feat(data): add CCC interpolator module (7x7 -> 8x8 expansion)

This commit is contained in:
Variet Agent
2026-03-11 16:07:30 +09:00
parent 2b94cc802d
commit a406d98226

197
data/ccc_interpolator.py Normal file
View File

@@ -0,0 +1,197 @@
# -*- coding: utf-8 -*-
"""
CCC interpolation module: 7x7 -> 8x8
B and D rows/columns are used to create a synthetic CCC grade
via geometric mean (log-interpolation) of transition probabilities.
This module runs AFTER Zt estimation (which uses 7x7 matrices)
to produce the final 8x8 matrices for Lifetime PD projection.
Usage:
from data.ccc_interpolator import expand_to_8x8
tm_8x8 = expand_to_8x8(tm_7x7)
"""
import numpy as np
from typing import Optional
# 7x7 index: AAA=0, AA=1, A=2, BBB=3, BB=4, B=5, D=6
# 8x8 index: AAA=0, AA=1, A=2, BBB=3, BB=4, B=5, CCC=6, D=7
GRADES_7 = ["AAA", "AA", "A", "BBB", "BB", "B", "D"]
GRADES_8 = ["AAA", "AA", "A", "BBB", "BB", "B", "CCC", "D"]
def expand_to_8x8(
tm_7x7: np.ndarray,
alpha: float = 0.5,
method: str = "geometric"
) -> np.ndarray:
"""
7x7 transition matrix -> 8x8 with CCC interpolated between B and D.
The CCC row is interpolated from B row and D row.
The CCC column is created by splitting the D column for grades above CCC.
Parameters
----------
tm_7x7 : np.ndarray
7x7 (AAA, AA, A, BBB, BB, B, D) probability matrix
alpha : float
Interpolation weight (0.5 = geometric midpoint between B and D)
method : str
'geometric': log-interpolation (default)
'linear': linear interpolation
Returns
-------
np.ndarray
8x8 (AAA, AA, A, BBB, BB, B, CCC, D) probability matrix
"""
assert tm_7x7.shape == (7, 7), f"Expected (7,7), got {tm_7x7.shape}"
tm_8x8 = np.zeros((8, 8))
# --- Step 1: Copy existing grades (AAA~B) rows/cols ---
# 7x7 index mapping: 0-5 -> 0-5 (AAA~B), 6 -> 7 (D)
for i in range(6): # AAA~B rows
for j in range(6): # AAA~B cols
tm_8x8[i, j] = tm_7x7[i, j]
# D col: 7x7 col6 -> 8x8 col7
tm_8x8[i, 7] = tm_7x7[i, 6]
# --- Step 2: CCC column (col6) for existing grades ---
# For each grade AAA~B, split some probability from D column to CCC
# Rationale: some firms default through CCC before reaching D
for i in range(6):
pd_i = tm_7x7[i, 6] # P(i -> D) in 7x7
if pd_i > 0:
# B row: larger CCC fraction (B is adjacent to CCC)
# Higher grades: smaller CCC fraction
grade_distance_from_b = max(5 - i, 0)
# B->CCC gets ~30%, BB->CCC ~20%, BBB->CCC ~10%, A->CCC ~5%
ccc_fraction = max(0.30 - grade_distance_from_b * 0.06, 0.02)
ccc_prob = pd_i * ccc_fraction
tm_8x8[i, 6] = ccc_prob # to CCC
tm_8x8[i, 7] = pd_i - ccc_prob # remaining to D
else:
tm_8x8[i, 6] = 0.0
# --- Step 3: CCC row (row 6) via interpolation ---
b_row = np.zeros(8)
d_row = np.zeros(8)
# Expand B row (7x7 row5) to 8x8 space
for j in range(6):
b_row[j] = tm_7x7[5, j]
b_row[6] = 0.0 # placeholder for CCC
b_row[7] = tm_7x7[5, 6]
# D row in 8x8: absorbing state
d_row[7] = 1.0
if method == "geometric":
# Geometric interpolation in log space
ccc_row = _geometric_interp(b_row, d_row, alpha)
else:
# Linear interpolation
ccc_row = alpha * b_row + (1 - alpha) * d_row
# Ensure CCC PD is between B PD and 1.0
# CCC should default more than B
ccc_pd = max(ccc_row[7], b_row[7] * 1.5)
ccc_pd = min(ccc_pd, 0.60) # cap at 60%
# CCC stay rate
ccc_stay = max(1.0 - ccc_pd - ccc_row[:6].sum() - ccc_row[6], 0.30)
# Reassemble CCC row
# Upgrade probabilities from B row, scaled down
for j in range(5): # AAA~BB: very small upgrade from CCC
ccc_row[j] = b_row[j] * 0.3 # CCC upgrades less than B
ccc_row[5] = b_row[5] * 0.5 # CCC -> B (upgrade)
ccc_row[6] = ccc_stay # CCC -> CCC (stay)
ccc_row[7] = ccc_pd # CCC -> D
tm_8x8[6, :] = ccc_row
# --- Step 4: D row (absorbing state) ---
tm_8x8[7, :] = 0.0
tm_8x8[7, 7] = 1.0
# --- Step 5: Normalize rows ---
for i in range(8):
s = tm_8x8[i].sum()
if s > 0:
tm_8x8[i] /= s
return tm_8x8
def _geometric_interp(
row_a: np.ndarray,
row_b: np.ndarray,
alpha: float = 0.5,
eps: float = 1e-10
) -> np.ndarray:
"""Geometric (log-space) interpolation between two probability rows."""
result = np.zeros_like(row_a)
for j in range(len(row_a)):
a = max(row_a[j], eps)
b = max(row_b[j], eps)
result[j] = np.exp(alpha * np.log(a) + (1 - alpha) * np.log(b))
return result
def expand_conditional_tm(
cond_7x7: np.ndarray,
ttc_8x8: np.ndarray = None
) -> np.ndarray:
"""
Expand a Z-conditional 7x7 TM to 8x8 using the same interpolation.
This is used in the lifetime PD projection pipeline:
1. Estimate Zt from 7x7 matrices
2. Generate Z-conditional 7x7 TM
3. Expand to 8x8 for lifetime PD calculation
Parameters
----------
cond_7x7 : np.ndarray
Z-conditional 7x7 transition matrix
ttc_8x8 : np.ndarray, optional
Reference TTC 8x8 for CCC structure (if available)
"""
return expand_to_8x8(cond_7x7)
if __name__ == "__main__":
import sys
sys.path.insert(0, ".")
from data.transition_matrices import load_transition_matrices, compute_ttc_matrix
matrices = load_transition_matrices(source="real")
ttc_7x7 = compute_ttc_matrix(matrices)
print("=== TTC 7x7 ===")
for i, g in enumerate(GRADES_7):
print(f" {g:>4}: [{', '.join(f'{v:.4f}' for v in ttc_7x7[i])}]")
ttc_8x8 = expand_to_8x8(ttc_7x7)
print("\n=== TTC 8x8 (CCC interpolated) ===")
for i, g in enumerate(GRADES_8):
print(f" {g:>4}: [{', '.join(f'{v:.4f}' for v in ttc_8x8[i])}]")
# Verify: PD ordering
print("\n=== PD ordering check ===")
for i, g in enumerate(GRADES_8[:-1]):
print(f" {g:>4}: PD = {ttc_8x8[i, -1]*10000:.1f}bp")
# Check row sums
print("\n=== Row sum check ===")
for i in range(8):
print(f" {GRADES_8[i]:>4}: sum = {ttc_8x8[i].sum():.6f}")