feat: Lifetime PD (50yr) - Belkin & Suchower + Vasicek model

- Belkin & Suchower (1998) credit cycle index (Zt) estimation via WLS - Vasicek single-factor conditional PD/TM model - Macro-Zt OLS regression with stepwise variable selection - 3-scenario (boom/neutral/recession) 50yr PD projection - Statistical validation suite (ADF, Ljung-Box, R2, ARCH) - BOK ECOS API integration with fallback data - Visualization module (7 chart types) - Detailed theoretical methodology docs/methodology.md
2026-03-10 21:57:34 +09:00
commit 3a9374c61a
39 changed files with 4671 additions and 0 deletions
--- a/projection/init.py
+++ b/projection/init.py
@@ -0,0 +1 @@
+# Projection: 50년 Lifetime PD 산출
--- a/projection/lifetime_pd.py
+++ b/projection/lifetime_pd.py
@@ -0,0 +1,281 @@
+"""
+50년 Lifetime PD 산출 엔진
+
+시나리오별 Zt 경로와 TTC 전이행렬을 결합하여:
+1. 연도별 조건부 전이행렬 산출
+2. 순차적 행렬 곱으로 누적 전이확률 계산
+3. Marginal PD / Cumulative PD / 시나리오 가중평균 PD 산출
+
+IFRS 9 ECL에 직접 사용 가능한 PD Term Structure 출력
+
+참고문헌:
+- IFRS 9 Financial Instruments (IASB, 2014)
+- EBA Guidelines on IFRS 9 implementation
+- Basel Committee BCBS 350 (Credit Risk)
+"""
+
+import numpy as np
+import pandas as pd
+from typing import Dict, List, Optional, Tuple
+from models.vasicek import conditional_transition_matrix
+from data.transition_matrices import RATING_GRADES
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class LifetimePDEngine:
+    """
+    50년 Lifetime PD 산출 엔진
+    
+    Process:
+    1. 각 연도 t에 대해 Zt로 조건부 전이행렬 TM(Zt) 산출
+    2. 누적 전이행렬 = TM(Z1) × TM(Z2) × ... × TM(Zt)
+    3. 누적 전이행렬의 D열이 누적 PD
+    4. 한계 PD = Cumulative PD(t) - Cumulative PD(t-1)
+    """
+    
+    def __init__(
+        self,
+        ttc_matrix: np.ndarray,
+        rho: float = 0.20,
+        rating_grades: List[str] = None
+    ):
+        """
+        Parameters
+        ----------
+        ttc_matrix : np.ndarray
+            N×N TTC 전이행렬
+        rho : float
+            자산상관계수
+        rating_grades : List[str]
+            등급 레이블
+        """
+        self.ttc_matrix = ttc_matrix
+        self.rho = rho
+        self.n_grades = ttc_matrix.shape[0]
+        self.grades = rating_grades or RATING_GRADES
+        self.non_default_grades = self.grades[:-1]  # D 제외
+    
+    def compute_lifetime_pd(
+        self,
+        z_path: np.ndarray,
+        horizon: Optional[int] = None
+    ) -> Dict[str, np.ndarray]:
+        """
+        단일 시나리오의 Lifetime PD 산출
+        
+        Parameters
+        ----------
+        z_path : np.ndarray
+            Zt 경로 (길이 = horizon)
+        horizon : int
+            예측 기간 (기본: z_path 길이)
+        
+        Returns
+        -------
+        dict with keys:
+            - "cumulative_pd": shape (horizon, N-1) — 등급별 누적 PD
+            - "marginal_pd": shape (horizon, N-1) — 등급별 한계 PD
+            - "survival_prob": shape (horizon, N-1) — 등급별 생존확률
+            - "conditional_tms": list of 전이행렬 (디버깅용)
+        """
+        if horizon is None:
+            horizon = len(z_path)
+        
+        cumulative_tm = np.eye(self.n_grades)
+        cumulative_pds = []
+        conditional_tms = []
+        
+        for t in range(horizon):
+            z_t = z_path[t] if t < len(z_path) else 0.0
+            
+            # 조건부 전이행렬 산출
+            cond_tm = conditional_transition_matrix(self.ttc_matrix, z_t, self.rho)
+            conditional_tms.append(cond_tm)
+            
+            # 누적 전이행렬
+            cumulative_tm = cumulative_tm @ cond_tm
+            
+            # 누적 PD = D열 (마지막 열)
+            cum_pd = cumulative_tm[:-1, -1].copy()
+            cumulative_pds.append(cum_pd)
+        
+        cumulative_pds = np.array(cumulative_pds)  # shape: (horizon, N-1)
+        
+        # 한계 PD
+        marginal_pds = np.zeros_like(cumulative_pds)
+        marginal_pds[0] = cumulative_pds[0]
+        for t in range(1, horizon):
+            marginal_pds[t] = np.maximum(cumulative_pds[t] - cumulative_pds[t - 1], 0.0)
+        
+        # 생존확률
+        survival_probs = 1.0 - cumulative_pds
+        
+        return {
+            "cumulative_pd": cumulative_pds,
+            "marginal_pd": marginal_pds,
+            "survival_prob": survival_probs,
+            "conditional_tms": conditional_tms,
+        }
+    
+    def compute_all_scenarios(
+        self,
+        z_paths: Dict[str, np.ndarray],
+        scenario_weights: Dict[str, float],
+        horizon: Optional[int] = None
+    ) -> Dict[str, any]:
+        """
+        전체 시나리오 Lifetime PD 산출 + 가중평균
+        
+        Parameters
+        ----------
+        z_paths : Dict[str, np.ndarray]
+            시나리오별 Zt 경로
+        scenario_weights : Dict[str, float]
+            시나리오별 확률가중치
+        horizon : int
+            예측 기간
+        
+        Returns
+        -------
+        dict with keys:
+            - "by_scenario": {scenario: {cumulative_pd, marginal_pd, ...}}
+            - "weighted_cumulative_pd": shape (horizon, N-1)
+            - "weighted_marginal_pd": shape (horizon, N-1)
+        """
+        results = {"by_scenario": {}}
+        
+        weighted_cum = None
+        weighted_marginal = None
+        
+        for scenario_name, z_path in z_paths.items():
+            logger.info(f"시나리오 '{scenario_name}' PD 산출 중...")
+            
+            result = self.compute_lifetime_pd(z_path, horizon)
+            results["by_scenario"][scenario_name] = result
+            
+            weight = scenario_weights.get(scenario_name, 1.0 / len(z_paths))
+            
+            if weighted_cum is None:
+                weighted_cum = weight * result["cumulative_pd"]
+                weighted_marginal = weight * result["marginal_pd"]
+            else:
+                weighted_cum += weight * result["cumulative_pd"]
+                weighted_marginal += weight * result["marginal_pd"]
+        
+        results["weighted_cumulative_pd"] = weighted_cum
+        results["weighted_marginal_pd"] = weighted_marginal
+        results["weighted_survival_prob"] = 1.0 - weighted_cum
+        
+        return results
+    
+    def format_pd_table(
+        self,
+        results: Dict,
+        years: List[int] = None,
+        scenario: str = None
+    ) -> pd.DataFrame:
+        """
+        PD 결과를 DataFrame 테이블로 포매팅
+        
+        Parameters
+        ----------
+        results : dict
+            compute_all_scenarios() 결과
+        years : List[int]
+            표시할 연도 목록 (기본: 1,2,3,5,7,10,15,20,30,50)
+        scenario : str
+            특정 시나리오 (None이면 가중평균)
+        
+        Returns
+        -------
+        pd.DataFrame
+            index=연도, columns=등급
+        """
+        if years is None:
+            years = [1, 2, 3, 5, 7, 10, 15, 20, 30, 50]
+        
+        if scenario is not None:
+            cum_pd = results["by_scenario"][scenario]["cumulative_pd"]
+        else:
+            cum_pd = results["weighted_cumulative_pd"]
+        
+        # 호라이즌 범위 내 연도만 선택
+        max_t = cum_pd.shape[0]
+        valid_years = [y for y in years if y <= max_t]
+        
+        data = {}
+        for y in valid_years:
+            data[y] = cum_pd[y - 1]  # 0-indexed
+        
+        df = pd.DataFrame(data, index=self.non_default_grades).T
+        df.index.name = "년"
+        
+        return df
+    
+    def format_marginal_pd_table(
+        self,
+        results: Dict,
+        years: List[int] = None,
+        scenario: str = None
+    ) -> pd.DataFrame:
+        """한계 PD를 DataFrame으로 포매팅"""
+        if years is None:
+            years = [1, 2, 3, 5, 7, 10, 15, 20, 30, 50]
+        
+        if scenario is not None:
+            m_pd = results["by_scenario"][scenario]["marginal_pd"]
+        else:
+            m_pd = results["weighted_marginal_pd"]
+        
+        max_t = m_pd.shape[0]
+        valid_years = [y for y in years if y <= max_t]
+        
+        data = {}
+        for y in valid_years:
+            data[y] = m_pd[y - 1]
+        
+        df = pd.DataFrame(data, index=self.non_default_grades).T
+        df.index.name = "년"
+        
+        return df
+
+
+def compute_ecl_weights(
+    marginal_pds: np.ndarray,
+    lgd: float = 0.45,
+    discount_rate: float = 0.03,
+    horizon: int = None
+) -> np.ndarray:
+    """
+    ECL (Expected Credit Loss) 계산 보조 함수
+    
+    ECL = Σ_t [PD_marginal(t) × LGD × DF(t)]
+    
+    Parameters
+    ----------
+    marginal_pds : np.ndarray
+        한계 PD 배열 (등급별)
+    lgd : float
+        부도시 손실률 (LGD), 기본 45% (Basel IRB)
+    discount_rate : float
+        할인율, 기본 3%
+    
+    Returns
+    -------
+    np.ndarray : 등급별 누적 ECL
+    """
+    if horizon is None:
+        horizon = marginal_pds.shape[0]
+    
+    ecl = np.zeros(marginal_pds.shape[1] if marginal_pds.ndim > 1 else 1)
+    
+    for t in range(horizon):
+        df = 1.0 / (1.0 + discount_rate) ** (t + 1)
+        if marginal_pds.ndim > 1:
+            ecl += marginal_pds[t] * lgd * df
+        else:
+            ecl += marginal_pds[t] * lgd * df
+    
+    return ecl