Fine-grained emotion classification is a challenging task in natural language processing because emotional categories often overlap semantically and exhibit severe class imbalance. The difficulty becomes even greater in multi-label settings, where a single text instance may simultaneously express multiple emotional states. This study presents a comparative evaluation of optimisation strategies for BERT-based fine-grained emotion classification on the Go Emotions dataset. Five experimental settings are examined under a unified training protocol: Baseline BERT with Binary Cross Entropy, BERT with Weighted Binary Cross Entropy, BERT with Focal Loss, BERT with Focal Loss and threshold optimisation, and Baseline BERT with threshold optimisation. The objective is to determine whether performance gains are better obtained through modifying the loss function during training or calibrating decision thresholds after training. The findings show that threshold optimisation provides the strongest overall improvement, with BERT plus threshold optimisation achieving the best Micro-F1 score of 61.44%. While Weighted Binary Cross Entropy substantially improves recall, it also causes a pronounced drop in precision and accuracy, indicating over-prediction of positive labels. Focal Loss produces more balanced behaviour than Weighted Binary Cross Entropy but remains inferior to threshold-optimised BERT. The combination of Focal Loss and threshold optimisation improves recall further, yet it still does not outperform the simpler baseline model with optimised thresholds. These results indicate that decision-level calibration can be more effective than loss-function engineering for multi-label fine-grained emotion classification on imbalanced emotion corpora. The study therefore argues that threshold optimisation is a practical, low-complexity, and architecture-preserving strategy for improving BERT-based emotion classification systems.