🏪 Predict RTO — X5 Tech Hackathon

A solution for forecasting RTO (Retail Turnover) of X5 store locations across two hackathon stages.

---

📋 Problem Statement

The goal was to predict the next month's RTO for each store using historical sales data, store characteristics, and external factors (foot/car traffic, demographics, nearby competitors).

Target metric: MAPE (Mean Absolute Percentage Error)

---

🗂 Repository Structure

├── first stage/          # Stage 1
│   ├── train_boost.py    # CatBoost training
│   ├── grid.py           # Hyperparameter search (GridSearch)
│   ├── create_test.py    # Test set construction with lag features
│   ├── predict.py        # Inference and export
│   └── data_normalize.csv
│
├── second_stage/         # Stage 2
│   ├── edit_data.py      # Feature engineering
│   ├── create_predict.py # March 2025 inference pipeline
│   ├── utils.py          # Shared utilities
│   ├── cat_boost/        # CatBoostRegressor (final model)
│   ├── gradboost/        # GradientBoostingRegressor
│   └── RandomForest/     # Baseline — RandomForestRegressor
│
└── requirements.txt

---

⚙️ Approach

Stage 1

• Target RTO log-transformed via log1p to reduce skewness
• Built lag features (lag1–lag6) per store
• Added rolling mean (mean3) and trend (lag1 − lag3)
• Seasonality encoded with sin/cos of month index
• Categorical features (region, city, store area, opening date) passed natively to CatBoost without encoding
• Hyperparameters tuned with ParameterGrid

Stage 2

• Extended lag set: lag1, lag2, lag3, lag6, lag12
• Added mean3, mean6, EMA3, std3, trend, trend6
• Log-normalized numerical features: population, households, RTO
• Feature selection via CatBoost SHAP (RecursiveByShapValues), tested with 14, 18, 22 and full feature sets
• Trained three models for comparison: CatBoost, GradientBoosting, RandomForest (baseline)
• Separate inference script create_predict.py for March 2025 with exact lag reproducibility

---

📊 Results

Stage 1

Model	MAPE
CatBoostRegressor	14%

Stage 2

Model	MAPE	R²
RandomForest (baseline)	13.47%
CatBoost (final)	5.24%

---

🛠 Stack

Python · CatBoost · scikit-learn · pandas · numpy · matplotlib