Abstract
Time-series forecasting supports decisions in finance, energy, transportation, public health, and industrial monitoring. Recent foundation models improve transfer across forecasting tasks, but many depend on centralized data and Transformer attention, which restricts their use for long, high-dimensional, and privacy-sensitive signals. This paper presents QuantFlow, a probabilistic forecasting framework that combines inverted sequence embedding, bidirectional Mamba state-space decoders, quantile regression, and federated learning.
Each variable is embedded over the complete observation window, processed in forward and reverse directions, and projected to five conditional quantiles. TSMixup expands temporal diversity through Dirichlet-weighted interpolation while preserving sequence structure. Experiments cover cryptocurrency, traffic, electricity, Electricity Transformer Temperature, influenza, and weather data. QuantFlow obtains mean squared errors of 0.2834 on ETTm1 and 0.2218 on Weather, and a 20-client non-IID deployment retains useful accuracy after three communication rounds without centralizing raw records.
The results indicate that selective state-space modeling is a promising basis for scalable, uncertainty-aware, and privacy-conscious time-series prediction, while also revealing limitations on irregular epidemiological signals and long-horizon generalization.
Blogger's Review: The innovation of QuantFlow lies in its integration of federated learning and state-space models, enabling efficient time-series forecasting while safeguarding data privacy. However, further optimization is needed when dealing with long horizons and complex signals.