Telemarketing remains a critical customer outreach channel in the financial industry, yet connection rates typically range between 12% and 18% across large-scale financial telemarketing operations. Traditional rule-based dialing strategies that call all users during fixed “golden hours” fail to account for individual timing preferences, leading to substantial operational inefficiency. We propose T2D-CNet (Temporal-aware Decoupling and Data-aware Debiasing Coordination Network), a production-ready deep learning system for personalized financial call-time optimization at scale, deployed at a major financial institution processing over millions of calls daily. We formalize the problem as a discrete-choice prediction task and identify three core modeling challenges: (1) extreme data sparsity under strict calling limits, (2) severe temporal imbalance in call distribution, and (3) heterogeneous user sensitivity to timing. T2D-CNet addresses these through a coordinated architecture that integrates heterogeneous features, decouples global and personalized timing patterns, corrects distributional bias, and captures user-specific sensitivity. We conduct extensive experiments on two large-scale financial datasets and demonstrate consistent superiority over strong baselines. Rigorous online A/B testing with million users shows a statistically significant 5.85% improvement in connection rate (p-value < 0.001). We distill key modeling insights regarding feature separation for sparse data, data-aware debiasing, and coordinated learning for sensitivity modeling.