This research project considers the problem of predicting throughput for link adaptation and resource scheduling in a multi-carrier, multi-technology cellular mobile network while preserving user privacy. It aims to develop advanced machine learning models that can accurately predict throughput in real-time, leveraging the diverse fine-grained data available from operational 4G and 5G networks that offer LTE and 5G NR connectivity across multiple operators. The project improves state-of-the-art research in this domain by investigating federated throughput prediction across various user devices, operating on different cellular technologies leveraging data from a large-scale measurement campaign across two countries. The approach focuses on developing a time series Federated Learning (FL) model that predicts network throughput at a future point in time across various mobile network operators and edge devices. Model weights are shared across devices to ensure user data remains localized, enhancing privacy and security. By utilizing FL, the aim is to collaboratively train models across different scenarios, user equipment, and network operators without sharing raw data, addressing the challenges of data heterogeneity, resource constraints, and communication overhead.