From d37ad8b7ba189bbae122870a30152d0668199bad Mon Sep 17 00:00:00 2001 From: user040131 Date: Sun, 12 Apr 2026 20:21:33 +0900 Subject: [PATCH] feat: add replay calibration and exit timing analysis --- .env.example | 2 +- .../replay/weekly-replay-client.tsx | 235 +++++++++++- packages/contracts/src/index.ts | 60 +++- scripts/seed_reference_data.py | 1 + services/api/app/core/config.py | 2 +- services/api/app/repositories/dashboard.py | 2 + services/api/app/schemas/api.py | 60 ++++ services/api/app/services/agents/router.py | 7 +- .../api/app/services/replay/calibration.py | 338 ++++++++++++++++++ .../api/app/services/replay/trade_timing.py | 162 +++++++++ .../api/app/services/replay/weekly_replay.py | 175 +++++++++ .../api/tests/test_replay_analysis_helpers.py | 98 +++++ 12 files changed, 1129 insertions(+), 13 deletions(-) create mode 100644 services/api/app/services/replay/calibration.py create mode 100644 services/api/app/services/replay/trade_timing.py create mode 100644 services/api/tests/test_replay_analysis_helpers.py diff --git a/.env.example b/.env.example index 337e7a6..2869213 100644 --- a/.env.example +++ b/.env.example @@ -22,6 +22,6 @@ MODEL_MICROSTRUCTURE_ANALYST=gpt-5.4-mini MODEL_SKEPTIC_ANALYST=gpt-5.4-mini MODEL_KOREA_TRANSLATOR=gpt-5.4 MODEL_FINAL_JUDGE=gpt-5.4 +MODEL_REPLAY_CALIBRATOR=gpt-5.4-mini ALPHA_VANTAGE_API_KEY= POLYGON_API_KEY= - diff --git a/apps/web/components/replay/weekly-replay-client.tsx b/apps/web/components/replay/weekly-replay-client.tsx index e8b65bd..aa59e31 100644 --- a/apps/web/components/replay/weekly-replay-client.tsx +++ b/apps/web/components/replay/weekly-replay-client.tsx @@ -14,7 +14,7 @@ const API_BASE_URL = const REPLAY_TIMEOUT_MS = 90000; function outcomeVariant(label: string): "ok" | "warn" | "danger" { - if (label === "강한 반응" || label === "양호") { + if (label === "강한 적중" || label === "양호") { return "ok"; } if (label === "보통") { @@ -23,6 +23,16 @@ function outcomeVariant(label: string): "ok" | "warn" | "danger" { return "danger"; } +function deviationVariant(score: number): "ok" | "warn" | "danger" { + if (score >= 0.75) { + return "danger"; + } + if (score >= 0.45) { + return "warn"; + } + return "ok"; +} + function tierLabel(tier: string) { return tier === "leader" ? "주도주" : "2등주"; } @@ -84,7 +94,8 @@ export function WeeklyReplayClient() {

Loading

- 리플레이 데이터를 불러오는 중입니다. 최근 7거래일 미국 이벤트와 한국장 결과를 정리하고 있습니다. + 리플레이 데이터를 불러오는 중입니다. 최근 7거래일 미국 이벤트, 실제 한국장 반응, + 보정 분석을 함께 계산하고 있습니다.

); @@ -107,12 +118,12 @@ export function WeeklyReplayClient() { return ( <> -
+

Window

{data.windowLabel}

- 현재 운영 중인 추론 규칙과 동일한 방식으로 과거 구간을 재구성한 결과입니다. + 현재 추론 규칙과 동일한 방식으로 최근 구간을 다시 돌린 결과입니다.

@@ -133,8 +144,106 @@ export function WeeklyReplayClient() { 전체 리플레이 기준 주도주의 평균 종가 수익률입니다.

+ +

Weekly Drift

+

+ {formatScore(data.calibrationSummary.averageDeviationScore)} +

+

+ 의미 있는 오차 {data.calibrationSummary.daysWithMeaningfulMismatch}일 / 분석 모드{" "} + {data.calibrationSummary.analysisMode === "model_augmented" ? "모델 보강" : "규칙 기반"} +

+
+ +
+
+

Calibration Summary

+

+ 왜 빗나갔는지와 다음 분석에서 조정할 포인트 +

+

+ {data.calibrationSummary.summary} +

+
+
+

역할 {data.calibrationSummary.role}

+

모델 {data.calibrationSummary.model}

+

신뢰도 {formatScore(data.calibrationSummary.confidence)}

+
+
+ +
+ {data.calibrationSummary.topDriverLabels.map((label) => ( + + {label} + + ))} +
+ +
+
+

권장 보정

+
+ {data.calibrationSummary.suggestedAdjustments.map((action) => ( +
+

+ {action.target} · {action.adjustment} +

+

+ {action.rationale} +

+
+ ))} +
+
+ +
+

다음 확인 포인트

+
+ {data.calibrationSummary.watchouts.map((item) => ( +

+ {item} +

+ ))} +
+
+
+ + + +
+
+

Exit Timing

+

+ 시가 진입 가정 시 언제 파는 것이 가장 유리했는가 +

+

+ {data.tradeTimingSummary.summary} +

+
+
+

평균 추가 수익 {formatPct(data.tradeTimingSummary.averageExtraReturnPct)}

+

기준 5분봉 / 시가 진입 가정

+
+
+ +
+ {data.tradeTimingSummary.topSignalLabels.map((label) => ( + + {label} + + ))} +
+ +
{orderedDays.map((day) => ( @@ -159,7 +268,7 @@ export function WeeklyReplayClient() {

{day.summary}

-
+
{day.predictedThemes.length > 0 ? ( day.predictedThemes.map((theme) => ( @@ -197,13 +306,13 @@ export function WeeklyReplayClient() {

예상 주도주 / 2등주

- 주도주: {theme.leaders.map((item) => `${item.ticker} ${item.name}`).join(", ") || "-"} + 주도주 {theme.leaders.map((item) => `${item.ticker} ${item.name}`).join(", ") || "-"}

- 2등주: {theme.secondTier.map((item) => `${item.ticker} ${item.name}`).join(", ") || "-"} + 2등주 {theme.secondTier.map((item) => `${item.ticker} ${item.name}`).join(", ") || "-"}

- 무효화 조건: {theme.invalidationCondition} + 무효화 조건 {theme.invalidationCondition}

@@ -241,12 +350,120 @@ export function WeeklyReplayClient() { )) ) : (
- 해당 날짜에는 현재 규칙 기준으로 의미 있는 한국 테마 후보가 생성되지 않았습니다. + 해당 날짜에는 현재 규칙 기준으로 의미 있는 한국장 테마 후보가 생성되지 않았습니다.
)}
+ +
+
+

오차 원인 분석

+

+ {day.calibrationReview.summary} +

+
+ + 오차 {formatScore(day.calibrationReview.deviationScore)} + +
+ +
+ + {day.calibrationReview.mismatchLabel} + + + 분석 신뢰도 {formatScore(day.calibrationReview.calibrationConfidence)} + +
+ +
+ {day.calibrationReview.drivers.map((driver) => ( +
+
+

{driver.label}

+ + {driver.impactLabel} + +
+

+ {driver.reason} +

+

+ 근거 {driver.evidence} +

+
+ ))} +
+ +
+

다음 분석에서의 보정 제안

+
+ {day.calibrationReview.suggestedActions.map((action) => ( +
+

+ {action.target} · {action.adjustment} +

+

+ {action.rationale} +

+
+ ))} +
+
+ + + +
+
+

최적 매도 시점 분석

+

+ {day.tradeOptimization.executionNote} +

+
+ 0.5 + ? "ok" + : "warn" + } + > + {day.tradeOptimization.chartSignal} + +
+ +
+

+ 분석 종목 {day.tradeOptimization.ticker} {day.tradeOptimization.name} +

+

{day.tradeOptimization.entryAssumption}

+

+ 최적 매도 시점 {day.tradeOptimization.bestExitTimeLabel ?? "-"} / 최고 수익률{" "} + {day.tradeOptimization.bestExitReturnPct !== null + ? formatPct(day.tradeOptimization.bestExitReturnPct) + : "-"} +

+

+ 종가 보유 수익률{" "} + {day.tradeOptimization.closeReturnPct !== null + ? formatPct(day.tradeOptimization.closeReturnPct) + : "-"}{" "} + / 추가 확보 가능 수익{" "} + {day.tradeOptimization.extraReturnPct !== null + ? formatPct(day.tradeOptimization.extraReturnPct) + : "-"} +

+
+ +

근거 이벤트

diff --git a/packages/contracts/src/index.ts b/packages/contracts/src/index.ts index dbd80f9..92ce251 100644 --- a/packages/contracts/src/index.ts +++ b/packages/contracts/src/index.ts @@ -153,6 +153,29 @@ export const WeeklyReplayPayloadSchema = z.object({ positiveHitDays: z.number(), avgLeaderCloseReturnPct: z.number() }), + calibrationSummary: z.object({ + analysisMode: z.string(), + role: z.string(), + model: z.string(), + summary: z.string(), + daysWithMeaningfulMismatch: z.number(), + averageDeviationScore: z.number(), + topDriverLabels: z.array(z.string()), + suggestedAdjustments: z.array( + z.object({ + target: z.string(), + adjustment: z.string(), + rationale: z.string() + }) + ), + watchouts: z.array(z.string()), + confidence: z.number() + }), + tradeTimingSummary: z.object({ + summary: z.string(), + averageExtraReturnPct: z.number(), + topSignalLabels: z.array(z.string()) + }), days: z.array( z.object({ marketDateLabel: z.string(), @@ -212,7 +235,42 @@ export const WeeklyReplayPayloadSchema = z.object({ url: z.string() }) ), - summary: z.string() + summary: z.string(), + calibrationReview: z.object({ + mismatchLabel: z.string(), + deviationScore: z.number(), + calibrationConfidence: z.number(), + summary: z.string(), + drivers: z.array( + z.object({ + code: z.string(), + label: z.string(), + impactLabel: z.string(), + reason: z.string(), + evidence: z.string() + }) + ), + suggestedActions: z.array( + z.object({ + target: z.string(), + adjustment: z.string(), + rationale: z.string() + }) + ) + }), + tradeOptimization: z.object({ + availabilityLabel: z.string(), + ticker: z.string(), + name: z.string(), + barInterval: z.string(), + entryAssumption: z.string(), + bestExitTimeLabel: z.string().nullable(), + bestExitReturnPct: z.number().nullable(), + closeReturnPct: z.number().nullable(), + extraReturnPct: z.number().nullable(), + chartSignal: z.string(), + executionNote: z.string() + }) }) ) }); diff --git a/scripts/seed_reference_data.py b/scripts/seed_reference_data.py index 63aad44..7758196 100644 --- a/scripts/seed_reference_data.py +++ b/scripts/seed_reference_data.py @@ -74,6 +74,7 @@ ("skeptic_analyst", "gpt-5.4-mini"), ("korea_translator", "gpt-5.4"), ("final_judge", "gpt-5.4"), + ("replay_calibrator", "gpt-5.4-mini"), ] THEME_MAPS = [ diff --git a/services/api/app/core/config.py b/services/api/app/core/config.py index 88f2701..e2a0168 100644 --- a/services/api/app/core/config.py +++ b/services/api/app/core/config.py @@ -36,6 +36,7 @@ class Settings(BaseSettings): model_skeptic_analyst: str = "gpt-5.4-mini" model_korea_translator: str = "gpt-5.4" model_final_judge: str = "gpt-5.4" + model_replay_calibrator: str = "gpt-5.4-mini" alpha_vantage_api_key: str | None = None polygon_api_key: str | None = None @@ -44,4 +45,3 @@ class Settings(BaseSettings): @lru_cache def get_settings() -> Settings: return Settings() - diff --git a/services/api/app/repositories/dashboard.py b/services/api/app/repositories/dashboard.py index 9aa62b3..7c3ad57 100644 --- a/services/api/app/repositories/dashboard.py +++ b/services/api/app/repositories/dashboard.py @@ -46,6 +46,7 @@ "skeptic_analyst": "리스크 분석가", "korea_translator": "한국장 번역 분석가", "final_judge": "최종 판단자", + "replay_calibrator": "리플레이 보정 분석가", } ROLE_ORDER = { @@ -55,6 +56,7 @@ "skeptic_analyst": 3, "korea_translator": 4, "final_judge": 5, + "replay_calibrator": 6, } diff --git a/services/api/app/schemas/api.py b/services/api/app/schemas/api.py index 24672a1..42b8d7a 100644 --- a/services/api/app/schemas/api.py +++ b/services/api/app/schemas/api.py @@ -172,6 +172,29 @@ class WeeklyReplayThemeOutcome(BaseModel): stockResults: list[WeeklyReplayStockResult] +class WeeklyReplayCalibrationDriver(BaseModel): + code: str + label: str + impactLabel: str + reason: str + evidence: str + + +class WeeklyReplayCalibrationAction(BaseModel): + target: str + adjustment: str + rationale: str + + +class WeeklyReplayCalibrationReview(BaseModel): + mismatchLabel: str + deviationScore: float + calibrationConfidence: float + summary: str + drivers: list[WeeklyReplayCalibrationDriver] + suggestedActions: list[WeeklyReplayCalibrationAction] + + class WeeklyReplayThemeItem(BaseModel): name: str confidence: float @@ -208,6 +231,8 @@ class WeeklyReplayDay(BaseModel): predictedThemes: list[WeeklyReplayThemeItem] evidenceItems: list[WeeklyReplayEvidenceItem] summary: str + calibrationReview: WeeklyReplayCalibrationReview + tradeOptimization: WeeklyReplayTradeOptimization class WeeklyReplayAggregate(BaseModel): @@ -216,10 +241,45 @@ class WeeklyReplayAggregate(BaseModel): avgLeaderCloseReturnPct: float +class WeeklyReplayCalibrationSummary(BaseModel): + analysisMode: str + role: str + model: str + summary: str + daysWithMeaningfulMismatch: int + averageDeviationScore: float + topDriverLabels: list[str] + suggestedAdjustments: list[WeeklyReplayCalibrationAction] + watchouts: list[str] + confidence: float + + +class WeeklyReplayTradeOptimization(BaseModel): + availabilityLabel: str + ticker: str + name: str + barInterval: str + entryAssumption: str + bestExitTimeLabel: str | None + bestExitReturnPct: float | None + closeReturnPct: float | None + extraReturnPct: float | None + chartSignal: str + executionNote: str + + +class WeeklyReplayTradeTimingSummary(BaseModel): + summary: str + averageExtraReturnPct: float + topSignalLabels: list[str] + + class WeeklyReplayResponse(BaseModel): windowLabel: str promptVersion: str aggregate: WeeklyReplayAggregate + calibrationSummary: WeeklyReplayCalibrationSummary + tradeTimingSummary: WeeklyReplayTradeTimingSummary days: list[WeeklyReplayDay] diff --git a/services/api/app/services/agents/router.py b/services/api/app/services/agents/router.py index b1aedd9..64aefc9 100644 --- a/services/api/app/services/agents/router.py +++ b/services/api/app/services/agents/router.py @@ -13,6 +13,7 @@ def __init__(self, settings: Settings) -> None: "skeptic_analyst": settings.model_skeptic_analyst, "korea_translator": settings.model_korea_translator, "final_judge": settings.model_final_judge, + "replay_calibrator": settings.model_replay_calibrator, } def model_for(self, role: str) -> str: @@ -26,5 +27,9 @@ def model_for(self, role: str) -> str: "skeptic_analyst": "과적합, 선반영, 데이터 누락, 반례 가능성을 제기하고 확신을 낮출 요소를 정리한다.", "korea_translator": "미국 검증 이벤트를 한국 상장사와 테마로 명시적으로 번역하고 1등주/2등주를 구분한다.", "final_judge": "모든 에이전트의 합의와 이견을 비교하고 근거 부족 영역을 불확실성으로 남긴다.", + "replay_calibrator": ( + "리플레이 결과에서 예측과 실제 한국장 반응이 왜 달라졌는지 분석한다. " + "반드시 JSON 객체만 반환하고 summary, keyDrivers, recommendedAdjustments, watchouts, confidence를 포함한다. " + "recommendedAdjustments는 target, adjustment, rationale 필드를 가진 객체 배열이어야 한다." + ), } - diff --git a/services/api/app/services/replay/calibration.py b/services/api/app/services/replay/calibration.py new file mode 100644 index 0000000..f6444cc --- /dev/null +++ b/services/api/app/services/replay/calibration.py @@ -0,0 +1,338 @@ +from __future__ import annotations + +from collections import Counter +from statistics import mean +from typing import Any + +from pydantic import BaseModel, Field, ValidationError + + +class CalibrationActionPayload(BaseModel): + target: str + adjustment: str + rationale: str + + +class CalibrationAgentPayload(BaseModel): + summary: str + keyDrivers: list[str] = Field(default_factory=list) + recommendedAdjustments: list[CalibrationActionPayload] = Field(default_factory=list) + watchouts: list[str] = Field(default_factory=list) + confidence: float = 0.55 + + +def build_day_calibration_review(day_item: dict[str, Any]) -> dict[str, Any]: + predicted_themes = day_item.get("predictedThemes", []) + market_context = day_item.get("marketContext", {}) + evidence_items = day_item.get("evidenceItems", []) + + if not predicted_themes: + return { + "mismatchLabel": "신호 부족", + "deviationScore": 0.18, + "calibrationConfidence": 0.52, + "summary": "당일에는 유의미한 예측 테마가 없어 보정에 활용할 만한 신호가 제한적이었습니다.", + "drivers": [ + { + "code": "no_theme_generated", + "label": "테마 신호 부족", + "impactLabel": "중간", + "reason": "장전 시점 근거 패킷에서 한국장으로 번역할 만한 강한 테마가 충분히 형성되지 않았습니다.", + "evidence": market_context.get("summary", "시장 요약 없음"), + } + ], + "suggestedActions": [ + { + "target": "테마 생성 임계치", + "adjustment": "모니터링 유지", + "rationale": "신호가 적은 날에는 억지 예측보다 공백을 허용하는 쪽이 안전합니다.", + } + ], + } + + top_theme = predicted_themes[0] + actual_outcome = top_theme.get("actualOutcome", {}) + leader_results = [item for item in actual_outcome.get("stockResults", []) if item.get("tier") == "leader"] + second_tier_results = [ + item for item in actual_outcome.get("stockResults", []) if item.get("tier") == "second_tier" + ] + avg_leader_return = float(actual_outcome.get("avgLeaderCloseReturnPct", 0.0)) + confidence = float(top_theme.get("confidence", 0.0)) + avg_confirmation = ( + mean(float(item.get("marketConfirmationScore", 0.0)) for item in evidence_items) + if evidence_items + else 0.0 + ) + kospi_return = float(market_context.get("kospiCloseReturnPct", 0.0)) + kosdaq_return = float(market_context.get("kosdaqCloseReturnPct", 0.0)) + avg_open_gap = mean(float(item.get("openGapPct", 0.0)) for item in leader_results) if leader_results else 0.0 + avg_intraday = mean(float(item.get("intradayMovePct", 0.0)) for item in leader_results) if leader_results else 0.0 + avg_second_tier_return = ( + mean(float(item.get("closeReturnPct", 0.0)) for item in second_tier_results) + if second_tier_results + else None + ) + + drivers: list[dict[str, str]] = [] + actions: list[dict[str, str]] = [] + severity = 0.22 + + def add_driver( + code: str, + label: str, + impact_label: str, + reason: str, + evidence: str, + action_target: str, + action_adjustment: str, + action_rationale: str, + floor: float, + ) -> None: + nonlocal severity + if any(existing["code"] == code for existing in drivers): + return + drivers.append( + { + "code": code, + "label": label, + "impactLabel": impact_label, + "reason": reason, + "evidence": evidence, + } + ) + actions.append( + { + "target": action_target, + "adjustment": action_adjustment, + "rationale": action_rationale, + } + ) + severity = max(severity, floor) + + if min(kospi_return, kosdaq_return) <= -2.0 and avg_leader_return <= 0.5: + add_driver( + code="broad_market_headwind", + label="시장 전반 역풍", + impact_label="높음", + reason="테마 자체보다 한국장 전체 리스크 오프 흐름이 더 강하게 작용했습니다.", + evidence=f"KOSPI {kospi_return:+.2f}% / KOSDAQ {kosdaq_return:+.2f}%", + action_target="시장 테이프 필터", + action_adjustment="하향 보정 강화", + action_rationale="장전 예측 전에 KOSPI·KOSDAQ 선행 리스크 오프 신호가 강하면 테마 점수를 감산해야 합니다.", + floor=0.82, + ) + + if confidence >= 0.72 and avg_leader_return < 1.0: + add_driver( + code="confidence_overshoot", + label="확신 과대", + impact_label="높음", + reason="예측 확신에 비해 실제 주도주 성과가 충분히 따라오지 못했습니다.", + evidence=f"예측 확신 {confidence:.2f} / 실제 평균 수익률 {avg_leader_return:+.2f}%", + action_target="최종 확신 점수", + action_adjustment="상단 캡 축소", + action_rationale="확인 점수가 약한 날에는 높은 확신을 제한해야 합니다.", + floor=0.76, + ) + + if avg_confirmation < 0.35 and confidence >= 0.65: + add_driver( + code="weak_confirmation", + label="미국 반응 확인 부족", + impact_label="중간", + reason="뉴스는 있었지만 미국 시장 반응이 충분히 동조하지 않았습니다.", + evidence=f"평균 시장 확인 점수 {avg_confirmation:.2f}", + action_target="시장 확인 점수", + action_adjustment="하한 상향", + action_rationale="교차 자산 확인이 약하면 한국 테마 번역 점수를 낮춰야 합니다.", + floor=0.68, + ) + + if avg_open_gap >= 2.5 and avg_intraday <= -1.0: + add_driver( + code="gap_fade", + label="갭상승 후 이탈", + impact_label="높음", + reason="시가에서 이미 선반영된 뒤 장중에 수급이 빠지며 수익이 축소됐습니다.", + evidence=f"평균 시가 갭 {avg_open_gap:+.2f}% / 장중 변동 {avg_intraday:+.2f}%", + action_target="갭 페이드 리스크", + action_adjustment="상향", + action_rationale="장전 강세 예측이라도 시가 과열 패턴은 별도 경고로 분리해야 합니다.", + floor=0.8, + ) + + if avg_second_tier_return is not None and avg_second_tier_return - avg_leader_return >= 2.0: + add_driver( + code="leader_selection_miss", + label="주도주 선별 미스", + impact_label="중간", + reason="같은 테마 안에서도 2등주가 주도주보다 더 강했습니다.", + evidence=f"주도주 {avg_leader_return:+.2f}% / 2등주 {avg_second_tier_return:+.2f}%", + action_target="리더 랭킹 가중치", + action_adjustment="재조정", + action_rationale="한국 수급 민감도와 단기 탄력 점수를 더 크게 반영할 필요가 있습니다.", + floor=0.64, + ) + + if confidence < 0.58 and avg_leader_return >= 3.0: + add_driver( + code="underweighted_signal", + label="신호 과소평가", + impact_label="중간", + reason="낮게 본 신호가 실제로는 강한 테마 수익으로 이어졌습니다.", + evidence=f"예측 확신 {confidence:.2f} / 실제 평균 수익률 {avg_leader_return:+.2f}%", + action_target="테마 번역 점수", + action_adjustment="상향 검토", + action_rationale="같은 유형의 사건에서 한국 민감도가 더 높았는지 재평가해야 합니다.", + floor=0.61, + ) + + if not drivers: + drivers.append( + { + "code": "aligned_result", + "label": "예측과 결과 정합", + "impactLabel": "낮음", + "reason": "주요 예측과 실제 테마 성과가 대체로 같은 방향으로 움직였습니다.", + "evidence": f"주도주 평균 수익률 {avg_leader_return:+.2f}%", + } + ) + actions.append( + { + "target": "현재 스코어링", + "adjustment": "유지", + "rationale": "현 규칙을 유지하되 같은 패턴이 반복되는지 추가 관찰하면 됩니다.", + } + ) + + mismatch_label = _mismatch_label(drivers[0]["code"], avg_leader_return) + calibration_confidence = min( + 0.92, + round(0.52 + min(len(evidence_items), 4) * 0.05 + (0.08 if leader_results else 0.0), 2), + ) + summary = ( + f"{top_theme.get('name', '상위 테마')} 예측은 {mismatch_label}으로 평가됩니다. " + f"핵심 원인은 {drivers[0]['label']}이며, {drivers[0]['evidence']}가 가장 큰 차이 신호였습니다." + ) + + return { + "mismatchLabel": mismatch_label, + "deviationScore": round(severity, 2), + "calibrationConfidence": calibration_confidence, + "summary": summary, + "drivers": drivers[:3], + "suggestedActions": _dedupe_actions(actions)[:3], + } + + +def build_weekly_calibration_summary( + day_items: list[dict[str, Any]], + role: str, + model: str, + model_payload: dict[str, Any] | None = None, +) -> dict[str, Any]: + reviews = [item["calibrationReview"] for item in day_items if item.get("calibrationReview")] + deviation_scores = [float(review["deviationScore"]) for review in reviews] + meaningful_days = [review for review in reviews if float(review["deviationScore"]) >= 0.55] + driver_counts = Counter( + driver["label"] + for review in reviews + for driver in review.get("drivers", [])[:1] + ) + top_driver_labels = [label for label, _count in driver_counts.most_common(3)] + action_pool = _dedupe_actions( + [ + action + for review in meaningful_days + for action in review.get("suggestedActions", []) + ] + )[:4] + + base_summary = ( + f"최근 {len(reviews)}거래일 중 {len(meaningful_days)}일에서 의미 있는 오차가 확인됐습니다. " + f"반복 원인은 {', '.join(top_driver_labels) if top_driver_labels else '뚜렷한 반복 없음'}입니다." + ) + watchouts = [ + "장전 강세 신호라도 KOSDAQ 급락 구간에서는 감산 필터를 우선 적용하세요.", + "시가 갭이 과하면 추격보다 15분 후 체결 강도 확인이 필요합니다.", + ] + + analysis_mode = "deterministic" + summary = base_summary + confidence = 0.58 + + if model_payload: + analysis_mode = "model_augmented" + summary = model_payload.get("summary") or base_summary + confidence = float(model_payload.get("confidence", 0.58)) + if not top_driver_labels: + top_driver_labels = model_payload.get("keyDrivers", [])[:3] + if model_payload.get("recommendedAdjustments"): + action_pool = _dedupe_actions( + [*action_pool, *model_payload["recommendedAdjustments"]] + )[:4] + if model_payload.get("watchouts"): + watchouts = model_payload["watchouts"][:4] + + return { + "analysisMode": analysis_mode, + "role": role, + "model": model, + "summary": summary, + "daysWithMeaningfulMismatch": len(meaningful_days), + "averageDeviationScore": round(mean(deviation_scores), 2) if deviation_scores else 0.0, + "topDriverLabels": top_driver_labels, + "suggestedAdjustments": action_pool, + "watchouts": watchouts, + "confidence": round(confidence, 2), + } + + +def normalize_calibration_agent_payload(raw: dict[str, Any]) -> dict[str, Any] | None: + try: + payload = CalibrationAgentPayload.model_validate(raw) + except ValidationError: + return None + + return { + "summary": payload.summary, + "keyDrivers": payload.keyDrivers[:4], + "recommendedAdjustments": [ + { + "target": action.target, + "adjustment": action.adjustment, + "rationale": action.rationale, + } + for action in payload.recommendedAdjustments[:4] + ], + "watchouts": payload.watchouts[:4], + "confidence": round(float(payload.confidence), 2), + } + + +def _dedupe_actions(actions: list[dict[str, str]]) -> list[dict[str, str]]: + deduped: list[dict[str, str]] = [] + seen: set[tuple[str, str]] = set() + for action in actions: + key = (action["target"], action["adjustment"]) + if key in seen: + continue + seen.add(key) + deduped.append(action) + return deduped + + +def _mismatch_label(primary_code: str, avg_leader_return: float) -> str: + if primary_code == "broad_market_headwind": + return "시장 역풍" + if primary_code == "gap_fade": + return "갭상승 후 이탈" + if primary_code == "leader_selection_miss": + return "주도주 선별 미스" + if primary_code == "underweighted_signal": + return "신호 과소평가" + if primary_code in {"confidence_overshoot", "weak_confirmation"}: + return "과신 실패" + if avg_leader_return >= 1.0: + return "정합" + return "보통 오차" diff --git a/services/api/app/services/replay/trade_timing.py b/services/api/app/services/replay/trade_timing.py new file mode 100644 index 0000000..32d0813 --- /dev/null +++ b/services/api/app/services/replay/trade_timing.py @@ -0,0 +1,162 @@ +from __future__ import annotations + +from statistics import mean +from typing import Any + +import pandas as pd + + +def analyze_trade_optimization( + market_date_label: str, + ticker: str, + name: str, + intraday: pd.DataFrame, +) -> dict[str, Any]: + if intraday.empty: + return _unavailable_trade_optimization(ticker=ticker, name=name) + + bars = intraday.sort_index().copy() + if "Open" not in bars or "High" not in bars or "Close" not in bars: + return _unavailable_trade_optimization(ticker=ticker, name=name) + + entry_price = float(bars.iloc[0]["Open"]) + if entry_price <= 0: + return _unavailable_trade_optimization(ticker=ticker, name=name) + + high_at = bars["High"].astype(float).idxmax() + high_row = bars.loc[high_at] + best_exit_price = float(high_row["High"]) + close_price = float(bars.iloc[-1]["Close"]) + best_exit_return = round(((best_exit_price - entry_price) / entry_price) * 100, 2) + close_return = round(((close_price - entry_price) / entry_price) * 100, 2) + extra_return = round(best_exit_return - close_return, 2) + + vwap = _vwap_series(bars) + post_peak = bars.loc[high_at:] + vwap_break = None + if not post_peak.empty: + for timestamp, row in post_peak.iloc[1:].iterrows(): + if float(row["Close"]) < float(vwap.loc[timestamp]): + vwap_break = timestamp + break + + signal_label = _signal_label(high_at, bars.index[-1], extra_return, vwap_break is not None) + execution_note = _execution_note( + market_date_label=market_date_label, + high_at=high_at, + best_exit_return=best_exit_return, + close_return=close_return, + extra_return=extra_return, + vwap_break=vwap_break, + ) + + return { + "availabilityLabel": "available", + "ticker": ticker, + "name": name, + "barInterval": "5m", + "entryAssumption": "시가 진입 가정", + "bestExitTimeLabel": high_at.tz_convert("Asia/Seoul").strftime("%H:%M KST") + if getattr(high_at, "tzinfo", None) + else high_at.strftime("%H:%M KST"), + "bestExitReturnPct": best_exit_return, + "closeReturnPct": close_return, + "extraReturnPct": extra_return, + "chartSignal": signal_label, + "executionNote": execution_note, + } + + +def summarize_trade_optimizations(day_items: list[dict[str, Any]]) -> dict[str, Any]: + available = [ + item["tradeOptimization"] + for item in day_items + if item.get("tradeOptimization", {}).get("availabilityLabel") == "available" + ] + if not available: + return { + "summary": "장중 5분봉 데이터가 부족해 최적 매도 타이밍을 요약하지 못했습니다.", + "averageExtraReturnPct": 0.0, + "topSignalLabels": [], + } + + signal_counts: dict[str, int] = {} + for item in available: + signal_counts[item["chartSignal"]] = signal_counts.get(item["chartSignal"], 0) + 1 + ordered_signals = [label for label, _count in sorted(signal_counts.items(), key=lambda pair: (-pair[1], pair[0]))] + avg_extra_return = round(mean(float(item["extraReturnPct"]) for item in available), 2) + + return { + "summary": ( + f"시가 진입 가정 기준으로 종가까지 보유하는 것보다 평균 {avg_extra_return:+.2f}%p 더 유리했던 " + f"매도 힌트는 {', '.join(ordered_signals[:2]) if ordered_signals else '없음'}이었습니다." + ), + "averageExtraReturnPct": avg_extra_return, + "topSignalLabels": ordered_signals[:3], + } + + +def _signal_label(high_at, last_at, extra_return: float, has_vwap_break: bool) -> str: + if extra_return <= 0.3: + return "종가 보유 우위" + hour = int(high_at.strftime("%H")) + minute = int(high_at.strftime("%M")) + total_minutes = hour * 60 + minute + + if total_minutes <= 10 * 60: + return "시초 급등 고점" + if has_vwap_break: + return "고점 후 VWAP 이탈" + if total_minutes <= 11 * 60 + 30: + return "오전 추세 둔화" + if high_at == last_at: + return "종가 직전 재가속" + return "오후 재돌파" + + +def _execution_note( + market_date_label: str, + high_at, + best_exit_return: float, + close_return: float, + extra_return: float, + vwap_break, +) -> str: + base = ( + f"{market_date_label}에는 5분봉 기준 {high_at.strftime('%H:%M')} 부근이 최대 수익 구간이었고 " + f"시가 대비 최고 {best_exit_return:+.2f}%까지 도달했습니다." + ) + if extra_return <= 0.3: + return f"{base} 종가 수익률도 {close_return:+.2f}%로 비슷해 종가 보유가 크게 불리하지 않았습니다." + if vwap_break is not None: + return ( + f"{base} 이후 {vwap_break.strftime('%H:%M')} 무렵 VWAP 하향 이탈이 나타나 " + f"종가 보유 대비 {extra_return:+.2f}%p 더 유리했습니다." + ) + return f"{base} 종가까지 보유했을 때보다 {extra_return:+.2f}%p 더 높은 수익을 확보할 수 있었습니다." + + +def _unavailable_trade_optimization(ticker: str, name: str) -> dict[str, Any]: + return { + "availabilityLabel": "unavailable", + "ticker": ticker, + "name": name, + "barInterval": "5m", + "entryAssumption": "시가 진입 가정", + "bestExitTimeLabel": None, + "bestExitReturnPct": None, + "closeReturnPct": None, + "extraReturnPct": None, + "chartSignal": "데이터 부족", + "executionNote": "해당 일자의 장중 데이터를 확보하지 못해 최적 매도 시점을 계산하지 못했습니다.", + } + + +def _vwap_series(bars: pd.DataFrame) -> pd.Series: + volume = bars["Volume"].fillna(0).astype(float) + typical_price = ( + bars["High"].astype(float) + bars["Low"].astype(float) + bars["Close"].astype(float) + ) / 3 + cumulative_volume = volume.cumsum().replace(0, pd.NA).ffill() + vwap = (typical_price * volume).cumsum() / cumulative_volume + return vwap.ffill().fillna(bars["Close"].astype(float)) diff --git a/services/api/app/services/replay/weekly_replay.py b/services/api/app/services/replay/weekly_replay.py index 2e3fe3d..06d2bd9 100644 --- a/services/api/app/services/replay/weekly_replay.py +++ b/services/api/app/services/replay/weekly_replay.py @@ -32,6 +32,14 @@ ) from app.services.korea.translation_engine import KoreaTranslationEngine, RankedTheme, merge_ranked_themes from app.services.market.reaction_engine import YahooMarketReactionEngine, _numeric_series +from app.services.replay.calibration import ( + build_day_calibration_review, + build_weekly_calibration_summary, + normalize_calibration_agent_payload, +) +from app.services.replay.trade_timing import analyze_trade_optimization, summarize_trade_optimizations +from app.services.agents.openai_client import AgentOrchestrator +from app.core.config import get_settings SEOUL = ZoneInfo("Asia/Seoul") HTTP_HEADERS = {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36"} @@ -106,10 +114,13 @@ def sort_day_items_latest_first(day_items: list[dict[str, object]]) -> list[dict class WeeklyReplayService: def __init__(self) -> None: + self.settings = get_settings() self.reaction_engine = YahooMarketReactionEngine() self.translation_engine = KoreaTranslationEngine() + self.agent_orchestrator = AgentOrchestrator(self.settings) self._price_cache: dict[str, pd.DataFrame] = {} self._price_cache_bounds: dict[str, tuple[date, date]] = {} + self._intraday_cache: dict[tuple[str, date], pd.DataFrame] = {} async def build(self, session: AsyncSession, days: int = 7) -> dict[str, object]: cache_key = await self._cache_key(session, days) @@ -128,6 +139,23 @@ async def build(self, session: AsyncSession, days: int = 7) -> dict[str, object] "windowLabel": "최근 0거래일", "promptVersion": prompt_version, "aggregate": {"daysAnalyzed": 0, "positiveHitDays": 0, "avgLeaderCloseReturnPct": 0.0}, + "calibrationSummary": { + "analysisMode": "deterministic", + "role": "replay_calibrator", + "model": self.agent_orchestrator.router.model_for("replay_calibrator"), + "summary": "분석할 거래일이 없어 보정 요약을 만들지 않았습니다.", + "daysWithMeaningfulMismatch": 0, + "averageDeviationScore": 0.0, + "topDriverLabels": [], + "suggestedAdjustments": [], + "watchouts": [], + "confidence": 0.0, + }, + "tradeTimingSummary": { + "summary": "분석할 거래일이 없어 매도 타이밍 요약을 만들지 않았습니다.", + "averageExtraReturnPct": 0.0, + "topSignalLabels": [], + }, "days": [], } _WEEKLY_REPLAY_CACHE[cache_key] = (now, result) @@ -208,15 +236,133 @@ async def build(self, session: AsyncSession, days: int = 7) -> dict[str, object] ) ordered_day_items = sort_day_items_latest_first(day_items) + for day_item in ordered_day_items: + day_item["calibrationReview"] = build_day_calibration_review(day_item) + await self._attach_trade_optimizations(ordered_day_items, security_map) + + calibration_summary = await self._build_calibration_summary( + prompt_version=prompt_version, + day_items=ordered_day_items, + ) + trade_timing_summary = summarize_trade_optimizations(ordered_day_items) result = { "windowLabel": f"최근 {len(day_items)}거래일", "promptVersion": prompt_version, "aggregate": self._aggregate(ordered_day_items), + "calibrationSummary": calibration_summary, + "tradeTimingSummary": trade_timing_summary, "days": ordered_day_items, } _WEEKLY_REPLAY_CACHE[cache_key] = (now, result) return result + async def _build_calibration_summary( + self, + prompt_version: str, + day_items: list[dict[str, object]], + ) -> dict[str, object]: + role = "replay_calibrator" + model = self.agent_orchestrator.router.model_for(role) + evidence_pack = { + "prompt_version": prompt_version, + "days": [ + { + "market_date": item["marketDateLabel"], + "summary": item["summary"], + "market_context": item["marketContext"], + "top_theme": item["predictedThemes"][0]["name"] if item["predictedThemes"] else None, + "calibration_review": item["calibrationReview"], + } + for item in day_items + ], + "aggregate": self._aggregate(day_items), + } + + model_payload: dict[str, object] | None = None + if self.settings.openai_api_key: + result = await self.agent_orchestrator.run_role( + role=role, + evidence_pack=evidence_pack, + background=False, + ) + model_payload = normalize_calibration_agent_payload(result.content) + + return build_weekly_calibration_summary( + day_items=day_items, + role=role, + model=model, + model_payload=model_payload, + ) + + async def _attach_trade_optimizations( + self, + day_items: list[dict[str, object]], + security_map: dict[str, str], + ) -> None: + targets: list[tuple[dict[str, object], str, str, str, date]] = [] + for day_item in day_items: + market_date = datetime.strptime(str(day_item["marketDateLabel"]), "%Y-%m-%d").date() + top_theme = day_item["predictedThemes"][0] if day_item["predictedThemes"] else None + if not top_theme: + day_item["tradeOptimization"] = { + "availabilityLabel": "unavailable", + "ticker": "-", + "name": "데이터 없음", + "barInterval": "5m", + "entryAssumption": "시가 진입 가정", + "bestExitTimeLabel": None, + "bestExitReturnPct": None, + "closeReturnPct": None, + "extraReturnPct": None, + "chartSignal": "테마 부재", + "executionNote": "상위 예측 테마가 없어 매도 최적화 분석을 생략했습니다.", + } + continue + + leader_results = [ + stock for stock in top_theme["actualOutcome"]["stockResults"] if stock["tier"] == "leader" + ] + selected = max( + leader_results or top_theme["actualOutcome"]["stockResults"], + key=lambda stock: stock["closeReturnPct"], + default=None, + ) + if selected is None: + day_item["tradeOptimization"] = { + "availabilityLabel": "unavailable", + "ticker": "-", + "name": "데이터 없음", + "barInterval": "5m", + "entryAssumption": "시가 진입 가정", + "bestExitTimeLabel": None, + "bestExitReturnPct": None, + "closeReturnPct": None, + "extraReturnPct": None, + "chartSignal": "데이터 부족", + "executionNote": "장중 가격 데이터가 없어 최적 매도 시점을 계산하지 못했습니다.", + } + continue + + market = security_map.get(selected["ticker"], "KOSPI") + symbol = f"{selected['ticker']}.KS" if market == "KOSPI" else f"{selected['ticker']}.KQ" + targets.append((day_item, symbol, selected["ticker"], selected["name"], market_date)) + + await asyncio.gather( + *[ + asyncio.to_thread(self._intraday_for_symbol_day, symbol, market_date) + for _day_item, symbol, _ticker, _name, market_date in targets + ] + ) + + for day_item, symbol, ticker, name, market_date in targets: + intraday = self._intraday_for_symbol_day(symbol, market_date) + day_item["tradeOptimization"] = analyze_trade_optimization( + market_date_label=str(day_item["marketDateLabel"]), + ticker=ticker, + name=name, + intraday=intraday, + ) + async def _prefetch_histories(self, symbols: set[str], start: date, end: date) -> None: await asyncio.gather( *[ @@ -606,6 +752,35 @@ def _history_for_symbol(self, symbol: str, start: date, end: date) -> pd.DataFra self._price_cache_bounds[symbol] = (fetch_start, fetch_end) return frame[(frame.index >= start) & (frame.index < end)] + def _intraday_for_symbol_day(self, symbol: str, market_date: date) -> pd.DataFrame: + cache_key = (symbol, market_date) + cached = self._intraday_cache.get(cache_key) + if cached is not None: + return cached + + ticker = yf.Ticker(symbol) + frame = ticker.history( + interval="5m", + start=market_date.isoformat(), + end=(market_date + timedelta(days=1)).isoformat(), + auto_adjust=False, + actions=False, + prepost=False, + ) + if frame.empty: + self._intraday_cache[cache_key] = frame + return frame + + frame = frame.copy() + frame.columns = [str(column) for column in frame.columns] + if getattr(frame.index, "tz", None) is None: + frame.index = pd.to_datetime(frame.index).tz_localize("Asia/Seoul") + else: + frame.index = pd.to_datetime(frame.index).tz_convert("Asia/Seoul") + frame = frame[frame.index.date == market_date] + self._intraday_cache[cache_key] = frame + return frame + def _evidence_pack_hash(self, evidence_items: list[dict[str, object]]) -> str: encoded = json.dumps(evidence_items, ensure_ascii=False, sort_keys=True).encode("utf-8") return hashlib.sha256(encoded).hexdigest() diff --git a/services/api/tests/test_replay_analysis_helpers.py b/services/api/tests/test_replay_analysis_helpers.py new file mode 100644 index 0000000..393af2f --- /dev/null +++ b/services/api/tests/test_replay_analysis_helpers.py @@ -0,0 +1,98 @@ +from __future__ import annotations + +import pandas as pd + +from app.services.replay.calibration import build_day_calibration_review +from app.services.replay.trade_timing import analyze_trade_optimization, summarize_trade_optimizations + + +def test_build_day_calibration_review_flags_market_headwind_and_gap_fade() -> None: + review = build_day_calibration_review( + { + "marketContext": { + "kospiCloseReturnPct": -2.8, + "kosdaqCloseReturnPct": -3.6, + "summary": "KOSPI -2.80% / KOSDAQ -3.60%", + }, + "evidenceItems": [{"marketConfirmationScore": 0.21}], + "predictedThemes": [ + { + "name": "HBM/반도체 장비", + "confidence": 0.82, + "actualOutcome": { + "avgLeaderCloseReturnPct": -2.4, + "stockResults": [ + { + "ticker": "042700", + "name": "한미반도체", + "tier": "leader", + "openGapPct": 3.4, + "closeReturnPct": -2.4, + "intradayMovePct": -4.6, + } + ], + }, + } + ], + } + ) + + assert review["mismatchLabel"] == "시장 역풍" + assert any(driver["code"] == "broad_market_headwind" for driver in review["drivers"]) + assert any(action["target"] == "시장 테이프 필터" for action in review["suggestedActions"]) + + +def test_analyze_trade_optimization_finds_best_exit_before_close() -> None: + bars = pd.DataFrame( + { + "Open": [100.0, 104.0, 105.0, 103.0], + "High": [104.0, 108.0, 106.0, 103.0], + "Low": [99.5, 103.5, 102.0, 101.0], + "Close": [104.0, 105.0, 103.0, 102.0], + "Volume": [1000, 1400, 1300, 1200], + }, + index=pd.date_range("2026-04-08 09:05", periods=4, freq="5min", tz="Asia/Seoul"), + ) + + optimization = analyze_trade_optimization( + market_date_label="2026-04-08", + ticker="042700", + name="한미반도체", + intraday=bars, + ) + + assert optimization["availabilityLabel"] == "available" + assert optimization["bestExitTimeLabel"] == "09:10 KST" + assert optimization["bestExitReturnPct"] > optimization["closeReturnPct"] + assert optimization["extraReturnPct"] > 0 + + +def test_summarize_trade_optimizations_aggregates_signal_labels() -> None: + summary = summarize_trade_optimizations( + [ + { + "tradeOptimization": { + "availabilityLabel": "available", + "chartSignal": "고점 후 VWAP 이탈", + "extraReturnPct": 2.4, + } + }, + { + "tradeOptimization": { + "availabilityLabel": "available", + "chartSignal": "고점 후 VWAP 이탈", + "extraReturnPct": 1.1, + } + }, + { + "tradeOptimization": { + "availabilityLabel": "available", + "chartSignal": "시초 급등 고점", + "extraReturnPct": 0.7, + } + }, + ] + ) + + assert summary["topSignalLabels"][0] == "고점 후 VWAP 이탈" + assert summary["averageExtraReturnPct"] == 1.4