QuantFlow: Federated Mamba Model for Privacy-Preserving Time-Series Forecasting.

This paper introduces QuantFlow, an innovative probabilistic forecasting framework designed for time-series data. It integrates several advanced techniques: inverted sequence embedding, bidirectional Mamba state-space decoders, quantile regression, and federated learning. This combination allows for robust predictions while addressing critical concerns like data privacy and scalability, especially for long, high-dimensional, and sensitive signals. QuantFlow processes each variable across its entire observation window, using both forward and reverse directions, and projects the output to five conditional quantiles. A technique called TSMixup enhances temporal diversity. Experimental results across diverse datasets, including cryptocurrency, traffic, and electricity, show strong performance, with the federated learning aspect maintaining accuracy even in non-IID deployments without centralizing raw data. This highlights the potential of selective state-space modeling for privacy-preserving and uncertainty-aware time-series prediction.