A Machine Learning-Based Weather Forecasting System with Live Data Anchoring and Uncertainty Estimation

This preprint presents a machine learning weather forecasting system that combines historical pattern learning with live data anchoring and uncertainty estimation. An XGBoost regression model is trained on lag-based temperature features and anchored to real-time observations from the Open-Meteo API. It produces recursive 7-day forecasts with residual-based 95% confidence intervals, so users can see how forecast reliability decreases with horizon. The system is served through a FastAPI backend and Streamlit frontend. Evaluated on Greater Noida, India data against a persistence baseline, the model reduced RMSE at every horizon tested (Day 1: 1.12 vs 2.45, Day 7: 3.15 vs 5.50). Unpublished manuscript; code and reproduction instructions are available at https://github.com/Yokiatch/weather-forecasting.