veikkasj/bert-finnish-online-financial-sentiment

Sentiment classification model used in my 2026 Master's thesis "Finnish online investor sentiment and stock returns" for Aalto University.

Link to thesis: https://aaltodoc.aalto.fi/items/cb84fbc3-ff4a-4d46-b60d-70136716b038

See files for data source, scraping scripts, pseudo-labeling script, and model training & evaluation notebook.

Some eval metrics:

── Mean ± Std ──

• —Accuracy : 0.633 ± 0.029
• —F1 weighted : 0.632 ± 0.030
• —F1 macro : 0.628 ± 0.030
• —MAEM : 0.446 ± 0.042

── Mean Per-class Metrics (across folds) ──

• —negative: precision=0.561 recall=0.627 f1=0.587
• —neutral: precision=0.654 recall=0.609 f1=0.626
• —positive: precision=0.682 recall=0.668 f1=0.672