This paper introduces Faithful TimeSieve (FTS), a framework designed to enhance the faithfulness and robustness of the TimeSieve model in time series forecasting. TimeSieve, while effective, suffers from instability due to sensitivity to random seeds, input perturbations, and parameter variations. To address these issues, the authors propose FTS, which defines three key attributes: similarity in IB space (Sib), consistency in prediction space (Cps), and stability in noise perturbations (Snp). These attributes ensure that the model's predictions remain consistent and reliable under various conditions. The FTS framework incorporates a minimax optimization problem that integrates these attributes, aiming to improve the model's performance and robustness. The framework is validated through extensive experiments on real-world datasets, demonstrating improved faithfulness and reduced sensitivity to perturbations. The results show that FTS outperforms the original TimeSieve model in terms of prediction accuracy and stability, particularly under input and parameter perturbations. Additionally, an ablation study of the loss functions confirms the effectiveness of combining all three loss components for enhancing model faithfulness. The study concludes that FTS provides a robust and reliable framework for time series forecasting, contributing to the development of more trustworthy and accurate forecasting methods.This paper introduces Faithful TimeSieve (FTS), a framework designed to enhance the faithfulness and robustness of the TimeSieve model in time series forecasting. TimeSieve, while effective, suffers from instability due to sensitivity to random seeds, input perturbations, and parameter variations. To address these issues, the authors propose FTS, which defines three key attributes: similarity in IB space (Sib), consistency in prediction space (Cps), and stability in noise perturbations (Snp). These attributes ensure that the model's predictions remain consistent and reliable under various conditions. The FTS framework incorporates a minimax optimization problem that integrates these attributes, aiming to improve the model's performance and robustness. The framework is validated through extensive experiments on real-world datasets, demonstrating improved faithfulness and reduced sensitivity to perturbations. The results show that FTS outperforms the original TimeSieve model in terms of prediction accuracy and stability, particularly under input and parameter perturbations. Additionally, an ablation study of the loss functions confirms the effectiveness of combining all three loss components for enhancing model faithfulness. The study concludes that FTS provides a robust and reliable framework for time series forecasting, contributing to the development of more trustworthy and accurate forecasting methods.