Dynamic Sales Forecasting with Multiple-Multivariate Time Series

This Python, ML-based API with an impressive Normalised Mean Squared Error of 0.039 and Normalised Mean Absolute Error of 0.10 for forecasts of this multiple, multivariate timeseries data on a 15-day window. This POC aims to empowers retailers with data-driven sales forecasting to optimize inventory management and improve profitability. This comprehensive solution predicts short-term sales (up to 30 days) with outstanding accuracy and performance metrics. By integrating seamlessly with existing workflows, it enables retailers to:

• Make informed inventory decisions.
• Reduce stockouts and overstocking.
• Improve operational efficiency.
• Boost customer satisfaction.

Screenshots and Demo

Model Performance Metrics (15-day forecast)

MSE (on normalised data)	MAE (on normalised data)
0.03967	0.10960

The Approach

• Fetching of the data from the MongoDB database using Python's PyMongo library.
• Performing in-depth analysis of data in a Jupyter Environment.
• Performing Feature Engineering to prepare the data for modelling.
• Forecasting oil prices using a LightGBM Model for preparing the covariates for the actual model.
• Fitting another LightGBM Model to forecast sales with the target values and covariates.
• Converting the Jupyter code to modular format for building code pipelines efficiently.
• Setting up the github repository, dockerhub repository using git and docker dekstop.
• Running DVC and Git on the built artifacts during the training process for file versioning and tracking.
• Creating the REST API using Python's FastAPI.
• Creating a Dockerfile for building the docker image for containerization of the api.
• Signing up and logging in on Render Cloud and creating a web service with the configuration of dockerized deployments.
• Setting up Github Actions and creating a workflow with the jobs of building the docker image, pushing the image to Dockerhub, and finally use Render Cloud's webhook to trigger deployment of the api once a push is made to dockerhub.

Tech Stack

Built on a robust open-source ecosystem, it leverages:

• Darts : For efficient time series operations and forecasting.
• MongoDB : For storage and retrieval of data.
• LightGBM : To accurately predict covariate and target features.
• Scikit-learn : For creating data pipelines.
• DVC, Git, and Github : For seamless data and code versioning.
• Evidently AI: To check for data drift/target drift.
• FastAPI : For building a user-friendly API for model accessibility.
• Docker and Dockerhub : For secure and streamlined deployment to the Render Cloud Platform.
• Github Actions : For automating the CI/CD pipeline.
• Render Cloud : A PaaS for deployment of web apps, api's, etc.

Run Locally

Clone the project

  git clone https://github.com/rauhanahmed/store-sales-forecasting

Go to the project directory

  cd store-sales-forecasting

Install dependencies

  pip install -r requirements.txt

Start the server

  uvicorn app:app --reload

Run the Train Pipeline

For running the train pipeline, follow the below steps:

• Create a MongoDB Atlas account, or sign in if already exists.
• Create an environment file, named secrets.env in the project's main directory with the keys of environment variables as MONGODB_USERNAME and MONGODB_PASSWORD and fill their respective values in URL encoded format
• Install the project's requirements by using the command:

pip install -r requirements.txt

• run the train_pipeline by using the command:

!python /src/pipelines/train_pipeline.py

Deployment

For deployment of the application via dockers, follow the below steps:

• Create a docker image of the application using the below command, replacing <IMAGE_NAME> and after installing docker desktop and signing up on DockerHub.

docker build -t <IMAGE_NAME>:<TAG> .

• Test the container locally

docker run -p 8000:8000 <IMAGE_NAME>:<TAG>

• Push to DockerHub replacing <DOCKERHUB_USERNAME>, <IMAGE_NAME> and the with the dockerhub username, image name and the image tag respectively

docker push <DOCKERHUB_USERNAME>/<IMAGE_NAME>:<TAG>

• Sign up on Render Cloud
• Create a Web Service on Render and choose to deploy via dockers
• Specify the url of the dockerhub repository and start build
• Use the URL provided by render after successful deployment of application after few minutes :)

Connect with me

Support

For support, email [email protected]. I would be happy to help!

License

MIT