FinesseBERT

FinesseBERT is a fine-tuned sequence classification model based on DistilBERT, built to predict the sentiment of stock market and crypto news articles from the headline and metadata alone — no full article body required. Built by SentientMerchant, a platform exploring AI-driven tools for financial intelligence.

Purpose

Given the high volume and velocity of financial news, FinesseBERT enables fast, scalable sentiment analysis at the point of discovery — making it well-suited for real-time trading signals, news aggregators, and financial dashboards.

The model classifies inputs into three sentiment categories:

• Positive-Outlook-On-Stock-News
• Neutral-Outlook-On-Stock-News
• Negative-Outlook-On-Stock-News

Attribution Requirement

This model is licensed under CC-BY. If you use this model in your research, application, or product, you must provide attribution by linking back to sentientmerchant.com.

Training Data

FinesseBERT was fine-tuned on a dataset of 2,900 labeled stock news examples. Each example was structured as a single concatenated string capturing four fields sourced from financial news articles:

{
    "reference": "https://finance.yahoo.com/...",
    "text": "security: LULULEMON ATHLETICA INC (LULU); title: Lululemon (LULU) Dips More Than Broader Market: What You Should Know; description: In the closing of the recent trading day, Lululemon (LULU) stood at $138.16, denoting a -2.97% move from the preceding trading day.; author: yahoo.com",
    "labels": 2
}

Field	Description
security	The full company name and ticker symbol
title	The article headline
description	A short article summary or lede sentence
author	The publishing source

The labels field maps to: 0 → Positive, 1 → Neutral, 2 → Negative.

Optimal Inference Format

To get the best results from FinesseBERT, structure your input text to mirror the training data format exactly — a semicolon-delimited string with the four named fields in the same order:

security: <COMPANY NAME> (<TICKER>); title: <HEADLINE>; description: <DESCRIPTION>; author: <SOURCE>

Example:

security: LULULEMON ATHLETICA INC (LULU); title: Lululemon (LULU) Dips More Than Broader Market: What You Should Know; description: In the closing of the recent trading day, Lululemon (LULU) stood at $138.16, denoting a -2.97% move from the preceding trading day.; author: yahoo.com

Deviating from this format — such as passing a raw headline string alone — may degrade classification accuracy, as the model learned the sentiment signal from the full structured context it was trained on.

How to Use

You can load and use this model directly from the Hugging Face Hub using the transformers library.

1. Install Dependencies

Make sure you have the required libraries installed:

pip install transformers torch

2. Loading the Model

Use the Auto classes to load the model and tokenizer directly from the Hugging Face Hub.

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "MattELab/FinesseBERT"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

3. Running Inference

Here is a quick example of how to pass text through the model to get predictions.

# 1. Define your input text (see Optimal Inference Format above)
text = "security: LULULEMON ATHLETICA INC (LULU); title: Lululemon (LULU) Dips More Than Broader Market: What You Should Know; description: In the closing of the recent trading day, Lululemon (LULU) stood at $138.16, denoting a -2.97% move from the preceding trading day.; author: yahoo.com"

# 2. Tokenize the input
# Note: DistilBERT has a maximum input length of 512 tokens. Inputs longer than
# this will be silently truncated, which may degrade prediction quality.
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512, padding=True)

# 3. Run the model (using torch.no_grad() for faster, memory-efficient inference)
with torch.no_grad():
    outputs = model(**inputs)

# 4. Convert logits to probabilities
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)

# 5. Get the predicted class
predicted_class = torch.argmax(probabilities, dim=-1).item()

# 6. Map the predicted class ID to a human-readable label
label_map = {0: "Positive-Outlook-On-Stock-News", 1: "Neutral-Outlook-On-Stock-News", 2: "Negative-Outlook-On-Stock-News"}

print(f"Probabilities: {probabilities}")
print(f"Predicted Class ID: {predicted_class}")
print(f"Predicted Sentiment: {label_map[predicted_class]}")

Model Details

• Architecture: DistilBERT (AutoModelForSequenceClassification)
• Task: Text Classification
• Classes: 3 (Positive-Outlook-On-Stock-News, Neutral-Outlook-On-Stock-News, Negative-Outlook-On-Stock-News)
• Creator: MattELab

About SentientMerchant

SentientMerchant provides real-time stock, crypto, and international market data to keep you up-to-date. Find top news headlines, individual and overall news sentiment across various timelines, build a watchlist, buy US & SG stocks, and create and manage your portfolio.