FRIDAYS: A Financial Risk Information Detecting and Analyzing System

Chi-Han Du, Yi-Shyuan Chiang, Kun-Che Tsai, Liang-Chih Liu, Ming-Feng Tsai, Chuan-Ju WangAAAI 2019 · DOI

Introduction

Financial professionals, including auditors and investors, face a daunting task when navigating SEC-mandated 10-K annual reports. These documents are notoriously dense, filled with complex jargon, and often structured to be opaque. The primary challenges identified by the authors are twofold: information overload, where manual sifting through hundreds of pages is required to assess operational risk, and the "context gap," where traditional keyword-based sentiment analysis fails to capture the nuance of financial language. For instance, a competitor's "technological breakthrough" contains positive words but represents a significant risk to the reporting company.

System Architecture

The Financial Risk Information Detecting and Analyzing System (FRIDAYS) addresses these challenges through an integrated NLP pipeline operating at three levels of granularity:

  1. Word and Multi-word Level: The system identifies Financial Words (FWs) and Financial Multiword Expressions (FMWEs), classifying them into four sentiment categories: positive, negative, litigious, and uncertain. This classification leverages established financial dictionaries and sequence labeling models.
  2. Sentence Level: Using a classification model based on the fastText architecture, FRIDAYS categorizes Financial Sentences (FSs) into high, natural, or low risk levels.
  3. Visualization Engine: The system's core is a dual-interface dashboard designed for both macro and micro analysis.
The FRIDAYS Company Overview dashboard, aggregating multi-year financial trends and textual risk metrics.

The "Company Overview" interface (Figure 1) aggregates data over multiple years, allowing users to observe textual risk trends alongside quantitative market data, such as stock prices and post-event return volatilities.

Data and Synergistic Analysis

FRIDAYS is specifically tailored for 10-K filings, building upon foundational work such as the Loughran and McDonald (2011) financial sentiment dictionary and RiskFinder (2018) for sentence-level risk detection.

A key innovation of the system is the "Synergistic Filter." This feature allows users to filter reports by both sentence risk and word sentiment simultaneously. This is particularly useful for identifying "false positives"—sentences that appear positive due to word choice but are flagged as high-risk by the sentence-level model.

The Single Report Assessment interface, demonstrating synergistic filtering of high-risk sentences with positive word sentiment.

As shown in Figure 2, the "Single Report Assessment" interface enables users to drill down into specific reports. By highlighting word-level sentiment within a risk-classified sentence, the system provides transparency, helping users understand the rationale behind the model's predictions.

Results and Utility

As a system demonstration, FRIDAYS showcases significant qualitative utility:

  • Efficiency: It enables "quick navigation," allowing professionals to bypass filler content and focus on litigious or high-risk sections.
  • Contextual Accuracy: In case studies, such as with Saucony Inc., the system correctly identified sentences regarding competitor success as "high risk" despite the presence of positive-sentiment words.
  • Educational Utility: The system serves as a tool for accounting educators to validate research assumptions regarding the correlation between financial language and market performance.

Limitations and Future Directions

While FRIDAYS provides a powerful visualization framework, the authors note several limitations. Relying solely on word counts can introduce bias; thus, the system is intended to assist, rather than replace, human judgment. Furthermore, while the system overlays stock trends, establishing a direct causal link between specific textual elements and market volatility remains a complex research area. Future iterations could benefit from integrating more advanced Transformer-based models, such as FinBERT, to further enhance contextual accuracy.

Summary updated

BibTeX
@inproceedings{du2019fridays,
  author = {Chi-Han Du and Yi-Shyuan Chiang and Kun-Che Tsai and Liang-Chih Liu and Ming-Feng Tsai and Chuan-Ju Wang},
  title = {{FRIDAYS: A Financial Risk Information Detecting and Analyzing System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33019853},
}
Figure