musk1209/finsight-finbert

FinSight FinBERT — Financial Sentiment Classifier

A 3-class financial sentiment classifier (negative / neutral / positive) fine-tuned from bert-base-uncased on the Financial PhraseBank dataset.

Part of the FinSight project — a financial research AI assistant combining fine-tuned BERT, RAG, and multi-agent systems.

Performance (test split, sentences_50agree)

metric	value
loss	0.8388
accuracy	0.8433
f1_macro	0.8279
f1_weighted	0.8425
f1_negative	0.8308
f1_neutral	0.8789
f1_positive	0.7741

Training Setup

Setting	Value
Base model	bert-base-uncased
Dataset	atrost/financial_phrasebank (sentences_50agree, 64/16/20 split)
Train / Val / Test	3100 / 776 / 970
Epochs	3
Batch size	16
Learning rate	2e-5 (linear warmup over 10% of steps)
Weight decay	0.01
Max sequence length	128
Optimizer	AdamW (default)
Mixed precision	fp16
Seed	42
Hardware	NVIDIA Tesla T4 (Kaggle)

Label mapping

ID	Label
0	negative
1	neutral
2	positive

Usage

from transformers import pipeline

clf = pipeline("text-classification", model="musk1209/finsight-finbert", top_k=None)
clf("The company beat analyst expectations on revenue and profit.")
# [{'label': 'positive', 'score': 0.97}, ...]

Known limitations

• Strong on prototypical patterns ("losses widened", "record revenue") but can be fooled by inverted patterns sharing surface form with positive examples — e.g. "operating margin contracted from X% to Y%" may be mispredicted as positive because the training set's positive examples skew heavily toward margin-expansion sentences.
• Trained only on the sentences_50agree variant; not validated on the more ambiguous sentences excluded from that subset.

Citation

@article{Malo2014,

title = {Good debt or bad debt: Detecting semantic orientations in economic texts},

author = {Malo, P. and Sinha, A. and Korhonen, P. and Wallenius, J. and Takala, P.},

journal = {Journal of the Association for Information Science and Technology},

year = {2014}

}