Discovery News: A Generic Framework for Financial News Recommendation

Chong Wang, Lisa Kim, Grace Bang, Himani Singh, Russell Kociuba, Steven Pomerville, Xiaomo LiuAAAI 2020 · DOI

Introduction

Financial analysts are increasingly overwhelmed by the sheer volume of digital information. Traditional keyword-based alerts (e.g., searching for "Apple") often result in significant noise, returning irrelevant articles about agriculture or tangential mentions that lack financial materiality. Furthermore, the redundancy problem—where dozens of outlets report the same major event—creates a massive drain on productivity as analysts manually sift through duplicate content.

To address these challenges, researchers at S&P Global Ratings developed Discovery News, a modular, full-stack framework designed to automate the ingestion, filtering, and clustering of financial news. Unlike traditional recommendation engines that rely on Collaborative Filtering, Discovery News is built for the specialized, low-interaction environment of credit analysis, where "black swan" events are more critical than popular trends.

System Architecture

The Discovery News framework is designed for high scalability and modularity, utilizing a subscription-based parallelization strategy. The system treats every entity—whether a company, topic, or geographic location—as a separate subscription. This allows the workload to be distributed across multiple CPUs using a greedy allocation algorithm, ensuring efficient processing of massive data streams from sources like Acquire Media, Google News, and S&P Market Intelligence.

The Discovery News system architecture, illustrating the pipeline from textual data storage through the relevance and clustering layers to the final UI.

Key Pipeline Components

  • Smart De-duplication: The system identifies "near-duplicate" articles by generating Locally Sensitive Hashing (LSH) signatures and processing them through a Support Vector Machine (SVM) that evaluates Levenshtein distances and token sort ratios.
  • Entity Relevance Modeling: To determine if a company is the primary focus of an article rather than a passing mention, the framework employs an XGBoost classifier. This model analyzes features such as name position in the headline and mention frequency in the body.
  • Contextual Clustering: Discovery News utilizes ELMo (Embeddings from Language Models) to group similar stories. Unlike traditional LDA methods, ELMo provides contextualized word representations that capture the nuances of financial terminology, allowing the system to distinguish between distinct events that share similar vocabularies.

Performance and Results

The framework was evaluated using manually labeled training data and a news collection spanning three days. The results demonstrate significant improvements over baseline methods in both relevance classification and clustering accuracy.

Comparison of performance metrics between Logistic Regression and the selected XGBoost model for entity relevance.

As shown in the performance data, the fine-tuned XGBoost model achieved an F1-score of 0.866, notably outperforming the Logistic Regression baseline of 0.803. In terms of clustering, the integration of ELMo embeddings resulted in an Adjusted Rand Score of 0.3294, nearly doubling the performance of LDA-based methods (0.2843).

Beyond accuracy, the system proved highly efficient in production environments. The parallelized architecture allows the pipeline to process 1,500 subscriptions in approximately 15 minutes, scaling to nearly 5,000 subscriptions in under 45 minutes.

Conclusion and Future Work

Discovery News successfully shifts the focus of financial news recommendation from popularity to materiality. By filtering out routine noise—such as daily stock price fluctuations or minor patent filings—the system allows analysts to focus on events that truly impact creditworthiness.

Future iterations of the framework aim to incorporate direct analyst feedback loops via a deep neural network to further refine topic relevance. Additionally, the team plans to transition from a morning batch process to a near real-time update cycle and develop models that predict the specific direction of credit rating impacts based on news content.

Summary updated

BibTeX
@inproceedings{wang2020discovery,
  author = {Chong Wang and Lisa Kim and Grace Bang and Himani Singh and Russell Kociuba and Steven Pomerville and Xiaomo Liu},
  title = {{Discovery News: A Generic Framework for Financial News Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i08.7054},
}
Figure