Paper: arXiv 2403.12285

Authors: Thanos Konstantinidis, Giorgos Iacovides, Mingxue Xu, Tony G. Constantinides, Danilo Mandic

Abstract

There are multiple sources of financial news online which influence market movements and trader’s decisions. This highlights the need for accurate sentiment analysis, in addition to having appropriate algorithmic trading techniques, to arrive at better informed trading decisions. Standard lexicon based sentiment approaches have demonstrated their power in aiding financial decisions. However, they are known to suffer from issues related to context sensitivity and word ordering. Large Language Models (LLMs) can also be used in this context, but they are not finance-specific and tend to require significant computational resources. To facilitate a finance specific LLM framework, we introduce a novel approach based on the Llama 2 7B foundational model, in order to benefit from its generative nature and comprehensive language manipulation. This is achieved by fine-tuning the Llama2 7B model on a small portion of supervised financial sentiment analysis data, so as to jointly handle the complexities of financial lexicon and context, and further equipping it with a neural network based decision mechanism. Such a generator-classifier scheme, referred to as FinLlama, is trained not only to classify the sentiment valence but also quantify its strength, thus offering traders a nuanced insight into financial news articles. Complementing this, the implementation of parameter-efficient fine-tuning through LoRA optimises trainable parameters, thus minimising computational and memory requirements, without sacrificing accuracy. Simulation results demonstrate the ability of the proposed FinLlama to provide a framework for enhanced portfolio management decisions and increased market returns. These results underpin the ability of FinLlama to construct high-return portfolios which exhibit enhanced resilience, even during volatile periods and unpredictable market events.

Complexity vs Empirical Score

  • Math Complexity: 3.5/10
  • Empirical Rigor: 6.0/10 (capped from 7.5: transaction costs not modelled; no out-of-sample test; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 7.5)

Why this score: The paper focuses on implementing and evaluating a specific NLP architecture (LLM fine-tuning) using concrete datasets and portfolio backtests, indicating high empirical rigor but relatively low mathematical formalism compared to theoretical derivations.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Develop Finance-Specific<br>Sentiment Classification LLM"] --> B
  subgraph B ["Methodology"]
      direction LR
      B1["Fine-tune Llama 2 7B<br>via LoRA"] --> B2["Generator-Classify Scheme<br>Sentiment & Strength"]
  end
  B --> C["Input<br>Financial News Text"]
  C --> D["Computational Process<br>FinLlama Analysis"]
  D --> E["Output<br>Sentiment Valence &<br>Strength Quantification"]
  E --> F["Outcome<br>Enhanced Portfolio Management<br>Higher Market Returns"]

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 FinLlama, a fine‑tuned Llama 2 7B model for financial sentiment classification and strength quantification.

Source sentence
Our proposed FinLlama is one such solution, which is obtained by fine-tuning a pre-trained LLM (namely Llama 2 7B [2]) on specialised, labelled and publicly available financial news datasets.

Assumes 1

  • A risk‑free rate of 0% is used for Sharpe ratio calculations. market
    Source sentence
    a 0% rate is commonly used and is adopted in our analysis.

Data

  • Four labelled publicly available financial news datasets text-news public
    Source sentence
    Our training data was a combination of four labelled publicly available financial news datasets, resulting in a comprehensive collection of 34,180 labelled samples.
  • Financial news articles text-news public
    2015-2021
    Source sentence
    204,017 articles dating between 2015 to 2021 were collected from online sources.
  • Yahoo Finance S&P 500 daily returns · Yahoo Finance market-prices public
    2015-2021 · S&P 500 constituents
    Source sentence
    Financial market data were collected for the same time period from Yahoo Finance. These collected market data contained daily stock returns for the 500 companies in our Investable Universe (S&P 500), resulting in 1,672 days of stock returns data for each company.

Recognised sources: Bloomberg

Costs

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

Finds 3

  • FinLlama’s 35% long‑short portfolio achieved the highest cumulative return of 308.2%, outperforming all other sentiment methods. performance
    Source sentence
    Cumulative Returns (%) 204.6 100.4 130.6 213.0 308.2 83.1
  • FinLlama generated significantly higher returns while reducing portfolio risk, as shown by higher Sharpe and lower annualized volatility. performance
    Source sentence
    Overall, our FinLlama model successfully generated significantly higher returns for investors, whilst simultaneously reducing portfolio risk, as indicated by the higher Sharpe ratio and lower annualized volatility.
  • Deep learning approaches, including FinLlama, outperformed lexicon‑based methods in cumulative returns, especially during high‑volatility periods. performance
    Source sentence
    Observe that the deep learning approaches outperformed the lexicon-based approaches in terms of cumulative returns, particularly those relying on general-purpose dictionaries (HIV-4 and VADER).

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 6 papers in the archive (6 by independent authors): Execution Realism and Reproducibility in LLM-Based Trading … (uses its method or data); Large Language Model Agent in Financial Trading: A Survey (uses its method or data); From Deep Learning to LLMs: A survey of AI in Quantitative … (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.

Cited References (48)

  1. Llama 2: Open foundation and fine-tuned chat models (2023) — link
  2. LoRA : Low-rank adaptation of large language models (2021) — link
  3. Efficient capital markets: A review of theory and empirical work (1970)
  4. Evaluation of Sentiment Analysis in Finance: From Lexicons to Transformers (2020) — link
  5. An introduction to support vector machines and other kernel-based learning methods (2000)
  6. Hierarchical Attention Networks for Document Classification (2016) — link
  7. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2019) — link
  8. Instruct- FinGPT : Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models — → 2306.12659
and 40 more
  1. BloombergGPT : A Large Language Model for Finance (2023) — link
  2. Fixing Weight Decay Regularization in Adam (2017) — link
  3. When Is a Liability NOT a Liability? Textual Analysis, Dictionaries, and 10- Ks (2011) — link
  4. The General Inquirer: A Computer Approach to Content Analysis (1966)
  5. VADER : A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text (2015)
  6. Predicting Returns With Text Data (2019) — link
  7. Corporate Finance (2019)
  8. FinBERT-FOMC : Fine-Tuned FinBERT Model with Sentiment Focus Method for Enhancing Sentiment Analysis of FOMC Minutes (2023)
  9. Portfolio selection (1952)
  10. When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks (2011)
  11. News impact on stock price return via sentiment analysis (2014)
  12. Predicting returns with text data (2019)
  13. Convolutional networks for images, speech, and time series (1995)
  14. Learning task-dependent distributed representations by backpropagation through structure (1996)
  15. Long short-term memory (1997)
  16. A neural probabilistic language model (2000)
  17. Recurrent neural networks for prediction: learning algorithms, architectures and stability (2001)
  18. Recurrent neural network based language model. (2010)
  19. Generating sequences with recurrent neural networks (2013) — link
  20. Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014) — link
  21. Convolutional neural networks for sentence classification (2014) — link
  22. Sequence to sequence learning with neural networks (2014)
  23. Neural Machine Translation by Jointly Learning to Align and Translate (2015) — link
  24. Attention Is All You Need (2017)
  25. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018) — link
  26. Improving Language Understanding by Generative Pre-Training (2018)
  27. Language models are unsupervised multitask learners (2019)
  28. Language models are few-shot learners (2020) — link
  29. Flashattention: Fast and memory-efficient exact attention with io-awareness (2022)
  30. Gptq: Accurate post-training quantization for generative pre-trained transformers (2022) — link
  31. Distilling the knowledge in a neural network (2015) — link
  32. ReAugKD: Retrieval-augmented knowledge distillation for pre-trained language models (2023)
  33. FinBERT : Financial sentiment analysis with pre-trained language models (2019) — link
  34. When flue meets flang: Benchmarks and large pre-trained language model for financial domain (2022) — link
  35. FinGPT : Open-Source Financial Large Language Models — → 2306.06031
  36. FinGPT : Democratizing Internet-scale Data for Financial Large Language Models (2023) — link
  37. PIXIU: A Large Language Model, Instruction Data and Evaluation Benchmark for Finance (2023) — link
  38. Treasury yield 10 years historical data. (2023) — link
  39. Training language models to follow instructions with human feedback (2022) — link
  40. TensorGPT : Efficient Compression of the Embedding Layer in LLMs based on the Tensor-Train Decomposition (2023) — link

2 of 48 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.