[AI] [FIX] infer_input_shape_from_legacy_h5() 메서드 추가 (#333)

* [AI] [FIX] infer_input_shape_from_legacy_h5() 메서드 추가

* [AI] [FIX] 매매로직 치명적 버그 수정

* [AI] [FEAT] 백테스트 코드 작업 및 로컬 LLM 모델 변경

* [AI] [REFACT] 데이터 정리/검증 동작 강화

Author: twq110

.gitignore

@@ -31,6 +31,7 @@ __pycache__/
 AI/.venv/
 AI/data/weights/tcn/
 AI/config/trading.local.json
+AI/tests/out/
 
 # ===== Backend =====
 backend/src/main/java/org/sejongisc/backend/stock/TestController.java

AI/config/watchlist.json

@@ -2,33 +2,33 @@
     "tickers": [
         "NVDA",
         "TSLA",
-        "MU",
-        "SNDK",
-        "MSFT",
-        "AVGO",
         "AAPL",
-        "AMZN",
+        "PLTR",
+        "MSFT",
         "META",
+        "AMZN",
         "GOOGL",
-        "PLTR",
-        "AMD",
-        "ORCL",
+        "AVGO",
+        "SMCI",
+        "APP",
         "GOOG",
-        "INTC",
-        "XOM",
         "NFLX",
-        "CVX",
-        "JPM",
-        "WDC",
-        "CRM",
+        "AMD",
+        "UNH",
         "LLY",
-        "WMT",
-        "LRCX",
-        "APP",
+        "CRM",
+        "BRK-B",
+        "V",
         "COIN",
+        "JPM",
+        "WMT",
+        "INTC",
+        "HOOD",
+        "COST",
+        "MU",
         "BAC",
-        "ADBE",
-        "BRK-B",
-        "V"
+        "BKNG",
+        "GEV",
+        "HD"
     ]
 }

AI/libs/database/repository.py

@@ -5,6 +5,7 @@
 - 기존의 개별 함수들을 PortfolioRepository 클래스로 통합하여 응집도를 높이고 DB 커넥션 관리를 최적화했습니다.
 """
 
+import os
 from typing import List, Tuple, Optional, Any, Dict
 import pandas as pd
 import psycopg2
@@ -108,7 +109,8 @@ def get_current_position(self, ticker: str, target_date: str = None, initial_cas
             params.append(target_date)
 
         # 시간순으로 정렬하여 정확한 롤링 계산을 수행합니다.
-        query += " ORDER BY fill_date ASC, created_at ASC"
+        # Deterministic ordering is important for reproducible position/cost accounting.
+        query += " ORDER BY fill_date ASC, created_at ASC, id ASC"
 
         try:
             cursor.execute(query, tuple(params))
@@ -171,42 +173,138 @@ def get_current_position(self, ticker: str, target_date: str = None, initial_cas
 
     def get_current_cash(self, target_date: str = None, initial_cash: float = 10000000) -> float:
         """
-        [현재 포트폴리오 현금 조회] 
-        포트폴리오 요약 테이블에서 target_date 이전의 가장 최근 현금 잔고를 조회합니다.
-        
-        Args:
-            target_date (str, optional): 기준 날짜
-            initial_cash (float): 내역이 없을 경우 반환할 초기 자본금
-            
-        Returns:
-            float: 계산된 현재 현금 잔고
+        executions 테이블의 cash_after를 우선 사용해 기준 시점 현금을 조회합니다.
+        데이터가 없거나 조회에 실패하면 보수적으로 초기 현금(initial_cash)을 반환합니다.
         """
         conn = self._get_connection()
         if conn is None:
-            return initial_cash
-            
-        cursor = conn.cursor()
-        
-        # 기준일 이전의 가장 최근 마감 데이터를 가져오는 서브쿼리 활용
+            return float(initial_cash)
+
+        try:
+            with conn.cursor() as cursor:
+                if target_date:
+                    exec_cash_query = """
+                        SELECT cash_after
+                        FROM public.executions
+                        WHERE fill_date <= %s
+                        ORDER BY fill_date DESC, created_at DESC, id DESC
+                        LIMIT 1
+                    """
+                    cursor.execute(exec_cash_query, (target_date,))
+                    exec_cash = cursor.fetchone()
+                    if exec_cash and exec_cash[0] is not None:
+                        return float(exec_cash[0])
+
+                summary_cash_query = """
+                    SELECT cash
+                    FROM public.portfolio_summary
+                    WHERE date = (
+                        SELECT MAX(date)
+                        FROM public.portfolio_summary
+                        WHERE date < %s
+                    )
+                    LIMIT 1
+                """
+                cursor.execute(summary_cash_query, (target_date,))
+                summary_cash = cursor.fetchone()
+                if summary_cash and summary_cash[0] is not None:
+                    return float(summary_cash[0])
+                return float(initial_cash)
+        except Exception as e:
+            print(f"[PortfolioRepository][Error] 현재 현금 조회 실패: {e}")
+            return float(initial_cash)
+        finally:
+            conn.close()
+
+    def get_open_tickers(self, target_date: str) -> List[str]:
+        """
+        target_date 이전까지의 누적 순수량(net 포지션)이 0보다 큰 티커 목록을 반환합니다.
+        """
+        conn = self._get_connection()
+        if conn is None:
+            return []
+
         query = """
-            SELECT cash 
-            FROM public.portfolio_summary 
-            WHERE date = (SELECT MAX(date) FROM public.portfolio_summary WHERE date < %s)
-            LIMIT 1;
+            SELECT ticker
+            FROM public.executions
+            WHERE fill_date <= %s
+            GROUP BY ticker
+            HAVING SUM(
+                CASE
+                    WHEN side = 'BUY' THEN qty
+                    WHEN side = 'SELL' THEN -qty
+                    ELSE 0
+                END
+            ) > 0
+            ORDER BY ticker
         """
-        
+
+        try:
+            with conn.cursor() as cursor:
+                cursor.execute(query, (target_date,))
+                rows = cursor.fetchall()
+            return [str(row[0]) for row in rows]
+        except Exception as e:
+            print(f"[PortfolioRepository][Error] Open 티커 목록 조회 실패: {e}")
+            return []
+        finally:
+            conn.close()
+
+    def reset_run_data(self, run_id: str, target_date: Optional[str] = None) -> None:
+        """
+        Remove stale simulation artifacts before a rerun.
+
+        If target_date is provided, downstream dates are also removed so cash/position
+        chains are recalculated consistently from target_date forward.
+        """
+        if not run_id:
+            return
+
+        conn = self._get_connection()
+        if conn is None:
+            return
+
         try:
-            cursor.execute(query, (target_date,))
-            result = cursor.fetchone()
-            if result:
-                return float(result[0])
+            with conn.cursor() as cursor:
+                if target_date:
+                    # Safety-first default: only clear current run artifacts.
+                    # Global chain reset can remove unrelated simulations sharing the same DB.
+                    allow_global_chain_reset = os.environ.get("AI_ALLOW_GLOBAL_CHAIN_RESET", "0") == "1"
+                    if allow_global_chain_reset:
+                        cursor.execute(
+                            "DELETE FROM public.executions WHERE fill_date >= %s AND run_id LIKE 'daily_%%'",
+                            (target_date,),
+                        )
+                        cursor.execute(
+                            "DELETE FROM public.xai_reports WHERE date >= %s AND run_id LIKE 'daily_%%'",
+                            (target_date,),
+                        )
+                        cursor.execute("DELETE FROM public.portfolio_positions WHERE date >= %s", (target_date,))
+                        cursor.execute("DELETE FROM public.portfolio_summary WHERE date >= %s", (target_date,))
+                    else:
+                        cursor.execute("DELETE FROM public.executions WHERE run_id = %s", (run_id,))
+                        cursor.execute("DELETE FROM public.xai_reports WHERE run_id = %s", (run_id,))
+                else:
+                    cursor.execute("DELETE FROM public.executions WHERE run_id = %s", (run_id,))
+                    cursor.execute("DELETE FROM public.xai_reports WHERE run_id = %s", (run_id,))
+            conn.commit()
+            if target_date:
+                if os.environ.get("AI_ALLOW_GLOBAL_CHAIN_RESET", "0") == "1":
+                    print(
+                        f"[PortfolioRepository] Reset simulation rows from {target_date} onward "
+                        f"(triggered by run_id={run_id})."
+                    )
+                else:
+                    print(
+                        f"[PortfolioRepository] Reset current run artifacts only "
+                        f"(run_id={run_id}, target_date={target_date})."
+                    )
             else:
-                return initial_cash
+                print(f"[PortfolioRepository] Reset existing records for run_id={run_id}.")
         except Exception as e:
-            print(f"[PortfolioRepository][Error] 포트폴리오 현금 조회 중 오류 발생: {e}")
-            return initial_cash
+            conn.rollback()
+            print(f"[PortfolioRepository][Error] reset_run_data failed: {e}")
         finally:
-            cursor.close()
             conn.close()
 
     def save_executions_to_db(self, fills_df: pd.DataFrame) -> None:
@@ -231,9 +329,31 @@ def save_executions_to_db(self, fills_df: pd.DataFrame) -> None:
 
         if not required_cols.issubset(fills_df.columns):
             missing = required_cols - set(fills_df.columns)
-            print(f"[PortfolioRepository][Error] 체결 내역 데이터에 필수 컬럼이 누락되었습니다: {missing}")
+            print(f"[PortfolioRepository][Error] Missing required execution columns: {missing}")
             return
 
+        def _normalize_run_id(value: Any) -> Optional[str]:
+            if pd.isna(value):
+                return None
+            normalized = str(value).strip()
+            if not normalized:
+                return None
+            if normalized.lower() in {"nan", "none", "null"}:
+                return None
+            return normalized
+
+        normalized_run_ids = fills_df["run_id"].apply(_normalize_run_id)
+        missing_mask = normalized_run_ids.isna()
+        if bool(missing_mask.any()):
+            sample_rows = fills_df.loc[missing_mask, ["ticker", "signal_date", "fill_date"]].head(5)
+            raise ValueError(
+                "[PortfolioRepository][Error] run_id must be non-empty for all execution rows. "
+                f"missing rows sample={sample_rows.to_dict(orient='records')}"
+            )
+
+        fills_df = fills_df.copy()
+        fills_df["run_id"] = normalized_run_ids
+
         conn = self._get_connection()
         if conn is None:
             print("[PortfolioRepository][Error] DB 연결에 실패하여 체결 내역을 저장할 수 없습니다.")
@@ -242,6 +362,15 @@ def save_executions_to_db(self, fills_df: pd.DataFrame) -> None:
         cursor = conn.cursor()
 
         try:
+            run_ids = sorted(
+                {
+                    str(run_id).strip()
+                    for run_id in fills_df["run_id"].tolist()
+                    if pd.notna(run_id) and str(run_id).strip()
+                }
+            )
+            if run_ids:
+                cursor.execute("DELETE FROM public.executions WHERE run_id = ANY(%s)", (run_ids,))
             # 다량의 데이터를 빠르게 넣기 위한 INSERT 구문 준비
             insert_query = """
                 INSERT INTO public.executions (
@@ -296,7 +425,7 @@ def save_executions_to_db(self, fills_df: pd.DataFrame) -> None:
             cursor.close()
             conn.close()
 
-    def save_reports_to_db(self, reports_tuple_list: list) -> list:
+    def save_reports_to_db(self, reports_tuple_list: list, run_id: Optional[str] = None) -> list:
         """
         [XAI 리포트 일괄 저장]
         설명 가능한 AI(XAI) 분석 리포트들을 DB에 저장하고, 생성된 Primary Key(ID) 리스트를 반환합니다.
@@ -320,16 +449,26 @@ def save_reports_to_db(self, reports_tuple_list: list) -> list:
             # RETURNING id 절을 사용하여 INSERT 후 자동 생성된 PK를 반환받습니다.
             insert_query = """
                 INSERT INTO public.xai_reports (
-                    ticker, signal, price, date, report
+                    ticker, signal, price, date, report, run_id
                 ) VALUES %s
+                ON CONFLICT (ticker, date, signal) DO UPDATE
+                SET price = EXCLUDED.price,
+                    report = EXCLUDED.report,
+                    run_id = EXCLUDED.run_id,
+                    created_at = NOW()
                 RETURNING id
             """
 
             # fetch=True 옵션으로 execute_values 실행 시 RETURNING 결과를 리스트 형태로 모아줍니다.
+            reports_with_run = []
+            for row in reports_tuple_list:
+                ticker, signal, price, date, report = row
+                reports_with_run.append((ticker, signal, price, date, report, run_id))
+
             result_ids = execute_values(
-                cursor, 
-                insert_query, 
-                reports_tuple_list,
+                cursor,
+                insert_query,
+                reports_with_run,
                 fetch=True
             )
 
@@ -419,4 +558,4 @@ def save_portfolio_positions(self, date: str, data_tuples: list):
             print(f"[PortfolioRepository][Error] 포지션 저장 실패: {e}")
         finally:
             cursor.close()
-            conn.close()
+            conn.close()

AI/libs/llm/base_client.py

@@ -1,36 +1,35 @@
 # AI/libs/llm/base_client.py
 """
-[LLM 클라이언트 인터페이스]
-- 모든 LLM 서비스(Groq, Ollama 등)가 준수해야 할 공통 규약을 정의합니다.
-- 이를 통해 모델 교체 시 코드 수정을 최소화할 수 있습니다.
+Shared base interface for all LLM clients.
 """
 
 from abc import ABC, abstractmethod
-from typing import Dict, Any, Optional
+from typing import Any, Optional
+
 
 class BaseLLMClient(ABC):
-    """모든 LLM 클라이언트의 추상 기본 클래스"""
+    """Abstract base class for all LLM clients."""
 
     def __init__(self, api_key: Optional[str] = None, model_name: str = "default"):
         self.api_key = api_key
         self.model_name = model_name
+        self.last_error: Optional[str] = None
+
+    def clear_last_error(self) -> None:
+        self.last_error = None
+
+    def set_last_error(self, error: Any) -> None:
+        if isinstance(error, Exception):
+            self.last_error = str(error) or error.__class__.__name__
+            return
+        self.last_error = str(error)
 
     @abstractmethod
     def generate_text(self, prompt: str, system_prompt: Optional[str] = None, **kwargs) -> str:
-        """
-        단일 프롬프트를 입력받아 텍스트 응답을 생성합니다.
-        
-        Args:
-            prompt (str): 사용자 입력 프롬프트
-            system_prompt (str, optional): 시스템 프롬프트 (역할 정의 등)
-            **kwargs: 모델별 추가 파라미터 (temperature 등)
-            
-        Returns:
-            str: 생성된 텍스트
-        """
-        pass
-    
+        """Generate text from prompt."""
+        raise NotImplementedError
+
     @abstractmethod
     def get_health(self) -> bool:
-        """서비스 상태를 확인합니다."""
-        pass
+        """Return whether the LLM service is available."""
+        raise NotImplementedError