Synchrony Datathon 2026 — Team Naveed

Problem

Predict intraday call center volume (CV, CCT, ABD) at 30-minute intervals for 4 Synchrony portfolios across a full month.

Setup

All development done inside the provided Docker environment (Classical ML + Boosting image).

pip install -r requirements.txt  # all packages pre-installed in Docker

How to Run

# Full pipeline (preprocess → features → train → predict)
python run_pipeline.py --portfolio all --output outputs/forecast_v01.csv

# Individual steps
python src/preprocess.py
python src/features.py
python src/train.py --portfolio portfolio_1
python src/predict.py --output outputs/forecast_v01.csv

Approach

• Prophet model per portfolio for baseline
• XGBoost with engineered time/lag features for iteration
• Asymmetric scoring addressed by slight upward bias on CV
• Experiment tracking via MLFlow

Results

Validation Performance

• Portfolio 3: Composite Score = 0.1725 (Best performing)
• Portfolio 4: Composite Score = 0.1742
• Portfolio 2: Composite Score = 0.1750
• Portfolio 1: Composite Score = 0.1761

Individual Metrics (MAPE)

• Call Volume (CV): ~12% MAPE across portfolios
• Customer Care Time (CCT): ~11% MAPE across portfolios
• Abandon Rate (ABD): ~41% MAPE across portfolios

Model Performance

• Prophet baseline models trained for each portfolio and metric
• All predictions clipped to non-negative values
• Forecast generated for August 2023 (31 days × 48 intervals × 4 portfolios)