federated probabilistic forecasting

To address weak privacy guarantees, severe overfitting on small samples, poor convergence under multi-party settings, and high hyperparameter-tuning complexity in vertical federated learning (VFL) for industrial time-series forecasting, this paper proposes STV—the first secret-sharing-based VFL framework for time-series prediction. STV introduces an *N*-party secure matrix multiplication and matrix inversion protocol, enabling direct parameter optimization while ensuring strong convergence and low tuning complexity. It integrates SARIMAX, autoregressive trees, and custom multi-party secure computation protocols—eliminating reliance on a trusted third party. Evaluated on six real-world datasets, STV matches centralized methods in accuracy and outperforms state-of-the-art diffusion models and LSTMs by 23.81%. Communication overhead analysis further provides theoretical foundations for practical deployment.

This work addresses the challenge of time-series forecasting in federated learning settings, where heterogeneity in temporal granularity and variable sets across nodes hinders effective collaboration. To tackle this issue, the authors propose PiXTime, a novel framework that employs personalized patch embedding to unify multi-granularity time-series representations and introduces a global Variable Embedding (VE) table to align semantic meanings of variables across nodes. Building upon this unified representation, PiXTime integrates a shared Transformer architecture with cross-attention mechanisms to enable efficient and collaborative modeling. Evaluated under realistic federated conditions, PiXTime achieves state-of-the-art performance on eight real-world time-series benchmarks, demonstrating its effectiveness in overcoming the challenges posed by heterogeneity in federated time-series forecasting.

This paper addresses the binary classification of financial market price trend volatility under privacy-preserving cross-institutional collaborative modeling. To tackle real-world challenges—including non-IID financial data and heterogeneous institutional requirements—we propose a differentially private LSTM framework integrated with personalized federated learning. Each participant trains an LSTM model locally; secure model aggregation is achieved via noisy gradient updates and a personalized adaptation mechanism. Theoretical analysis and empirical evaluation demonstrate that our approach achieves accuracy comparable to centralized training while strictly ensuring data locality—outperforming single-institution baselines by a significant margin. Our key contributions are threefold: (1) the first systematic integration of personalized federated learning and differential privacy into financial volatility forecasting; (2) rigorous validation of its feasibility, robustness against data heterogeneity, and collective performance gain; and (3) provision of a practical, regulation-compliant technical pathway for efficient, privacy-aware cross-institutional intelligent risk control.

Adaptive selection of the look-back horizon for time-series forecasting in federated learning remains challenging due to data decentralization, non-IID distributions, and client heterogeneity. Method: We propose the first intrinsic representation space adaptation framework tailored for federated settings. Our approach models temporal structural heterogeneity via a synthetic data generator, integrates intrinsic-space mapping, Bayesian error decomposition, and geometric statistical modeling, and achieves optimal trade-offs among prediction errors under decentralized constraints. Contribution/Results: We theoretically prove that the optimal look-back horizon corresponds to the minimal point where the irreducible loss saturates. Empirically, our method significantly improves both forecasting accuracy and communication efficiency across diverse heterogeneous time-series tasks, outperforming existing baselines in federated forecasting benchmarks.

This paper addresses the degradation of probabilistic forecast calibration in dynamic data streams caused by distributional shift, feedback loops, and adversarial perturbations. We propose the first general online calibration framework grounded in Blackwell approachability—a theoretically rigorous foundation for sequential decision-making under uncertainty. Our method provides strong calibration guarantees in compact output spaces (e.g., classification and bounded regression) and enables lossless post-hoc recalibration of arbitrary pre-trained predictors. Technically, it unifies insights from Blackwell approachability theory, online optimization, and gradient-based updates, and introduces task-specific efficient algorithms for both classification and regression. Empirical evaluation demonstrates substantial improvements in calibration quality for energy system forecasting, with marked gains in robustness and practical utility for downstream decision-making tasks.

This work addresses the challenge of forecasting in agricultural markets where data cannot be centralized due to regulatory or sovereignty constraints and exhibit heterogeneous distributions. The authors propose FedChronos, a framework that performs federated parameter-efficient fine-tuning on the pre-trained time series foundation model Chronos-T5. By integrating low-rank adaptation (LoRA) with differential privacy, the method transmits only approximately 384 KB of adapter parameters per communication round, marking the first integration of federated learning with efficient fine-tuning of time series foundation models. Experimental results on commodity price prediction across 15 Indian markets show that the injected differential privacy noise acts as implicit regularization, mitigating overfitting in low-data regimes. Under the optimal configuration (ε=5), FedChronos reduces MAPE by 31% compared to zero-shot inference and by 26% against conventional baselines, while providing (ε,δ)-differential privacy guarantees.

This work addresses key challenges in time series forecasting—namely, reliance on centralized data, computational inefficiency of Transformers in long-horizon and high-dimensional settings, and modeling under privacy-sensitive conditions—by proposing QuantFlow, a probabilistic forecasting framework that integrates inverted sequence embedding, a bidirectional Mamba state space decoder, conditional quantile regression, and federated learning. To the best of our knowledge, this is the first approach to incorporate Mamba into federated time series modeling, further enhanced by TSMixup to augment temporal diversity. The resulting method enables efficient, privacy-preserving, and uncertainty-aware multivariate forecasting. QuantFlow achieves MSE scores of 0.2834 and 0.2218 on the ETTm1 and Weather datasets, respectively, and maintains high prediction accuracy in a non-IID federated setting with 20 clients using only three communication rounds.

This work addresses the degradation in accuracy and calibration commonly observed in federated Bayesian neural networks due to the difficulty of specifying appropriate priors and likelihoods. To overcome this challenge without sharing local data, the authors propose a novel federated predictive Bayesian approach that integrates martingale posteriors, trainable data embeddings, and a federated learning framework. They introduce the first single-round, parallelized federated Markov posterior sampling protocol, wherein clients upload learned embeddings and the server performs centralized predictive sampling to recover parameter uncertainty—thereby preserving privacy and eliminating reliance on explicit prior distributions. Experimental results demonstrate that the method achieves performance nearly on par with centralized training on MNIST, CIFAR-10, and CIFAR-100, while significantly outperforming existing consensus-based baselines in terms of calibration.