FinBERT-FOMC Aspect Sentiment v5

This is a single aspect-aware sentiment model fine-tuned from ZiweiChen/FinBERT-FOMC on FOMC_sentences_expanded_zeroshot_labeled.xlsx.

Model Details

• Base model: ZiweiChen/FinBERT-FOMC
• Training mode: all
• Train rows: 28377
• Eval rows: 3123

Labels

• Negative (0)
• Neutral (1)
• Positive (2)

Inference Contract

Use one model for all targets:

• text = mode
• text_pair = FOMC sentence

Valid mode values:

• growth
• employment
• inflation

Python Usage

from transformers import BertForSequenceClassification, BertTokenizer, pipeline

model_id = "Wonseong/FinBERT-FOMC-aspects"
tokenizer = BertTokenizer.from_pretrained(model_id)
model = BertForSequenceClassification.from_pretrained(model_id)
clf = pipeline("text-classification", model=model, tokenizer=tokenizer)

sentence = "Spending on cars and light trucks increased somewhat in July."
mode = "growth"  # growth | employment | inflation

result = clf({"text": mode, "text_pair": sentence})
print(result)

Local Helper Usage

from predict_aspect import FinBERTFOMCModeClassifier

clf = FinBERTFOMCModeClassifier("Wonseong/FinBERT-FOMC-aspects")
print(clf.predict("Spending on cars and light trucks increased somewhat in July.", mode="growth"))

Metrics

Holdout evaluation:

• Accuracy: 0.8789625360230547
• Loss: 0.5731372237205505
• Macro F1: 0.7980614874405157
• Macro Precision: 0.7708534213383098
• Macro Recall: 0.8321941498302832
• MAE: 0.1399295549151457
• MSE: 0.1777137367915466
• MAPE: 8.09051126054008

Full training data fit:

• Accuracy: 0.9348888888888889
• Macro F1: 0.8977282349978499
• Macro Precision: 0.8621019267617939
• Macro Recall: 0.9424642336852936
• MAE: 0.06847619047619048
• MSE: 0.07520634920634921
• MAPE: 3.6275132275132274

Training Data Contract

Expected spreadsheet columns in sheet expanded_10500:

• sentence
• growth_sentiment_final
• employment_sentiment_final
• inflation_sentiment_final

Label values are normalized from negative, neutral, and positive.

Dataset Provenance

The training workbook FOMC_sentences_expanded_zeroshot_labeled.xlsx is publicly available in this repository. The expanded_10500 sheet contains 10,500 rows with growth, employment, and inflation labels, yielding 31,500 aspect-sentence examples before the training/validation split.

The workbook records mixed label provenance:

• 1,405 rows are marked is_human_labeled = 1, source_sheet = human label, and selection_note = existing_human_label.
• 9,095 rows are marked is_human_labeled = 0.
• The label-completion script preserves existing labels and fills blank or no majority found cells with zero-shot annotations from oss-20b, using zeroshot_fomc_sentiment_prompt.txt. Its checkpoint covers 9,150 rows, including 55 rows carrying human-label provenance that required at least one missing label to be completed.

The provenance flag identifies retained human annotations; it does not mean that the zero-shot completions were subsequently human-audited. The training supervision therefore combines retained annotations with model-generated completions. Reported holdout metrics measure agreement with these mixed-provenance labels and are not an independent evaluation against a fully human-annotated benchmark.

The additional sheets provide length distributions, label counts, and construction metadata. The is_oversampled_duplicate field identifies rows introduced by sampling with replacement.

The zero-shot prompt asks for one label per aspect:

• growth_sentiment_final
• employment_sentiment_final
• inflation_sentiment_final

Each aspect is restricted to:

• positive
• neutral
• negative

The prompt design is derived from the human-labeling framework described in:

Kim, W., Sporer, J., Lee, C. L., & Handschuh, S. (2024, November). Is small really beautiful for central Bank communication? Evaluating language models for finance: Llama-3-70B, GPT-4, FinBERT-FOMC, FinBERT, and VADER. In Proceedings of the 5th ACM International Conference on AI in Finance (pp. 626-633).

Intended Use

This model is intended for aspect-conditioned sentiment classification on FOMC-related sentences. The caller supplies the target aspect through mode and the sentence through text_pair.

Limitations

• This is not a general finance sentiment model for arbitrary entities or tasks.
• Reported metrics are from a holdout split of the labeled spreadsheet, not from external benchmark evaluation.
• The model predicts one of three ordinal sentiment classes only: Negative, Neutral, Positive.

Training Summary

• Runtime device: Tesla T4
• CUDA available: true
• BF16: true
• FP16: false

Additional artifacts in this repo:

• training_metadata.json
• metric_history.csv
• mae_mse_curve.png

Citation

@article{kim2023analyzing,
  title={Analyzing FOMC minutes: Accuracy and constraints of language models},
  author={Kim, Wonseong and Sp{\~A}{\c{k}}rer, Jan Frederic and Handschuh, Siegfried},
  journal={arXiv preprint arXiv:2304.10164},
  year={2023}
}

@inproceedings{kim2024small,
  author = {Kim, W. and Sporer, J. and Lee, C. L. and Handschuh, S.},
  title = {Is small really beautiful for central Bank communication? Evaluating language models for finance: Llama-3-70B, GPT-4, FinBERT-FOMC, FinBERT, and VADER},
  booktitle = {Proceedings of the 5th ACM International Conference on AI in Finance},
  pages = {626--633},
  year = {2024}
}