When prices climb and the cost of living feels like a slow squeeze, the way people talk about money changes. It’s not just about “what to buy” — it’s about how to cope. This study takes a closer look at that overlap between finance and self-care, digging into thousands of real posts from the FinTalk-19k dataset. Using a fine-tuned BERT-base model (optimized with Unsloth for speed and efficiency), I classified posts as self-care related or not, explored deeper themes with BERTopic, and checked how sentiment and uncertainty weave through different conversations.To keep things transparent, I used a train–validate–test split (60/20/20) with stratification, fixed seeds for reproducibility, confusion matrices to see exactly where the model messed up, and ROC curves to check performance trade-offs. SHAP explainability helped me peek inside the model’s “thought process” — why it made certain calls, and where it got confused.The results were strong on paper: ~93% accuracy, F1 scores around 0.90, and ROC-AUC near 0.96. But I also found possible data leakage (keywords and near-duplicate posts), meaning the model might look smarter than it really is. Crisis-related topics (scams, medical debt) had the lowest sentiment and the most uncertainty, while “good news” topics (lottery wins, tax refunds) were happier but still cautious.Looking ahead, I want to zoom in on just the “Personal Finance” posts and manually tag emotional nuances like sarcasm, masked anxiety, and quiet pride — the kind of things algorithms often miss.
Tanvi Dogra (Mon,) studied this question.