Compliance monitoring · Production

LSEG builds AI-powered Surveillance Guide on Amazon Bedrock to classify price-sensitive news and detect market abuse

The problem

LSEG's market surveillance analysts faced a time-consuming manual triage process: after automated alerts flagged suspicious trading, analysts had to manually collate evidence and assess news price sensitivity. High false-positive alert volumes degraded analyst quality time and risked operational delays.

First attempt

Existing automated rules-based surveillance systems were outdated and generated an increasing number of false-positive alerts, overwhelming analysts with low-value cases and becoming inadequate as market tactics evolved.

Workflow diagram · grounded in source
1
Surveillance alert triggers review
trigger
“LSEG's surveillance monitoring systems generate automated, customized alerts to flag suspicious trading activity to the Market Supervision team”
2
Ingest and preprocess RNS articles
integration
“Ingest raw RNS news documents in HTML format Preprocess and extract clean news text”
3
Claude classifies news price sensitivity
ai_action
“The system employs Anthropic's Claude Sonnet 3.5 model—the most price performant model at the time—through Amazon Bedrock to analyze news content from London Stock Exchange's Regulatory News Service (RNS) and classify articles based on t…”
4
Model outputs summary and justification
output
“The prompt templates elicited three key components from the model: - A concise summary of the news article - A price sensitivity classification - A chain-of-thought explanation justifying the classification decision”
5
Results visualized via Streamlit
output
“Present results through the visualization interface developed using Streamlit”
6
Analyst triages flagged cases
human_review
“Analysts then conduct initial triage assessments to determine whether the activity warrants further investigation”
Reported outcome

Over a 6-week evaluation period, the Surveillance Guide achieved 100% precision for identifying non-sensitive news and 100% recall for detecting price-sensitive content, reducing manual review time and enabling analysts to focus on the most critical cases.

Reported metrics
Precision for non-sensitive news classification100%
Recall for price-sensitive content detection100%
Manual review time for analystsReducing manual review time
Evaluation period duration6-week period
Reported stack
Claude Sonnet 3.5Amazon SageMakerJupyter NotebooksStreamlitInstructor libraryPythonRegulatory News Service (RNS)
Source
https://aws.amazon.com/blogs/machine-learning/how-london-stock-exchange-group-is-detecting-market-abuse-with-their-ai-powered-surveillance-guide-on-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Over a 6-week evaluation period, the Surveillance Guide achieved 100% precision for identifying non-sensitive news and 100% recall for detecting price-sensitive content, reducing manual review time and enabling analys…

What tools did this team use?

Claude Sonnet 3.5, Amazon SageMaker, Jupyter Notebooks, Streamlit, Instructor library, Python, Regulatory News Service (RNS).

What results were reported?

Precision for non-sensitive news classification: 100%; Recall for price-sensitive content detection: 100%; Manual review time for analysts: Reducing manual review time; Evaluation period duration: 6-week period (source-reported, not independently verified).

What failed first in this deployment?

Existing automated rules-based surveillance systems were outdated and generated an increasing number of false-positive alerts, overwhelming analysts with low-value cases and becoming inadequate as market tactics evolved.

How is this compliance monitoring AI workflow structured?

Surveillance alert triggers review → Ingest and preprocess RNS articles → Claude classifies news price sensitivity → Model outputs summary and justification → Results visualized via Streamlit → Analyst triages flagged cases.

Related compliance monitoring cases