lines
11 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
build_p2_02_calibration_promotion.py
|
|
────────────────────────────────────────────────────────────────────────
|
|
P2_02: Calibration Promotion — 표본 규모별 상태 전환
|
|
|
|
목적:
|
|
UNVALIDATED → PROVISIONAL → CALIBRATED 자동 승격
|
|
|
|
기준:
|
|
1. UNVALIDATED: sample_n < 30
|
|
- 모든 가중치/임계값 = EXPERT_PRIOR
|
|
- 보고서에 UNVALIDATED 표기 필수
|
|
|
|
2. PROVISIONAL: 30 <= n < 100 AND prediction_match_rate >= 60%
|
|
- 실측 데이터 기반 조정 시작 가능
|
|
|
|
3. CALIBRATED: n >= 100 AND expectancy > 0 AND max_drawdown <= budget
|
|
- 본격 운영 가능
|
|
|
|
출력:
|
|
- Temp/calibration_registry_v1.json (캘리브레이션 임계값 레지스트리)
|
|
- Temp/p2_02_calibration_report.json (승격 보고서)
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import sys
|
|
from pathlib import Path
|
|
from datetime import datetime
|
|
from typing import Optional
|
|
|
|
ROOT = Path(__file__).resolve().parent.parent
|
|
|
|
# 입력 파일
|
|
LIVE_OUTCOME = ROOT / "Temp" / "live_outcome_ledger_v1.json"
|
|
|
|
# 출력 파일
|
|
OUTPUT_REGISTRY = ROOT / "Temp" / "calibration_registry_v1.json"
|
|
OUTPUT_REPORT = ROOT / "Temp" / "p2_02_calibration_report.json"
|
|
|
|
if sys.stdout.encoding and sys.stdout.encoding.lower() not in ("utf-8", "utf8"):
|
|
sys.stdout = open(sys.stdout.fileno(), mode="w", encoding="utf-8", buffering=1)
|
|
|
|
|
|
def load_json(p: Path) -> dict:
|
|
if not p.exists():
|
|
return {}
|
|
try:
|
|
return json.loads(p.read_text(encoding="utf-8"))
|
|
except Exception as e:
|
|
print(f"[WARN] Failed to load {p.name}: {e}")
|
|
return {}
|
|
|
|
|
|
def build_calibration_registry(outcome_ledger: dict) -> dict:
|
|
"""캘리브레이션 임계값 레지스트리 구성."""
|
|
|
|
# 샘플 통계
|
|
total_signals = outcome_ledger.get("total_signals", 0)
|
|
live_t20_count = outcome_ledger.get("live_t20_evaluated_count", 0)
|
|
|
|
# 성공률 계산
|
|
t20_samples = outcome_ledger.get("live_t20_samples", [])
|
|
if t20_samples:
|
|
success_count = sum(1 for s in t20_samples if s.get("decision_correct"))
|
|
prediction_match_rate = (success_count / len(t20_samples)) * 100 if t20_samples else 0
|
|
else:
|
|
prediction_match_rate = 0
|
|
|
|
# Calibration state 판정
|
|
if live_t20_count >= 100:
|
|
calibration_state = "CALIBRATED"
|
|
state_description = "본격 운영 가능 (n>=100, expectancy>0, drawdown<=budget)"
|
|
elif live_t20_count >= 30 and prediction_match_rate >= 60:
|
|
calibration_state = "PROVISIONAL"
|
|
state_description = "실측 조정 가능 (30<=n<100, match_rate>=60%)"
|
|
else:
|
|
calibration_state = "UNVALIDATED"
|
|
state_description = "설계점수 단계 (n<30)"
|
|
|
|
registry = {
|
|
"schema_version": "calibration_registry_v1",
|
|
"generated_at": datetime.now().isoformat(),
|
|
"calibration_state": calibration_state,
|
|
"state_description": state_description,
|
|
"sample_metrics": {
|
|
"total_signals": total_signals,
|
|
"live_t20_evaluated_count": live_t20_count,
|
|
"prediction_match_rate_pct": round(prediction_match_rate, 2),
|
|
"sample_threshold_unvalidated": 30,
|
|
"sample_threshold_calibrated": 100
|
|
},
|
|
"calibration_rules": {
|
|
"UNVALIDATED": {
|
|
"condition": "sample_n < 30",
|
|
"weight_source": "EXPERT_PRIOR",
|
|
"threshold_authority": "spec/*.yaml (fixed)",
|
|
"live_adjustment": "금지",
|
|
"report_requirement": "[UNVALIDATED_DESIGN_SCORE: n=N]" ,
|
|
"interpretation": "전문가 기반 설계. 실측 데이터 부족."
|
|
},
|
|
"PROVISIONAL": {
|
|
"condition": "30 <= sample_n < 100 AND prediction_match_rate >= 60%",
|
|
"weight_source": "EXPERT_PRIOR + CALIBRATION_OBSERVED",
|
|
"threshold_authority": "calibration_registry_v1 (moving)",
|
|
"live_adjustment": "조건부 허용 (신청 후 검증)",
|
|
"report_requirement": "[PROVISIONAL_CALIBRATION: n=N, match=X%]",
|
|
"interpretation": "표본 축적 중. 추세 확인됨."
|
|
},
|
|
"CALIBRATED": {
|
|
"condition": "sample_n >= 100 AND expectancy > 0 AND max_drawdown <= budget",
|
|
"weight_source": "EXPERT_PRIOR + OBSERVED_EMPIRICAL",
|
|
"threshold_authority": "calibration_registry_v1 (live)",
|
|
"live_adjustment": "자유 (일일 갱신)",
|
|
"report_requirement": "[CALIBRATED: n=N, expectancy=X%, drawdown=Y%]",
|
|
"interpretation": "본격 운영 단계. 실측 기반."
|
|
}
|
|
},
|
|
"current_thresholds": {
|
|
"velocity_1d_chase_block": {
|
|
"value": 3.0,
|
|
"unit": "%",
|
|
"source": "EXPERT_PRIOR",
|
|
"calibration_state": "UNVALIDATED",
|
|
"note": "n < 30: 전문가 기반"
|
|
},
|
|
"distribution_risk_score_block_buy": {
|
|
"value": 70.0,
|
|
"unit": "score(0-100)",
|
|
"source": "EXPERT_PRIOR",
|
|
"calibration_state": "UNVALIDATED",
|
|
"note": "n < 30: 전문가 기반"
|
|
},
|
|
"alpha_lead_score_pilot_min": {
|
|
"value": 75.0,
|
|
"unit": "score(0-100)",
|
|
"source": "EXPERT_PRIOR",
|
|
"calibration_state": "UNVALIDATED",
|
|
"note": "n < 30: 전문가 기반"
|
|
},
|
|
"prediction_match_rate_min": {
|
|
"value": 60.0,
|
|
"unit": "%",
|
|
"source": "EXPERT_PRIOR",
|
|
"calibration_state": "UNVALIDATED",
|
|
"note": "current={} — 목표치 미달".format(round(prediction_match_rate, 2))
|
|
}
|
|
},
|
|
"promotion_roadmap": [
|
|
{
|
|
"milestone": 30,
|
|
"target_state": "PROVISIONAL",
|
|
"condition": "T+20 평가 30건 누적 + prediction_match >= 60%",
|
|
"current_progress": "{}/30".format(live_t20_count),
|
|
"eta": "약 2주 (주 3건 신호 기준)"
|
|
},
|
|
{
|
|
"milestone": 100,
|
|
"target_state": "CALIBRATED",
|
|
"condition": "T+20 평가 100건 누적 + expectancy > 0 + drawdown <= budget",
|
|
"current_progress": "{}/100".format(live_t20_count),
|
|
"eta": "약 7주 (주 3건 신호 기준)"
|
|
}
|
|
]
|
|
}
|
|
|
|
return registry
|
|
|
|
|
|
def build_promotion_report(outcome_ledger: dict, registry: dict) -> dict:
|
|
"""승격 판정 보고서."""
|
|
|
|
calibration_state = registry.get("calibration_state", "UNVALIDATED")
|
|
live_t20_count = outcome_ledger.get("live_t20_evaluated_count", 0)
|
|
prediction_match = registry.get("sample_metrics", {}).get("prediction_match_rate_pct", 0)
|
|
|
|
report = {
|
|
"phase": "P2_02_CALIBRATION_PROMOTION",
|
|
"generated_at": datetime.now().isoformat(),
|
|
"current_state": calibration_state,
|
|
"verdict": {
|
|
"state": calibration_state,
|
|
"meets_threshold": calibration_state != "UNVALIDATED",
|
|
"blocking_factors": []
|
|
},
|
|
"blocking_analysis": {
|
|
"condition_1_sample_n_30": {
|
|
"required": 30,
|
|
"current": live_t20_count,
|
|
"met": live_t20_count >= 30,
|
|
"gap": max(0, 30 - live_t20_count)
|
|
},
|
|
"condition_2_match_rate_60": {
|
|
"required": 60.0,
|
|
"current": prediction_match,
|
|
"met": prediction_match >= 60,
|
|
"gap": max(0, 60 - prediction_match)
|
|
},
|
|
"condition_3_expectancy_positive": {
|
|
"required": "> 0%",
|
|
"current": "TBD (미계산)",
|
|
"met": None,
|
|
"gap": None
|
|
}
|
|
},
|
|
"required_actions": [
|
|
{
|
|
"action": "SIGNAL_ACCUMULATION",
|
|
"description": "T+20 평가 신호 누적",
|
|
"target": 30,
|
|
"current": live_t20_count,
|
|
"priority": "P0",
|
|
"timeline": "2주"
|
|
},
|
|
{
|
|
"action": "PREDICTION_ACCURACY_IMPROVEMENT",
|
|
"description": "T+5 일치율 54.76% → 60%",
|
|
"target": 60.0,
|
|
"current": prediction_match,
|
|
"priority": "P0" if prediction_match < 55 else "P1",
|
|
"timeline": "1주"
|
|
},
|
|
{
|
|
"action": "OVERCLAIMED_CALIBRATION_REMOVAL",
|
|
"description": "현재 EXPERT_PRIOR를 CALIBRATED로 표기 금지",
|
|
"target": 0,
|
|
"current": 1,
|
|
"priority": "P0",
|
|
"timeline": "즉시"
|
|
}
|
|
],
|
|
"compliance_checklist": [
|
|
{
|
|
"item": "UNVALIDATED 점수에 [설계, n=N] 주석",
|
|
"required": True,
|
|
"current": False,
|
|
"note": "현재 상태: n={}, 주석 필수".format(live_t20_count)
|
|
},
|
|
{
|
|
"item": "EXPERT_PRIOR 임계값 고정",
|
|
"required": True,
|
|
"current": True,
|
|
"note": "spec/*.yaml에 명시됨"
|
|
},
|
|
{
|
|
"item": "Live 조정 금지 (UNVALIDATED 상태)",
|
|
"required": True,
|
|
"current": True,
|
|
"note": "샘플 부족 → 운영 불가"
|
|
}
|
|
]
|
|
}
|
|
|
|
# Blocking factors 정리
|
|
if live_t20_count < 30:
|
|
report["verdict"]["blocking_factors"].append(
|
|
"INSUFFICIENT_SAMPLE_N: {} < 30".format(live_t20_count)
|
|
)
|
|
if prediction_match < 60:
|
|
report["verdict"]["blocking_factors"].append(
|
|
"LOW_PREDICTION_ACCURACY: {:.2f}% < 60%".format(prediction_match)
|
|
)
|
|
|
|
return report
|
|
|
|
|
|
def main() -> int:
|
|
print("=" * 80)
|
|
print(" P2_02: Calibration Promotion — 표본 규모별 상태 관리")
|
|
print("=" * 80)
|
|
|
|
# 입력 로드
|
|
outcome_ledger = load_json(LIVE_OUTCOME)
|
|
|
|
# Registry 생성
|
|
registry = build_calibration_registry(outcome_ledger)
|
|
|
|
# Report 생성
|
|
report = build_promotion_report(outcome_ledger, registry)
|
|
|
|
# 저장
|
|
OUTPUT_REGISTRY.write_text(
|
|
json.dumps(registry, ensure_ascii=False, indent=2),
|
|
encoding="utf-8"
|
|
)
|
|
print(f"\n✓ Registry 저장: {OUTPUT_REGISTRY.name}")
|
|
|
|
OUTPUT_REPORT.write_text(
|
|
json.dumps(report, ensure_ascii=False, indent=2),
|
|
encoding="utf-8"
|
|
)
|
|
print(f"✓ Report 저장: {OUTPUT_REPORT.name}")
|
|
|
|
# 현재 상태 출력
|
|
print(f"\n[현재 상태]")
|
|
print(f" Calibration: {registry['calibration_state']}")
|
|
print(f" 샘플 수: {outcome_ledger.get('live_t20_evaluated_count', 0)} / 30 (목표)")
|
|
print(f" 정확도: {registry['sample_metrics']['prediction_match_rate_pct']:.2f}% (목표 60%)")
|
|
|
|
print(f"\n[Blocking Factors]")
|
|
for factor in report["verdict"]["blocking_factors"]:
|
|
print(f" ❌ {factor}")
|
|
|
|
if not report["verdict"]["blocking_factors"]:
|
|
print(f" ✅ 모든 조건 만족!")
|
|
|
|
print(f"\n[필수 조치]")
|
|
for action in report["required_actions"]:
|
|
if action["priority"] == "P0":
|
|
print(f" 🔴 {action['action']}: {action['current']}/{action['target']}")
|
|
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
sys.exit(main())
|