wtf is future-sales-forecasting-using-machine-learning?

This project predicts future daily sales for Rossmann, a European drugstore chain, using its historical sales records and store information. It walks through the whole process a data scientist would follow: cleaning the raw data, building useful features from it, trying out several prediction models, tuning the best one, and finally packaging it into a small web app anyone can use without writing code. The feature building step turns dates and past sales into signals a model can learn from, such as how long a competing store has been open nearby, how long a promotion has been running, the previous day's sales, seven and thirty day rolling averages, and how the same weekday or month performed a year earlier. The project first builds a simple baseline forecast to check whether the more complex models are actually worth using, then trains and compares Linear Regression, XGBoost, LightGBM, and CatBoost, before tuning the best performing one, which turned out to be XGBoost. Instead of splitting the data randomly, the last 70 days are held back for testing, which better mimics how the model would be used in practice, forecasting into days it has not seen yet. Accuracy is measured with several standard error metrics. The final trained model is deployed inside a Streamlit dashboard, a simple web app framework for Python, where a user can upload their own historical sales data, enter details about a store, generate forecasts for future dates, view how accurate the model has been, and download the results. There is a live demo link hosted on Streamlit's own cloud service, plus a Kaggle notebook showing the full training process. To run it on your own machine, you install the requirements with pip and start the dashboard with a single Streamlit command. The README notes plainly that the accuracy shown reflects past performance on this specific dataset and should not be taken as a guarantee for other businesses.