Sental: High-Performance Financial Sentiment Analysis

Sental is a state-of-the-art financial sentiment classifier fine-tuned from yiyanghkust/finbert-tone on corporate filings, financial news, analyst notes, and market tweets.

Model Performance

Evaluated on the official validation dataset (2,388 unseen samples):

Metric	Score
Accuracy	88.9% (Ensemble) / 88.2% (Standalone FinBERT)
Macro F1	0.85
Weighted F1	0.89

Per-Class Metrics

Class	Precision	Recall	F1-Score	Support
Bearish (2)	0.87	0.73	0.79	347
Bullish (1)	0.88	0.78	0.83	475
Neutral (0)	0.89	0.96	0.93	1,566

Label Mapping

The model outputs logits for 3 classes:

• 0: Neutral
• 1: Bullish
• 2: Bearish

Quickstart

Using transformers pipeline

from transformers import pipeline

classifier = pipeline("text-classification", model="muhalwan/sental")
result = classifier("Apple reports record quarterly profit, crushes Wall Street expectations!")
print(result)
# [{'label': 'Bullish', 'score': 0.98}]

Using PyTorch directly

import torch
from transformers import BertTokenizer, BertForSequenceClassification

tokenizer = BertTokenizer.from_pretrained("muhalwan/sental")
model = BertForSequenceClassification.from_pretrained("muhalwan/sental")

text = "Tesla plunges 12% as vehicle margins shrink and deliveries miss forecast."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)

predicted_idx = outputs.logits.argmax(dim=-1).item()
label = model.config.id2label[predicted_idx]
print(f"Sentiment: {label} (Confidence: {probs[0][predicted_idx]:.2%})")
# Sentiment: Bearish (Confidence: 96.4%)

Training Methodology

• Base Architecture: BERT-base uncased pre-trained on financial corpora (yiyanghkust/finbert-tone).
• Precision: Mixed Precision (fp16) with PyTorch AMP.
• Optimization: AdamW with linear warmup, cosine learning rate decay, and gradient clipping (1.0).
• Preprocessing: Semantics-preserving emoji demojization, handle/URL normalization, and dynamic token padding.