[QLRRTH]_Chuyên viên Kiểm định mô hình rủi ro tín dụng

## Ôn Thi & Chuẩn Bị Cho Vị Trí Kiểm Định Mô Hình RRTD ### 1. Lộ Trình Chuẩn Bị 2 Tuần ``` 📅 TUẦN 1: Củng cố Kiến thức Nền Ngày 1-2: Ôn Basel III ├── Basel I → II → III evolution ├── Capital Adequacy Ratio calculation ├── Pillars: Capital, Supervisory Review, Market Discipline ├── Tier 1 Capital vs Tier 2 Capital └── Capital buffer types (Conservation, Countercyclical) Ngày 3-4: Ôn IFRS 9 ├── 3-stage credit risk model │ ├── Stage 1: 12-month ECL (no significant increase) │ ├── Stage 2: Lifetime ECL (significant increase) │ └── Stage 3: Credit-impaired assets ├── PD term structure ├── Forward-looking information └── Đọc IFRS 9 technical summary Ngày 5-6: Ôn Mô hình RRTD ├── PD models: Scorecard, logistic regression, ML models ├── LGD models: Hurdle approach, beta regression ├── EAD models: CCF (Credit Conversion Factor) ├── EWS (Early Warning System) basics └── Correlation: Asset correlation, portfolio models Ngày 7: Ôn Validation Metrics ├── Discrimination: KS, AUC, Gini, Accuracy ratio ├── Calibration: Brier score, Hosmer-Lemeshow ├── Stability: PSI, CSI (Characteristic Stability Index) └── Business metrics: AR, Bad rate trending 📅 TUẦN 2: Luyện Kỹ Năng Thực Hành Ngày 8-9: Luyện SQL ├── Window functions (LAG, LEAD, running totals) ├── CASE WHEN cho variable creation ├── GROUP BY + HAVING cho aggregations ├── Self-join cho cohort analysis └── Practice: Viết query tính PD từ bảng transactions Ngày 10-11: Luyện Python/R ├── Pandas: data cleaning, grouping, merging ├── Scikit-learn: logistic regression, model evaluation ├── Statsmodels: hypothesis testing ├── Matplotlib/Seaborn: visualization └── Practice: Tính AUC, KS, PSI từ sample data Ngày 12-13: Mock Interview ├── Tự hỏi theo danh sách câu hỏi phỏng vấn ├── Record lại để tự đánh giá ├── Rèn giọng nói, ngôn ngữ body └── Timing: trả lời mỗi câu 2-3 phút Ngày 14: Tổng kết + Chuẩn bị hồ sơ ├── Review lại tất cả concepts ├── Chuẩn bị portfolio/cases đã làm ├── In ấn hồ sơ, giấy tờ └── Nghỉ ngơi, ngủ đủ giấc ``` ### 2. Tài Liệu Tham Khảo Bắt Buộc ``` 📚 SÁCH: 1. "An Introduction to Credit Risk Modeling" - Lyn C. Thomas → Nền tảng về PD, LGD, EAD 2. "Credit Risk Scorecards: Developing and Implementing Intelligent Credit Scoring" - Stephan Dirickx → Scorecard development methodology 3. "Basel IV: The Reason, The Rationale, The Results" - Tim Skeet → Cập nhật Basel mới nhất 4. "IFRS 9 Impairment: A Practitioner's Guide" - DV VMFA → IFRS 9 implementation thực tế 📄 TÀI LIỆU FREE: 1. Basel Committee Publications: → BCBS 239 (data governance) → BCBS 368 (validation principles) → BCBS 412 (stress testing) 2. VCB Annual Report: → Đọc phần Quản lý rủi ro → Hiểu chiến lược VCB 3. FRM Study Materials (GARP): → Part II: Credit Risk → Miễn phí preview một số chương 4. Khan Academy - Statistics: → Probability, hypothesis testing 💻 WEBSITE: 1. Credit Risk Modeling (github repos) 2. Towards Data Science (Medium) 3. Risk.net (industry news) 4. GARP (FRM certification) ``` ### 3. Các Khái Niệm Cần Nắm Vững ``` 🔑 CREDIT RISK MODELING: 1. PD (Probability of Default) ├── Point-in-time (PIT) vs Through-the-cycle (TTC) ├── Marginal PD vs Cumulative PD ├── Migration matrix (Markov chain) └── Rating scale design 2. LGD (Loss Given Default) ├── Direct vs Indirect LGD ├── Workout LGD vs Market LGD ├── Recovery rate modeling └── LGD confidence intervals 3. EAD (Exposure at Default) ├── CCF (Credit Conversion Factor) ├── On-balance sheet vs Off-balance sheet ├── Credit limit utilization └── Correlation with default 4. ECL (Expected Credit Loss) ├── 12-month ECL vs Lifetime ECL ├── Staging criteria ├── Forward-looking scenarios └── PD × LGD × EAD formula 🔑 VALIDATION METRICS: 1. Discrimination ├── KS (Kolmogorov-Smirnov) ├── AUC (Area Under ROC) ├── Gini coefficient └── Odds ratio 2. Calibration ├── Brier score ├── Hosmer-Lemeshow test ├── Reliability diagram └── Binomial test 3. Stability ├── PSI (Population Stability Index) ├── CSI (Characteristic Stability Index) └── Divergence analysis 🔑 REGULATORY: 1. Basel III/IV ├── CAR (Capital Adequacy Ratio) ≥ 8% ├── Tier 1 Ratio ≥ 6% ├── CET1 ≥ 4.5% ├── Capital conservation buffer 2.5% └── Liquidity ratios (LCR, NSFR) 2. IFRS 9 ├── 3-stage impairment model ├── Significant increase in credit risk (SICR) ├── Lifetime PD term structure └── Forward-looking information 3. Regulatory expectations ├── Model inventory ├── Model validation policy ├── Model risk governance └── Backtesting requirements ``` ### 4. Code Mẫu Ôn Tập ```python # 📌 Python: Tính PSI (Population Stability Index) import numpy as np import pandas as pd def calculate_psi(expected, actual, buckets=10): """ Calculate PSI for model validation PSI < 0.1: No significant shift PSI 0.1-0.25: Moderate shift, monitor PSI > 0.25: Significant shift, action required """ # Create bins based on expected distribution breakpoints = np.percentile(expected, np.linspace(0, 100, buckets + 1)) # Calculate actual and expected percentages in each bin expected_counts = np.histogram(expected, breakpoints)[0] actual_counts = np.histogram(actual, breakpoints)[0] # Avoid division by zero expected_pct = expected_counts / len(expected) actual_pct = actual_counts / len(actual) # Replace zeros to avoid log(0) expected_pct = np.where(expected_pct == 0, 0.0001, expected_pct) actual_pct = np.where(actual_pct == 0, 0.0001, actual_pct) # Calculate PSI psi = np.sum((actual_pct - expected_pct) * np.log(actual_pct / expected_pct)) return psi # Example usage expected_scores = np.random.normal(0.5, 0.2, 10000) actual_scores = np.random.normal(0.52, 0.22, 10000) psi_value = calculate_psi(expected_scores, actual_scores) print(f"PSI = {psi_value:.4f}") if psi_value < 0.1: print("✓ No significant population shift") elif psi_value < 0.25: print("⚠ Moderate shift - monitor closely") else: print("✗ Significant shift - action required") ``` ```sql -- 📌 SQL: Tính PD từ bảng account WITH default_flag AS ( SELECT account_id, date, CASE WHEN days_past_due >= 90 THEN 1 ELSE 0 END AS is_default, ROW_NUMBER() OVER (PARTITION BY account_id ORDER BY date) AS rn FROM account_transactions ), cohort_analysis AS ( SELECT DATE_TRUNC('month', date) AS month, COUNT(*) AS total_accounts, SUM(is_default) AS default_count, SUM(is_default) * 1.0 / COUNT(*) AS observed_pd FROM default_flag WHERE rn = 1 -- First observation per account GROUP BY DATE_TRUNC('month', date) ) SELECT month, total_accounts, default_count, observed_pd, AVG(observed_pd) OVER ( ORDER BY month ROWS BETWEEN 11 PRECEDING AND CURRENT ROW ) AS trailing_12m_pd FROM cohort_analysis ORDER BY month; ```