Assume the role of a Lead NLP Data Scientist. We have a fine-tuned BERT model currently deployed for sentiment analysis on general news text. However, it performs poorly on financial headlines containing sarcasm or idiomatic market slang. Provide a detailed technical roadmap to improve the model's F1-score on this specific subset. Include recommendations for: 1) Data augmentation techniques specific to financial text, 2) Advanced tokenization strategies, 3) Loss functions that handle class imbalance (positive vs. negative vs. neutral), and 4) Evaluation metrics beyond simple accuracy. Write the pseudo-code for a custom training loop that implements focal loss.