d2c-mp-sales-forecaster 0.0.3

An application for forecasting direct-to-consumer sales quantities using the NeuralProphet algorithm from Meta.

D2C Marketplace Sales Forecasting Application

Welcome to the D2C Marketplace Sales Forecaster! This Streamlit application utilizes the Neural Prophet model to predict sales for various products. Designed with non-technical users in mind, this app guides you through the entire process of loading data, selecting products, tuning and training the model, generating forecasts, and distributing size forecasts.

Table of Contents

• Overview
• Getting Started
• Application Pages
  • 1. Load Data
  • 2. Product Selection
  • 3. Tune and Train
  • 4. Forecast Product
  • 5. Distribute Size Forecasts
• Contributing
• License

Overview

The D2C Marketplace Sales Forecaster app is designed to help you predict sales for products in a direct-to-consumer (D2C) marketplace. With an easy-to-use interface and step-by-step guidance, you'll be able to generate sales quantity forecasts for your products. The application is built using Streamlit and leverages the [NeuralProphet] (https://neuralprophet.com/) algorithm developed by Meta research to provide flexible forecasting options and interpretable results.

Developed as a capstone for the UCSD Professional Extension, Machine Learning Engineering certification.

1. What is the problem?

Forecast product-level sales by day for footwear and apparel digital marketplace sellers.

2. Why solve for it?

Online sellers face challenges in tracking inventory across multiple platforms, which can lead to overselling and increased inventory costs. They are limited in how many times per day each platform can be updated with available inventory balances. By more accurately predicting sales, sellers can allocate inventory more efficiently, maximize revenue, and minimize costs. As an additional benefit, the by-day forecasts can assist in more accurately planning warehouse labor and save logistics operational costs for the seller.

3. Unique problems for D2C Digital Sales Forecasting

E-commerce sales forecasting faces specific challenges that this application aims to address leveraging similar-product alignment and hierarchical planning methods.

-Strong seasonality and trend to the sales -Cold Start problem for new products that have no sales history -Insufficient data at the lowest level of granularity such as a EAN/UPC or size-level to be able to train machine learning models reliably

Running the Application

The application is built wtih Streamlit framework for operating as a browser-based web application and a PostgreSQL database for storing forecast results. Docker can be used for easy-deployment of the complete application and database, and a Dockerfile and Docker Compose compose.yaml are included in this repository and available to support deployment. The compose setup uses the official Postgres image for deployment of the PostgreSQL database. The application container is available in public-facing Dockerhub repository for pull. It is recommended to setup your own database secrets to connect Streamlit with Postgres and manage these using Docker secrets management.

It is not required to use Docker to use the application. However, it is highly recommended as it will make setup and deployment easier to manage. Please check Docker documentation. to download and install Docker and Docker Compose (https://www.docker.com/) Sign up for a Docker Hub account to be able to pull and use the containerized application and official Postgres Docker contianer.

Anaconda was used for package management and is part of the application Docker image. It is recommended to download and use Anaconda virtual environment and take advantage of the environment.yaml for setup. miniconda3 should be suffucient. (https://www.anaconda.com/download) Docker container will take care of this for you in the container build.

This application is required to run from unix/linus operating system. Application tested with the following:

NAME="Ubuntu" VERSION="20.04.3 LTS (Focal Fossa)" ID=ubuntu ID_LIKE=debian PRETTY_NAME="Ubuntu 20.04.3 LTS" VERSION_ID="20.04" HOME_URL="https://www.ubuntu.com/"

To run this application using Docker Compose:

1. Copy compose.yaml file from this github repository to desired deployment directory.

cp compose.yaml your_directory
cd your_directory

2. Create a hidden folder ".streamlit" inside of this directory, where you will create your own deployment secrets configurations for the database.

mkdir .streamlit

3. Make two secrets files inside the .streamlit directory: secrets.toml and postgres_secrets.env. You can copy paste the content listed below into each, replacing the configurations with your own. Please ensure that your password is consistent across both files. The user, POSTGRES_USER and the database should all share the same name consistently across files too!

# .streamlit/secrets.toml

[postgresql]
host="postgres"
port = xxxx
database = xxxx
user = xxxx
password = xxx

# .streamlit/postgres_secrets.env
POSTGRES_PASSWORD=xxxx
POSTGRES_USER=xxx

4. After configuring your application-database secrets, ensure you are in the same directory as compose.yaml and exeucte docker compose up. If you are running locally, once the application and database have been pulled form docker hub and started, you should be able to access the application from a web browser on localhost:8501.

docker compose up

Application Pages

The application consists of five main pages that guide you through the sales forecasting process.

1. Load Data

In this first step, you'll upload your sales data to the application. The data should be in CSV format with columns representing dates, product identifiers, and sales quantities.

2. Product Selection

After loading the data, you'll be presented with a list of products from the dataset. Choose the product(s) you want to forecast sales for by selecting them from the list.

3. Tune and Train

In this step, you'll have the option to fine-tune the Neural Prophet model's hyperparameters to improve the sales forecast accuracy. You can either use the default settings or adjust the parameters as needed. Once you're satisfied with the configuration, click the "Train Model" button to train the model using your selected product data.

4. Forecast Product

With the trained model, you can now generate sales forecasts for the selected products. Choose a forecast horizon (e.g., next 30 days) and click the "Generate Forecast" button. The application will display the forecasted sales along with confidence intervals for the selected period.

5. Distribute Size Forecasts

In the final step, you can allocate the forecasted sales quantities across various product sizes based on historical data or custom distribution rules. You can input the percentage distribution for each size and click "Distribute Sizes" to generate the final size-wise sales forecast.

Contributing

Contributions are welcome to improve the D2C Marketplace Sales Forecaster. If you have any suggestions, bug reports, or feature requests, please open an issue or submit