Paper: arXiv 2401.11011

Authors: Valentina Aparicio, Daniel Gordon, Sebastian G. Huayamares, Yuhuai Luo

Abstract

Large language models (LLMs) are deep learning algorithms being used to perform natural language processing tasks in various fields, from social sciences to finance and biomedical sciences. Developing and training a new LLM can be very computationally expensive, so it is becoming a common practice to take existing LLMs and finetune them with carefully curated datasets for desired applications in different fields. Here, we present BioFinBERT, a finetuned LLM to perform financial sentiment analysis of public text associated with stocks of companies in the biotechnology sector. The stocks of biotech companies developing highly innovative and risky therapeutic drugs tend to respond very positively or negatively upon a successful or failed clinical readout or regulatory approval of their drug, respectively. These clinical or regulatory results are disclosed by the biotech companies via press releases, which are followed by a significant stock response in many cases. In our attempt to design a LLM capable of analyzing the sentiment of these press releases,we first finetuned BioBERT, a biomedical language representation model designed for biomedical text mining, using financial textual databases. Our finetuned model, termed BioFinBERT, was then used to perform financial sentiment analysis of various biotech-related press releases and financial text around inflection points that significantly affected the price of biotech stocks.

Complexity vs Empirical Score

  • Math Complexity: 4.0/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)

Why this score: The paper uses established deep learning techniques (fine-tuning BERT models) with minimal novel mathematical derivations, but demonstrates strong empirical rigor through detailed data collection, a concrete backtested trading strategy, and code availability.

Research Flowchart

  flowchart TD
  A["Research Goal:<br>Develop LLM for Sentiment Analysis<br>of Biotech Press Releases"] --> B["Input Data:<br>Financial Databases &<br>Biotech Press Releases"]
  B --> C["Methodology:<br>Finetune BioBERT using<br>Financial Text Databases"]
  C --> D["Computational Process:<br>Train BioFinBERT LLM<br>on Financial Sentiment"]
  D --> E["Outcome 1:<br>BioFinBERT Model Validated"]
  D --> F["Outcome 2:<br>Analysis of Sentiment Around<br>Biotech Inflection Points"]
  E --> G["Key Finding:<br>Effective LLM for Predicting<br>Stock Response to Clinical Readouts"]
  F --> G

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes BioFinBERT, a finetuned large language model for financial sentiment analysis of biotech press releases.

Source sentence
Here, we present BioFinBERT, a finetuned LLM to perform financial sentiment analysis of public text associated with stocks of companies in the biotechnology sector.

Assumes 2

  • Trading entries are triggered when the price difference between Open and Close on the release day exceeds 5%. execution
    Source sentence
    The benchmark trading strategy consists in entering a position based on a 5% threshold difference between the Open and Close price on the release day, or the next trading day if the press release wasn’t published on a trading day.
  • Positions are held for fixed periods of 1, 2, or 3 months after a positive event. execution
    Source sentence
    Specifically, our trading strategy implied holding biotech stocks following a positive event for 1, 2, or 3 months after the event’s announcement.

Data

  • financial_phrasebank · HuggingFace text-news public
    Source sentence
    For our first attempt at finetuning BioBERT, we used the financial_phrasebank dataset from HuggingFace, which contains ~4,000 samples of financial news sentiment classification already attributed with positive, negative, or neutral labels.
  • FiQA financial text sentiment dataset · Kaggle text-news public
    Source sentence
    we added the FiQA financial text sentiment dataset from Kaggle.
  • Biopharma‑Dive curated dataset · BioPharma‑Dive news publisher text-news public
    biotech companies
    Source sentence
    Our third attempt was to curate our own dataset to combine with financial_phrasebank to finetune BioBERT. This dataset was obtained from 96 news articles from the biotech section of BioPharma-Dive news publisher.
  • NASDAQ Biotechnology Index (NBI) · NASDAQ benchmark public
    2020-2023 · biotech stocks
    Source sentence
    Our market benchmark was the NASDAQ Biotechnology Index (NBI) and the back-testing period was between 2020 and 2023.

Recognised sources: Yahoo Finance FiQA Kaggle

Costs

Not modeled A trading or execution result is reported without costs.

Finds 3

  • The finetuned BioBERT (BioFinBERT) achieved an F1 score of 0.9104, validation loss of 0.7088, and success rate of 62.86%, comparable to FinBERT. predictive-accuracy
    Source sentence
    Ultimately, this finetuned BioBERT performed similarly to FinBERT for an F1 score of 0.9104, validation evaluation loss of 0.7088, and success rate of 62.86%.
  • For a 1‑month holding period, micro‑cap stocks following a positive event delivered the best performance among market‑cap categories. performance
    Source sentence
    For a 1-month holding period, holding micro-cap stocks following a positive event had the best performance, and holding l…
  • The BioPharma strategy yields the highest annualized Sharpe Ratio for Large‑Cap companies, while FinPhrase is best for Mid‑Cap and the benchmark for Small‑ and Micro‑Cap. performance
    Source sentence
    BioPharma strategy works best for Large-Cap companies • FinPhrase strategy works best for Mid-Cap companies •

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 12 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors): A Survey of Large Language Models for Financial … (uses its method or data)

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.