AWS re:Invent 2025 - How FSI Revolutionized HFT Analytics with Agentic AI (GBL302)

Summary of AWS re:Invent 2025 - How FSI Revolutionized HFT Analytics with Agentic AI (GBL302)

Overview of High-Frequency Trading (HFT) and Market Making

• In the trading ecosystem, there are buyers and sellers who place orders on exchanges
• Market makers play a crucial role by providing liquidity by continuously quoting buy and sell orders
• Market makers aim to execute trades quickly and at small profits, relying on high trade volume and turnover
• Volatility is a key factor that affects market makers' profitability, as they need to quickly adapt their orders based on market movements

Challenges in Analyzing Market-Moving News and Events

• Traditional approaches to sentiment analysis, such as dictionary-based methods and supervised machine learning, have limitations in handling context and require significant manual effort
• News and events, especially from influential figures like Elon Musk and Donald Trump, can have a significant and immediate impact on digital asset markets
• Interpreting these events and their market impact requires fast, context-aware analysis, which is difficult for manual or rule-based approaches

Agentic AI Approach to News Sentiment Analysis

• Leveraging large language models (LLMs) like GPT-3 for context-aware, multi-modal reasoning on news and events
• Optimizing the inference performance of the LLM to achieve low latency (180 tokens/second) through techniques like mixed precision, linear attention, and distributed expert parallelism
• Implementing a deduplication pipeline to efficiently process the high volume of duplicate news reports in the crypto market

Technical Implementation Details

1. Data Ingestion and Processing Pipeline:

   • Ingesting news streams directly into an S3 bucket and triggering Lambda functions for classification and metadata tagging
   • Utilizing Amazon Aurora Postgres and OpenSearch for storing and querying news data
   • Providing a QCI interface for traders and analysts to interact with the news data and query the LLM

2. Deduplication Pipeline:

   • Calculating news article embeddings using a fine-tuned BERT-based model (BGM3)
   • Performing similarity checks to identify and filter out duplicate news reports
   • Maintaining separate collections in OpenSearch for unique and duplicate news articles

3. Sentiment Analysis and Impact Assessment:

   • Using the LLM to perform context-aware sentiment analysis and assess the potential market impact of news
   • Generating spread widening recommendations and price movement probability estimates
   • Sending the analysis results to the trader desk for review and action

Key Takeaways and Business Impact

• Agentic AI architecture leveraging LLMs and fine-tuned embeddings enables faster, more accurate, and more cost-effective news sentiment analysis compared to traditional approaches
• The ability to rapidly process and analyze market-moving news and events in near real-time allows market makers to make more informed and timely decisions, improving their profitability and risk management
• The deduplication pipeline and fine-tuned embeddings address the unique challenges of the crypto market, where the same news is often reported across multiple channels within minutes
• The human-in-the-loop approach, with traders providing feedback to continuously improve the system, ensures a balance between automation and human oversight

Conclusion

The presented solution demonstrates how financial institutions can leverage agentic AI and LLMs to revolutionize their high-frequency trading analytics, enabling them to stay ahead of the curve in rapidly evolving digital asset markets.