Capstone

Project Name:

• SuperPlan: Forecasting and Demand Planning

Team Members:

• Estrella Ndrianasy, Jesse Miller, Jyoti Kumari, Kanika Mahajan, Kevin Lu

Project Objectives

Our project leverages the M5 Competition’s Walmart sales dataset, which covers several stores in multiple states in the US. The M5 is one of the most well-known forecasting competitions in the field of demand planning. We explore top performing models from the M5, particularly LightGBM, and state-of-the-art deep learning models like Temporal Fusion Transformer against industry standards such as Triple Exponential Smoothing, in order to provide key insights to business users including:

* Which aggregation levels do the models forecast best?
* Are the models effective at capturing price/promo changes?
* And do the model perform well in common business scenarios not covered by the M5?

The full writeup for our project is available here

Implementation

EDA:

• Exploratory Data Analysis

Models:

1. Baseline Models

   • Exponential Smoothing
   • Weekly forecast - Exponential Smoothing

2. LightGBM

   • LightGBM
   • LightGBM - Weekly

3. TFT

   • TFT
   • TFT - Weekly

Evaluation Script:

• Evaluation Script

Website link:

• Berkeley Intranet project link

References:

• the following repository was leveraged heavily in the development of our LightGBM model: https://github.com/monsaraida/kaggle-m5-forecasting-accuracy-4th-place