🇷🇺 Русская версия | 🇬🇧 English version

---

T-Bank Credit Attrition Prediction

University competition: Machine Learning Course (Central University)
Platform: Kaggle
Organizer: Central University & T-Bank
Timeline: November 2025

---

Challenge

Build a machine learning model to predict early loan repayment (attrition) using T-Bank credit product data. When customers repay loans early, banks earn less interest revenue — predicting this at the application stage is crucial for profitability optimization.

---

Data Overview

Key Characteristics

• Target variable: a6_flg (early repayment flag)
• Products: 4 credit products (product_1 — product_4)
• Time period: data split by months (month_dt)
• Features: ~100+ features (feature_0 — feature_N)

Main Challenge

Temporal distribution shift: test set significantly differs from training set in temporal distribution. Requires model stability monitoring across months and overfitting control.

---

Solution

Exploratory Data Analysis (EDA)

Missing values:

• Identified many features with high missing rates (50%+, 70%+)
• Analyzed feature importance for high-missing features using RandomForest
• Removed features with >70% missing values
• Applied median imputation (SimpleImputer) for remaining features

Class imbalance:

• Target variable shows imbalance (attrition is a rarer event)
• Used stratified validation to preserve class proportions

Modeling

Model choice: CatBoostClassifier

Rationale:

• Native handling of categorical features
• Resistant to overfitting (Ordered Boosting)
• High performance on tabular data
• Built-in missing value handling

Hyperparameters:

CatBoostClassifier(
    iterations=700,
    learning_rate=0.03,
    depth=6,
    eval_metric='AUC',
    random_state=42
)

Validation strategy:

• Train-test split with stratify=y (80/20)
• ROC-AUC monitoring on validation set
• Early stopping with use_best_model=True

---

Results

Metrics

• Baseline: 0.73707 ROC-AUC
• Kaggle leaderboard: 45th place, score: 0.75046

Key Findings

• Removing highly sparse features (>70% missing) improved model stability
• CatBoost demonstrated robustness to temporal shift through overfitting control
• Median imputation proved effective for numerical features with moderate missing rates

---

Tech Stack

ML Framework

• Data processing: pandas, numpy
• Visualization: matplotlib
• Modeling: CatBoost, RandomForestClassifier (for feature importance)
• Metrics: roc_auc_score

Environment

• Platform: Kaggle Notebooks (NVIDIA Tesla T4 GPU)
• Language: Python 3.11

---

About the Competition

Competition link: Kaggle — CU 2025 Scoring