Temporal reasoning in financial texts is essential for understanding event timing and claim validity, especially in earnings conference calls and social media discussions. While transformer-based models have advanced natural language processing, the comparative performance of fine-tuned encoder models and prompt-based decoder models in multilingual temporal classification remains underexplored. This study systematically compares model types, model sizes, and prompting strategies across two tasks: detecting temporal references in English texts and assessing claim validity in Chinese posts. Encoder models such as RoBERTa and BERT and decoder models such as GPT-4o, Mistral, and Gemma are evaluated using fine-tuning and few-shot prompting approaches. Results show that fine-tuned encoder models achieve consistently strong performance across both English and Chinese datasets. Mid-sized prompt-based decoder models also perform competitively under well-designed prompts, offering a practical alternative when fine-tuning is not feasible. In addition, decoder models are more robust to class imbalance, as reflected by smaller gaps between Micro-F1 and Macro-F1 scores. However, decoder models perform less effectively on Chinese tasks, indicating the need for language-specific adaptation. These findings provide practical guidance for selecting models and designing prompts for financial natural language processing under resource constraints.