Introduction

This repository contains code for the Walmart Sales Forecasting project. The project aims to forecast weekly sales for 45 Walmart stores located in different regions. The data includes historical sales data, holiday events, and store information. Project contains framework to run and test multiple models in a spark environment. It also contains code to build spark docker image and run the spark container locally if needed.

Installation

1. Clone the repository:

   git clone https://github.com/selewaut/forecast_forge.git
   cd forecast_forge

2. Create and activate a virtual environment:

   python3 -m venv .venv
   source .venv/bin/activate

3. Install the required dependencies:

   pip install -r requirements.txt

4. Install forecast_forge package

   pip install -e .

5. Install openjdk on local machine

   sudo apt-get update
   sudo apt-get install openjdk-8-jdk

Usage

Data Preparation

The data is stored in the data/ directory. The data is stored in the following files:

• train.csv: historical sales data for 45 Walmart stores
• test.csv: test data for forecasting
• features.csv: additional data related to the stores and regional activity
• stores.csv: store information

Data is originally downloaded using the following kaggle competition: https://www.kaggle.com/c/walmart-recruiting-store-sales-forecasting/data.

For downloading data from kaggle, you need to have a kaggle account and kaggle API key. You can download the data using the following command:

1. Install kaggle package

   pip install kaggle

2. Generate API key from kaggle account and save it in ~/.kaggle/kaggle.json

Running the Code

1. Move to directory containgin spark dockerfile.

   cd spark-setup

2. Start container.

   make run

   This will start a spark container with the code mounted in the container. The container will be running in the background.

3. Run univariate_weekly.py script to train and test univariate weekly sales forecasting models.

   spark-submit --master local[*] src/forecast_forge/univariate_weekly.py

4. Results are saved in parquet format in evaluation_output path.