General Financial Tasks

We consider six general financial tasks, in total 122.9k train questions and 31.7k test questions.

Table 8 Overview of general financial tasks.

Question sets	Type	#Train	#Test	Average Prompt Length	Metrics	Source	Train Data	Test Data
Financial Phrase Bank (FPB)	Sentiment Analysis	3.1k	970	56	Accuracy, F1	Hugging Face	Part of finlora_sentiment_train.jsonl	fpb_test.jsonl
Financial Question Answering (FiQA SA)	Sentiment Analysis	822	234	48	Accuracy, F1	Hugging Face	Part of finlora_sentiment_train.jsonl	fiqa_test.jsonl
Twitter Financial News Sentiment (TFNS)	Sentiment Analysis	9.5k	2.4k	52	Accuracy, F1	Hugging Face	Part of finlora_sentiment_train.jsonl	tfns_test.jsonl
News with GPT (NWGI)	Sentiment Analysis	12.9k	4.1k	81	Accuracy, F1	Hugging Face	Part of finlora_sentiment_train.jsonl	nwgi_test.jsonl
Headline	Headline Analysis	82.2k	20.5k	43	Accuracy, F1	Hugging Face	headline_train.jsonl	headline_test.jsonl
Named Entity Recognition (NER)	Named Entity Recognition	13.5k	3.5k	138	Accuracy, F1	Hugging Face	ner_train.jsonl	ner_test.jsonl

Financial Phrase Bank (FPB) (Sentiment Analysis) [fpb]

Financial Phrase Bank (FPB) contains sentences extracted from financial news and reports. These sentences are annotated with sentiment labels “positive”, “negative”, and “neutral”. We manually created the train/test split.

Instruction	What is the sentiment of this news? Please choose an answer from {negative/neutral/positive}.
Input	Pharmaceuticals group Orion Corp reported a fall in its third-quarter earnings that were hit by larger expenditures on R&D and marketing.
Output	negative

FiQA SA (Sentiment Analysis) [fiqa]

Financial question-answering sentiment analysis (FiQA SA) is another sentiment analysis dataset with the same labels as FPB from microblog headlines and financial news.

Instruction	What is the sentiment of this news? Please choose an answer from {negative/neutral/positive}.
Input	Johnson Matthey raises prospect of investor payout
Output	positive

Twitter Financial News Sentiment (TFNS) (Sentiment Analysis) [tfns]

Twitter financial news sentiment (TFNS) comprises annotated tweets related to financial news labeled with the same sentiment categories as FPB.

Instruction	What is the sentiment of this tweet? Please choose an answer from {negative/neutral/positive}.
Input	$BYND - JPMorgan reels in expectations on Beyond Meat https://t.co/bd0xbFGjkT
Output	negative

News with GPT (NWGI) (Sentiment Analysis)

News with GPT instruction (NWGI) comprises samples with seven labels ranging from strong negative to strong positive.

Instruction	What is the sentiment of this news? Please choose an answer from {strong negative/moderately negative/mildly negative/neutral/mildly positive/moderately positive/strong positive}.
Input	Amid a soft performance for the major equity indices on Tuesday, Nvidia (NASDAQ: NVDA ) posted a particularly glaring loss. Shares continued to fall in sympathy with fellow semiconductor specialist Micron Technology (NASDAQ: MU ) following its disappointing earnings results last week.
Output	moderately negative

Financial Headline Analysis (Headline Analysis) [headline-tasks]

The Headline dataset classifies headlines based on various questions into two classes: “yes” and “no”.

Instruction	Does the news headline talk about price? Please choose an answer from {Yes/No}.
Input	Gold futures edge up after two-session decline
Output	No

Named Entity Recognition (NER) (Named Entity Recognition) [ner-tasks]

The NER dataset annotates one entity per sentence, categorized into one of three classes: “location”, “person”, and “organization”.

Instruction	What is the entity type of ‘40 William St’ in the input sentence. Options: person, location, organization
Input	This LOAN AND SECURITY AGREEMENT dated January 27 , 1999 , between SILICON VALLEY BANK (” Bank “), a California - chartered bank with its principal place of business at 3003 Tasman Drive , Santa Clara , California 95054 with a loan production office located at 40 William St ., Ste .
Output	location

Fine-tuning for General Financial Tasks

To fine-tune a model for general financial tasks, you can use the configurations provided in the lora/finetune_configs.json file. Below are the configurations for each task:

Sentiment Analysis

To fine-tune a model for sentiment analysis tasks (FPB, FiQA SA, TFNS, NWGI), you can use one of the following configurations:

# Vanilla LoRA with 8-bit quantization and rank 8
python lora/finetune.py sentiment_llama_3_1_8b_8bits_r8

# QLoRA with 4-bit quantization and rank 4
python lora/finetune.py sentiment_llama_3_1_8b_4bits_r4

# DoRA with 8-bit quantization and rank 8
python lora/finetune.py sentiment_llama_3_1_8b_8bits_r8_dora

# RSLoRA with 8-bit quantization and rank 8
python lora/finetune.py sentiment_llama_3_1_8b_8bits_r8_rslora

These configurations use different combinations of quantization bits, rank, and LoRA methods:

• sentiment_llama_3_1_8b_8bits_r8: Vanilla LoRA with 8-bit quantization and rank 8, providing a good balance between performance and efficiency.

• sentiment_llama_3_1_8b_4bits_r4: QLoRA with 4-bit quantization and rank 4, reducing memory usage at the cost of some precision.

• sentiment_llama_3_1_8b_8bits_r8_dora: DoRA (Weight-Decomposed Low-Rank Adaptation) with 8-bit quantization and rank 8, which can improve performance by decomposing weights into magnitude and direction components.

• sentiment_llama_3_1_8b_8bits_r8_rslora: RSLoRA (Rank-Stabilized LoRA) with 8-bit quantization and rank 8, which uses a different scaling factor to improve stability.

Headline Analysis

To fine-tune a model for the Headline Analysis task, you can use one of the following configurations:

# Vanilla LoRA with 8-bit quantization and rank 8
python lora/finetune.py headline_llama_3_1_8b_8bits_r8

# QLoRA with 4-bit quantization and rank 4
python lora/finetune.py headline_llama_3_1_8b_4bits_r4

# DoRA with 8-bit quantization and rank 8
python lora/finetune.py headline_llama_3_1_8b_8bits_r8_dora

# RSLoRA with 8-bit quantization and rank 8
python lora/finetune.py headline_llama_3_1_8b_8bits_r8_rslora

These configurations use different combinations of quantization bits, rank, and LoRA methods:

• headline_llama_3_1_8b_8bits_r8: Vanilla LoRA with 8-bit quantization and rank 8, providing a good balance between performance and efficiency.

• headline_llama_3_1_8b_4bits_r4: QLoRA with 4-bit quantization and rank 4, reducing memory usage at the cost of some precision.

• headline_llama_3_1_8b_8bits_r8_dora: DoRA (Weight-Decomposed Low-Rank Adaptation) with 8-bit quantization and rank 8, which can improve performance by decomposing weights into magnitude and direction components.

• headline_llama_3_1_8b_8bits_r8_rslora: RSLoRA (Rank-Stabilized LoRA) with 8-bit quantization and rank 8, which uses a different scaling factor to improve stability.

Named Entity Recognition

To fine-tune a model for the Named Entity Recognition task, you can use one of the following configurations:

# Vanilla LoRA with 8-bit quantization and rank 8
python lora/finetune.py ner_llama_3_1_8b_8bits_r8

# QLoRA with 4-bit quantization and rank 4
python lora/finetune.py ner_llama_3_1_8b_4bits_r4

# DoRA with 8-bit quantization and rank 8
python lora/finetune.py ner_llama_3_1_8b_8bits_r8_dora

# RSLoRA with 8-bit quantization and rank 8
python lora/finetune.py ner_llama_3_1_8b_8bits_r8_rslora

These configurations use different combinations of quantization bits, rank, and LoRA methods:

• ner_llama_3_1_8b_8bits_r8: Vanilla LoRA with 8-bit quantization and rank 8, providing a good balance between performance and efficiency.

• ner_llama_3_1_8b_4bits_r4: QLoRA with 4-bit quantization and rank 4, reducing memory usage at the cost of some precision.

• ner_llama_3_1_8b_8bits_r8_dora: DoRA (Weight-Decomposed Low-Rank Adaptation) with 8-bit quantization and rank 8, which can improve performance by decomposing weights into magnitude and direction components.

• ner_llama_3_1_8b_8bits_r8_rslora: RSLoRA (Rank-Stabilized LoRA) with 8-bit quantization and rank 8, which uses a different scaling factor to improve stability.

Citations

[fpb]

Malo, P., H. Lu, M. Ahlgren, S. Rönnqvist, and P. Nyberg. (2014). FinancialPhraseBank-v1.0. Available at SSRN: https://ssrn.com/abstract=2512146 or http://dx.doi.org/10.2139/ssrn.2512146

[fiqa]

Sinha, A., Joglekar, M., & Murphy, F. (2018). FiQA: Financial Opinion Mining and Question Answering. arXiv preprint arXiv:1809.09431.

[tfns]

Araci, D. (2019). FinBERT: Financial Sentiment Analysis with Pre-trained Language Models. arXiv preprint arXiv:1908.10063.

[headline-tasks]

Sinha, A., & Khandait, P. (2020). Headline-Enhanced Financial Embedding. In Proceedings of the 2nd Workshop on Economics and Natural Language Processing (pp. 66-74).

[ner-tasks]

Salinas Alvarado, D., Rönnqvist, S., & Niklaus, J. (2015). Domain-Specific Named Entity Recognition: A Case Study in Finance. In Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing (pp. 110-115).