Key Insights: Built a country-specific forecasting model for call center demand, improving performance by up to 50% and achieving an RMSE under 0.72 for session rate forecasts.

About the Client

PedidosYa, part of Delivery Hero, is the leading food delivery platform in Latin America. Operating in multiple countries and expanding through acquisitions like Glovo, the company employs over 5,000 people and has a market valuation exceeding $3 billion.

The Challenge

PedidosYa needed to staff its call centers efficiently to handle fluctuating inbound call volumes, avoiding both overstaffing and understaffing.

Call rates depended on numerous operational variables that had to be analyzed to identify which factors had the greatest impact, and how they correlated. Accurate demand forecasting was essential to predict customer needs based on historical interaction data, enabling timely, efficient service while optimizing labor schedules.

Marvik’s Approach

To identify the best forecasting model, we benchmarked a variety of approaches:

• Traditional ML models including ARIMA, decision trees, and gradient boosting.
  
  
• Deep neural networks based on recurrent neural network architectures.
  
  

Our feature engineering process created new variables to capture specific user behaviors. The final solution was an assembled model fine-tuned for each country, capable of forecasting demand for six-hour shifts with high accuracy.

Built with Python, Google Dataproc, Google BigQuery, and MongoDB, the model achieved an RMSE of under 0.72 for session rate forecasts.

The Results & Impact

• Forecast accuracy improved by up to 50% compared to manual approaches.
  
  
• Reliable staffing schedules reduced operational inefficiencies.
  
  
• Country-specific tuning ensured performance across multiple markets.
  

Accurate forecasting allows firms to set realistic targets, allocate resources efficiently, and reduce forecasting errors, critical advantages in high-volume customer service operations.

Why This Matters

By combining advanced ML techniques with tailored feature engineering, this project gave PedidosYa a predictive edge, ensuring the right number of agents are ready at the right time, across every market they serve.