FPB (Financial PhraseBank)

Financial Sentiment Analysis Comparison - FPB Dataset
Financial Sentiment Analysis in FPB
Task: Analyze the sentiment of the following financial news sentence from an investor's perspective:
"According to Seppänen, the new technology UMTS900 solution network building costs are by one-third lower than that of the building of 3.5G networks, operating at 2,100 MHz frequency."
What is the sentiment of this statement?
A. Positive
B. Neutral
C. Negative
Dataset: Financial PhraseBank (FPB)
Domain: Financial news sentiment analysis

Human Evaluation
✓

Finance professionals analyze through systematic evaluation:

1 Identify Keywords
The core phrase is "costs are by one-third lower".
2 Business Meaning
A 33% reduction in network building costs is a significant financial advantage. This directly translates to higher profit margins, improved return on investment (ROI), and a stronger competitive position for the company.
3 Sentiment Score
From an investor's perspective, lower costs and higher potential profits are unequivocally positive news.
FINAL SENTIMENT: POSITIVE
Model reasoning (DeepSeek V3 Chat)
✓
Cost Efficiency: UMTS900 has building costs one-third lower than 3.5G networks. Lower costs imply higher profitability for deploying companies.
Technological Advantage: "New technology" suggests innovation and competitive advantages. Investors view tech advancements positively, especially with cost savings.
Operational Efficiency: 900 MHz operation may provide broader coverage, reducing operational expenses and improving service quality.
Conclusion: Cost savings + technological improvements = optimistic investor view. Sentiment: Positive
Model Performance: Model correctly identified the positive sentiment by recognizing key financial indicators and their implications for investors.

Figure 1. A sample financial sentiment analysis task from the FPB dataset. To solve it, financial professionals engage in systematic evaluation (middle panel): guided by domain knowledge, they identify key financial indicators and assess their investor implications. While DeepSeek correctly recognizes cost reduction as positive, it demonstrates strong domain understanding by systematically analyzing cost efficiency, technological advantages, and operational benefits, highlighting the importance of specialized knowledge in financial text analysis.

Goal

The Financial PhraseBank (FPB) is a benchmark dataset designed to advance sentiment analysis in financial and economic texts. Its primary goals are:

  • To provide a high-quality, domain-specific resource for training and evaluating sentiment analysis models in finance

  • To capture investor-centric sentiment (i.e., how news phrases may influence stock prices)

  • To address the challenge of context-dependent semantic orientations (e.g., “profit increase” vs. “loss increase”)

Description

FPB is a curated collection of English sentences from financial news and corporate disclosures, annotated for sentiment polarity.

Dataset Composition

  • Source: News articles covering companies listed on the OMX Helsinki Stock Exchange (Finland)

  • Time Period: Articles collected over a multi-year span (Unspecified)

  • Sample Size: Approximately 5,000 sentences, filtered to ensure relevance

  • Companies: Broad coverage across industries including banking, manufacturing, and technology

Annotation Process

  • Annotators: 16 finance professionals (researchers and Aalto University Business School graduates)

  • Labels: Each sentence labeled as Positive, Negative, or Neutral from an investor’s perspective

  • Agreement: - 74.9% overall agreement - 98.7% agreement on distinguishing Positive vs. Negative - Labels determined by majority voting (5-8 annotations per sentence)

Example

Below is a sample from the dataset (available on HuggingFace):

Table 2 Financial PhraseBank Examples

ID

Text

Label

Choices

fpb0

The five-storey, eco-efficient building will have a gross floor area of about 15,000 sq m. It will also include apartments.

neutral

positive, neutral, negative

fpb1

According to Seppänen, the new technology UMTS900 solution network building costs are by one-third lower than that of the building of 3.5G networks, operating at 2,100 MHz frequency.

positive

positive, neutral, negative

Task Description

Sentiment Classification

  • Input: Financial news sentence (e.g., “Profit surged by 20% due to strong demand”)

  • Output: Sentiment label (Positive/Negative/Neutral)

  • Challenge: Financial language often requires domain knowledge (e.g., “dividend cut” is negative, while “cost reduction” may be positive)

Key Use Cases

  1. Robo-advisors: Automating sentiment analysis for trading signals

  2. Academic Research: Benchmarking NLP models in finance

  3. Risk Management: Monitoring news sentiment for portfolio adjustments

Evaluation Metrics

  1. Accuracy: Primary metric for balanced class performance

  2. Macro-averaged F1-score: Accounts for class imbalance (Neutral sentences dominate)

Why Use FPB?

  • Domain-Specific: Tailored for finance, unlike generic sentiment datasets (e.g., IMDB, Yelp)

  • Rigorous Annotation: Expert-labeled with high agreement, reducing noise

  • Publicly Available: Freely accessible for non-commercial research

References

Pekka Malo, Ankur Sinha, Pekka Korhonen, Jyrki Wallenius, and Pyry Takala. “Good debt or bad debt: Detecting semantic orientations in economic texts.” Journal of the Association for Information Science and Technology 65, no. 4 (2014): 782-796.

For dataset access, check the official repository or HuggingFace link above.