FedChronos: Federated Fine-Tuning of Time-Series Foundation Models for Privacy-Preserving Commodity Price Forecasting

Research Context and Motivation

Time-series foundation models (TSFMs) like Chronos have shown impressive zero-shot forecasting capabilities. However, deploying these models in real-world scenarios—such as agricultural price forecasting across regional markets—is hindered by data fragmentation. Regulatory requirements, competitive sensitivities, and data sovereignty laws prevent the centralization of sensitive time-series data. FedChronos addresses this by enabling collaborative model fine-tuning across distributed clients without requiring raw data to leave its source.

The FedChronos Approach

FedChronos utilizes Low-Rank Adaptation (LoRA) to perform parameter-efficient fine-tuning on the Chronos-T5 backbone. By training only a small subset of parameters (approximately 1.16% of the total model), the framework reduces the data transmitted per communication round by 86% compared to full-model exchange. This makes the approach highly suitable for edge AI deployments where network bandwidth and client-side compute resources are constrained. The framework supports standard federated aggregation strategies like FedAvg and FedProx to update the global model based on local adapter weights.

Main Findings and Privacy

A critical challenge identified in the study is that naive LoRA fine-tuning on small, non-IID (non-independent and identically distributed) datasets leads to significant overfitting, causing the model to perform worse than the zero-shot baseline. The authors discover that integrating client-level differential privacy (DP) serves as an effective form of implicit regularization. By injecting noise into the adapter updates, the model avoids overfitting and achieves a 31% improvement in Mean Absolute Percentage Error (MAPE) over the zero-shot baseline. This demonstrates that privacy-preserving mechanisms can enhance, rather than hinder, model accuracy in federated TSFM applications.