QuantFlow: Federated Mamba Model Enables Privacy-First Time-Series Forecasting
Post-transformer approach with state-space decoders and federated learning matches centralized models.
The paper presents QuantFlow, a probabilistic forecasting framework that moves beyond Transformer-based architectures by leveraging bidirectional Mamba state-space decoders. Each variable is embedded over its full observation window, processed forward and reverse, then projected to five conditional quantiles for uncertainty estimation. TSMixup augments temporal diversity via Dirichlet-weighted interpolation without breaking sequence structure. This design allows long-horizon, high-dimensional modeling while keeping data decentralized—a key requirement for privacy-sensitive domains like healthcare and finance.
Experiments across cryptocurrency, traffic, electricity, influenza, and weather datasets show QuantFlow matches or beats Transformer baselines: MSE of 0.2834 on ETTm1 and 0.2218 on Weather. In a 20-client non-IID federated setting, it retains useful accuracy after only three rounds of communication, with no need to centralize raw records. Limitations appear on irregular epidemiological signals and very long horizons, but the results establish selective state-space modeling as a strong alternative for scalable, uncertainty-aware, and privacy-conscious time-series prediction.
- Combines bidirectional Mamba decoders, quantile regression, and federated learning for privacy-preserving forecasting.
- Achieves MSE 0.2834 on ETTm1 and 0.2218 on Weather, competing with centralized Transformer models.
- 20-client non-IID federated deployment holds accuracy after 3 communication rounds, avoiding raw data centralization.
Why It Matters
Enables scalable, privacy-compliant time-series predictions in finance, energy, and healthcare without compromising accuracy.