Skip to content

Demand Forecasting for PepsiCo

Project Title: Demand forecasting for PepsiCo
Client: PepsiCo (via Deloitte)
Segment: North American Beverages
Timeline: July 2024 – July 2025
Technologies: Python, PySpark, Databricks, Azure, MLflow, Random Forest, Hyperopt, SQL

Overview

Developed and deployed a scalable machine learning platform to forecast sales volume across customer, location, and segment dimensions for over 300 locations in North America. The solution integrated advanced forecasting models with robust MLOps practices to enhance planning accuracy and operational efficiency.

Key Contributions

  • Built over 10 thousand granular forecasting models achieving 90%+ accuracy across key segments.
  • Identified the best models by multimodel hyperparameter tuning using HyperOpt, Cross validation and Grid search algorithms.
  • Designed and implemented end-to-end MLOps pipelines in Azure, reducing model turnaround time from 15 hours to under 30 minutes.
  • Developed automated pre-checks and post-checks framework for data and model validation, ensuring high data integrity and model reliability.
  • Implementing CI/CD and Continuous Training using MLFlow for experiment tracking and model monitoring, enabling continuous performance evaluation and retraining.
  • Enabled real-time insights via a Power BI dashboard and a web-based tool, used by 600+ users with 3,000+ submissions per period.

Impact

  • Improved forecast accuracy and planning agility across sales, labor, and cost dimensions.
  • Reduced manual reporting time from 2 weeks to 30 minutes.
  • Enhanced collaboration across finance, merchandising, and operations teams.
Back to top