nishantnayar/earningsbert-semiconductor

EarningsBERT — Semiconductor Earnings Sentiment Classifier

Model ID: nishantnayar/earningsbert-semiconductor Base model: roberta-base Task: Text classification (6-class sentiment) Domain: Semiconductor sector earnings press releases

---

Model Description

EarningsBERT is a fine-tuned RoBERTa-base model for classifying sentences from semiconductor earnings press releases into a 6-class taxonomy. Unlike general-purpose finance sentiment models (positive/negative/neutral), EarningsBERT includes two guidance-specific classes — guidance_raise and guidance_lower — which are the strongest predictors of post-earnings stock direction.

Label Mapping

ID	Label	Economic signal
0	optimistic	Bullish tone
1	cautious	Bearish lean
2	hedging	Uncertainty / vagueness
3	guidance_raise	Strong bullish — explicit guidance increase
4	guidance_lower	Strong bearish — explicit guidance decrease
5	neutral	No signal (boilerplate, logistics)

---

Intended Use

• Sentence-level sentiment classification for semiconductor earnings press releases
• Feature generation for post-earnings stock direction models
• Research into language patterns in corporate communications

Not intended for: general financial text, non-semiconductor sectors, investment decisions.

---

Training Data

Source	Size	Labels
Financial PhraseBank (Malo et al. 2014)	2,264 sentences	positive → optimistic, negative → cautious, neutral → neutral
Silver-labeled semiconductor transcripts	~3,000 sentences	All 6 classes via rule-based labeler
Total	~5,000 sentences	6 classes, balanced

Silver labeling rules use Loughran-McDonald lexicon density, modal verb density, approximation density, and semiconductor-specific phrase matching ("guidance raised", "inventory digestion", etc.).

---

Performance

Evaluated on a held-out split (20% stratified by class):

Metric	Value
Macro F1	0.986
Accuracy	0.987
guidance_raise precision	>0.95
guidance_lower precision	>0.95

Training stopped at epoch 2 (early stopping on macro-F1) to avoid overfitting.

---

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="nishantnayar/earningsbert-semiconductor",
    return_all_scores=True,
)

result = classifier("We are raising our Q3 gross margin guidance to 57%, up from our prior range of 54-55%.")
# [{'label': 'guidance_raise', 'score': 0.94}, ...]

Or using the project wrapper:

from src.earningsbert.model_wrapper import EarningsBERT

model = EarningsBERT("nishantnayar/earningsbert-semiconductor")
scores = model.predict(["Record data center revenue driven by AI demand."])
# [{'optimistic': 0.89, 'cautious': 0.02, ...}]

---

Limitations

• Trained exclusively on semiconductor sector language. Performance on other sectors is untested.
• EDGAR 8-K press releases only — not trained on live earnings call Q&A transcripts.
• Silver labels carry ~15-20% noise. guidance_raise/guidance_lower rules are highest precision; hedging is noisiest.
• Base model cutoff: roberta-base (2019). Post-2023 terminology (CoWoS, HBM3e) may be underrepresented.

---

Citation

If you use this model, please cite:

@misc{earningsbert2024,
  title  = {EarningsBERT: Semiconductor Earnings Sentiment Classifier},
  author = {[Your Name]},
  year   = {2024},
  url    = {https://huggingface.co/nishantnayar/earningsbert-semiconductor}
}

---

How to Push

# 1. Set hub_model_id in config.yaml:
#    earningsbert:
#      hub_model_id: "your-username/earningsbert-semiconductor"

# 2. Login
huggingface-cli login

# 3. Push (after training)
python -m src.cli.main finetune --push-to-hub