Paper: arXiv 2306.12659

Authors: Boyu Zhang, Hongyang Yang, Xiao-Yang Liu

Abstract

Sentiment analysis is a vital tool for uncovering insights from financial articles, news, and social media, shaping our understanding of market movements. Despite the impressive capabilities of large language models (LLMs) in financial natural language processing (NLP), they still struggle with accurately interpreting numerical values and grasping financial context, limiting their effectiveness in predicting financial sentiment. In this paper, we introduce a simple yet effective instruction tuning approach to address these issues. By transforming a small portion of supervised financial sentiment analysis data into instruction data and fine-tuning a general-purpose LLM with this method, we achieve remarkable advancements in financial sentiment analysis. In the experiment, our approach outperforms state-of-the-art supervised sentiment analysis models, as well as widely used LLMs like ChatGPT and LLaMAs, particularly in scenarios where numerical understanding and contextual comprehension are vital.

Complexity vs Empirical Score

  • Math Complexity: 1.0/10
  • Empirical Rigor: 8.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper focuses on practical instruction tuning of LLMs with minimal mathematical formalism, relying heavily on experimental benchmarking against SOTA models and cost-effectiveness metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Improve Financial Sentiment Analysis<br>for Numerical & Contextual Understanding"] --> B["Method: Instruction Tuning"]
  B --> C["Input: Supervised Financial Sentiment Data"]
  C --> D["Process: Transform Data into Instruction Format"]
  D --> E["Process: Fine-Tune General-Purpose LLM<br>e.g., LLaMA or GPT"]
  E --> F["Outcomes: Instruct-FinGPT Model"]
  F --> G["Key Findings: Outperforms SOTA & LLMs<br>Superior Numerical/Contextual Comprehension"]

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 Instruct-FinGPT, an instruction‑tuned LLaMA model for financial sentiment analysis that transforms classification data into a generation task.

Source sentence
We propose Instruct-FinGPT by instruction tuning [Wei et al., 2022] a pre-trained LLM (namely LLaMA [Touvron et al., 2023]).

Assumes 2

  • A small amount of fine‑tuning data is sufficient to achieve superior performance. data
    Source sentence
    With only a small amount of fine-tuning data, our model consistently achieves superior performance in sentiment analysis compared to FinBERT and ChatGPT.
  • Numerical values and financial context are essential for accurate sentiment prediction. model
    Source sentence
    These challenges underline the need for improved models that can adeptly understand the intricate nuances of financial sentiment analysis.

Data

  • Twitter Financial News dataset text-news public
    Source sentence
    Twitter financial news sentiment training: This dataset is a corpus of news tweets that pertain to the financial sector and is exclusively in English.
  • FiQA dataset · HuggingFace text-news public
    Source sentence
    FiQA dataset: This dataset, which is readily accessible via HuggingFace, includes 961 samples.
  • Twitter Val (validation) dataset benchmark public
    Source sentence
    Face, contains 2,390 samples annotated with three labels: Bearish, Bullish, or Neutral.
  • Numerical sensitivity dataset benchmark public
    Source sentence
    Numerical sensitivity dataset (numerical): This dataset, which we automatically filtered from Twitter Val, includes 117 samples.
  • Contextual understanding dataset benchmark public
    Source sentence
    Contextual understanding dataset (contextual): This dataset, which we randomly selected from Twitter Val, includes 20 samples.
  • Financial PhraseBank (FPB) dataset · LexisNexis benchmark public
    Source sentence
    comprises 4,840 samples randomly extracted from financial news articles available on the LexisNexis database.

Recognised sources: Twitter / X FiQA Financial PhraseBank

Costs

Not applicable The paper reports no trading or execution result.

Finds 1

  • Overall, instruction‑tuned LLaMA‑7B consistently outperforms both FinBERT and the original LLaMA‑7B across all evaluated datasets. performance
    Source sentence
    our instruction tuned LLaMA-7B (Instruct-FinGPT-7B) consistently outperforms both FinBERT and LLaMA-7B across all three datasets in terms of accuracy and F1 score.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 4 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 31 papers in the archive (24 by independent authors): Alpha-R1: Alpha Screening with LLM Reasoning via … (uses its method or data); MarS: a Financial Market Simulation Engine Powered by … (uses its method or data); A Survey of Large Language Models for Financial … (uses its method or data)

Code

No repository by the authors found in the paper.

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