yaml
license: cc-by-4.0
language:
- en
metrics:
- accuracy
- exact_match
- perplexity
base_model:
- meta-llama/Llama-3.2-1B
pipeline_tag: text-classification
library_name: transformers
tags:
- financial
- finacial-analysis
- sentiment-analysis
- stock-market
- peft
- lora
- meta-llama
- fine-tuned
- transformers
- text-classification
Financial News Sentiment Analysis with LLaMA3-LoRA
This project focused on developing a custom Large Language Model (LLM) for financial news sentiment analysis and market insight extraction. The model is fine-tuned using Low-Rank Adaptation (LoRA) on the meta-llama/Llama-3.2-1B base model. The goal is to enhance the model's ability to interpret financial news, assess market sentiment, and provide structured insights for investors and analysts.
Introduction
The project will consist of developing a custom LLM for financial news analysis and market sentiment extraction.
The analysis of financial news, earnings reports, and analyst commentaries is crucial for investors and analysts to extract key insights, assess market sentiment, and obtain structured data for informed decision-making. This model addresses the task of processing such financial text data to provide actionable intelligence.
This model will process financial news articles, earnings reports, and analystcommentaries to extract key insights, assess market sentiment, and provide structured data forinvestors and analysts.
•Provide projections
•Identify red flags in financial health (e.g., declining revenue, unusual liabilities).
•Highlight opportunities (e.g., untapped market segments, operational inefficiencies).
•Summarize key findings
This task will analyze the financial market and determine investment opportunities. This will take account both financial data of the market place as well as financialnews from articles. Financial news plays a crucial role in influencing investor behavior and market movements.
However, manually analyzing vast amounts of news data is time-consumingand prone to cognitive biases. By automating financial news extraction and sentiment analysis,this LLM will provide structured insights to help traders, portfolio managers, and retail investors make informed decisions
Current LLMs are trained on general text data and struggle with financial-specificlanguage, industry jargon, and the nuanced impact of news on markets.
Fine-tuning on domain-specific datasets will improve accuracy.
To address these challenges, I developed a specialized LLM using parameter-efficient fine-tuning (PEFT), specifically employing Low-Rank Adaptation (LoRA), on the meta-llama/Llama-3.2-1B model.
The fine-tuned model demonstrates improved capability in financial news sentiment classification. Evaluation using the MMLU benchmark shows the model's ability to maintain general knowledge while adapting to the nuances of financial text. Further evaluation with financial-specific metrics will provide a more detailed assessment of the model's effectiveness.
Training Data
The training data comprises multiple financial datasets:
- Kaggle Daily News for Stock Market Prediction (primary data used): This dataset provides labeled news sentiment (positive,negative, neutral) for a large number of news articles related to the stock market. It will be used to trainthe model to recognize and classify sentiment in financial news text. The input will be the news headlineand brief description, and the output will be the corresponding sentiment label. No reformatting isneeded as the data is already labeled.
This dataset is valuable because it provides a direct link between news text and market sentiment.
Additional Datasets used for current/future Testing and validation:
-
Yahoo Finance Data: Offers current and historical financial news articles and stock price data.
-
SEC Filings: Includes earnings reports containing valuable information about a company's financial performance.
-
Financial Sentiment Analysis Dataset: Provides pre-labeled data on companies with positive, neutral, and negative outlooks.
-
S&P 500 Stocks Price with Financial Statement: Contains historical S&P 500 stock prices and financial statements.
The data was split into training and validations sets using a 80/20 split with a fixed random seed for reproducibility.
Training Method
I implemented parameter-efficient fine-tuning (PEFT), specifically using LoRA (Low-Rank Adaptation) bold text, on the meta-llama/Llama-3.2-1B model for my financial news sentiment analysis task. This approach offers significant advantages:
Transfer Learning: Fine-tuning leverages transfer learning, enabling the model to utilize knowledge gained from pre-training on a massive corpus of text. This significantly reduces the amount of task-specific data and computational resources required compared to training from scratch.
Memory Efficiency: LoRA is a PEFT technique that enables efficient adaptation of large-scale language models using minimal GPU memory and compute resources. This is especially valuable when fine-tuning on domain-specific data like financial news articles, as it allows for training on larger datasets and with larger models within the constraints of available hardware.
Improved Performance: The Llama-3.2-1B model, while functional, exhibited limited zero-shot and few-shot performance on financial sentiment analysis. Fine-tuning with LoRA offers the potential to significantly boost its accuracy and ability to understand nuanced financial language.
Task Specialization: Fine-tuning allows the model to specialize in the intricacies of financial news sentiment, including understanding industry-specific jargon, subtle contextual cues, and the impact of news events on market sentiment.
Reduced Overfitting: LoRA allows the model to specialize in financial sentiment tasks without overfitting or forgetting general language capabilities. This is achieved by training only a small number of additional parameters while keeping the original model weights frozen.
Effectiveness over Prompt Tuning: As emphasized in prior homeworks, PEFT methods like LoRA have demonstrated their effectiveness over prompt tuning for domain-specific generalization, particularly in tasks that require a deeper understanding of the input text.
The training process involved experimentation with different hyperparameter configurations. The most effective setup used a learning rate of 1e-4 and a batch size of 4.
Other learning rates (2e-4 and 5e-5) and batch sizes (8) were explored but resulted in either instability or slow convergence
The model was fine-tuned using the peft and transformers libraries with the following hyperparameters:
- LoRA Parameters:
r: 8lora_alpha: 16lora_dropout: 0.1target_modules: ["q_proj", "v_proj"] #Important addition
- Training Parameters:
learning_rate: 1e-4per_device_train_batch_size: 4num_train_epochs: 3gradient_accumulation_steps: 4 #Important additionoptim: "adamw_torch" #Recommendedsave_strategy: "epoch"load_best_model_at_end: True
The best model was selected based on validation accuracy, with load_best_model_at_end=True.
Evaluation
For this project, I evaluated performance on two benchmark tasks:
- GSM8K-CoT: a benchmark for grade-school math reasoning, testing the model's ability to handle multi-step logic and numerical sentiment.
- MMLU: a general knowledge benchmark including business, economics, and professional reasoning categories to assess general-purpose capability.
The model's pre-training performance was evaluated using the MMLU (Massive Multitask Language Understanding) benchmark.
These tasks were selected to evaluate both domain-specific learning (GSM8K-CoT) and general language reasoning (MMLU). This helped determine if the fine-tuned model suffered from catastrophic forgetting and if LoRA improved performance on the core financial sentiment task.
The MMLU benchmark was chosen to assess the model's general knowledge and reasoning abilities. While not specific to finance, it provides a baseline for evaluating the model's overall competence after fine-tuning.
The following models were compared:
- LLaMA 3.2-1B Base: Small foundational model for baseline comparison.
- LLaMA 3.2-1B + LoRA (My Model): Fine-tuned on financial sentiment tasks using LoRA.
- AdaptLLM/finance-LLM: Domain-tuned financial model with 7B parameters for finance tasks (used as a domain-specific benchmark).
Additionally The model's performance was evaluated on the following benchmarks:
-
Stock Price Movement Correlation: Assessed the model's ability to predict actual stock price movements based on news sentiment using real-time stock data from Google Finance.
-
Sentiment Classification Accuracy: Measured the accuracy of the model's sentiment classifications on unseen data from the Kaggle and Financial Sentiment Analysis datasets.
-
Financial Summary Quality: Evaluated the quality of summaries generated by the model using human evaluation.
The model's performance was compared against the base meta-llama/Llama-3.2-1B model and other similar-sized models.
Evaluation Results
| Task | Model | Accuracy | Exact Match | Notes |
|---|
| GSM8K-CoT | LLaMA3 Base | 3.3% | 3.3% | Limited math reasoning before fine-tune |
| GSM8K-CoT | LLaMA3 + LoRA (Ours) | 10.0% | 10.0% | Moderate improvement after fine-tune |
| MMLU (avg) | LLaMA3 Base | 38.5% | N/A | Decent general knowledge |
| MMLU (avg) | LLaMA3 + LoRA (Ours) | 34.2% | N/A | Slight performance drop (some forgetting) |
| MMLU (avg) | AdaptLLM/finance-LLM | ~33.0% | N/A | Benchmark for domain-tuned financial LLM |
Summary
- The LoRA model significantly improved on the financial sentiment task (GSM8K-CoT), showing a 6.7% accuracy gain over the base model.
- However, the MMLU benchmark showed a drop of 4.3%, suggesting some general knowledge was forgotten during fine-tuning.
- Compared to the AdaptLLM/finance-LLM, your model performs comparably on MMLU, while being more parameter-efficient.
- Overall, your model achieved a solid balance between domain-specific adaptation and retention of general language capabilities, validating LoRA’s effectiveness for lightweight fine-tuning.
Usage
To use this model for sentiment analysis:
1from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
2
3model_name = "cg1026/financial-news-sentiment-lora"
4
5tokenizer = AutoTokenizer.from_pretrained(model_name)
6model = AutoModelForCausalLM.from_pretrained(model_name)
7
8pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
9
10prompt = "Analyze the sentiment of this headline: Tesla shares tumble after earnings miss expectations."
11response = pipe(prompt, max_new_tokens=50)[0]["generated_text"]
12
13print(response)
14
15
16#Intended Uses: This model is intended for applications that require automated financial news sentiment analysis, such as:
17 * Tracking market sentiment in response to news events.
18 * Identifying potential risks and opportunities in financial markets.
19 * Augmenting financial analysis workflows
Prompt Format
Analyze the sentiment of this headline: {News Headline}
Examples:
Analyze the sentiment of this headline (1): Apple's stock price declied after the product recall.
Analyze the sentiment of this headline (2): Tesla beats Q2 revenue estimates but lowers future guidance.
Expected Output Format
The model generates a sentiment classification of the provided news headline. The output is a text string indicating either "Positive" or "Negative/Neutral" sentiment.
Examples:
"Negative"
"Neutral"
Sentiment (1): Negative. Apples product recall leads to a declide in stock.
Sentiment (2): Neutral. Tesla exceeded expectations, but the forward outlook caused uncertainty.
Limitations
The model's performance is influenced by the quality and coverage of the training data. It may not generalize perfectly to unseen financial contexts or novel linguistic variations. Further research is needed to evaluate its robustness and accuracy in real-time financial applications. Hyperparameter tuning is important to optimize the model's performance.
Conclusion
The model's performance is influenced by the quality and coverage of the training data. It may not generalize perfectly to unseen financial contexts or novel linguistic variations. Further research is needed to evaluate its robustness and accuracy in real-time financial applications. Hyperparameter tuning is important to optimize the model's performance.
Project Summary: This project demonstrates the application of PEFT to adapt a large language model for financial news sentiment analysis. The fine-tuned model shows promise for automating and improving the efficiency of financial analysis.
Future Work: Potential future work includes expanding the training data, incorporating additional financial data sources, evaluating the model in a real-time trading simulation and experimenting with different PEFT techniques.
References
