Beginning in May 2024, the FCA ran a 3-month Market Abuse Surveillance TechSprint to explore how advanced solutions leveraging artificial intelligence (AI) and machine learning (ML) could help detect evolving forms of market abuse.
The TechSprint aimed to improve market surveillance in firms and make processes more resilient.
We were particularly interested in seeing how new technologies, such as AI can help accurately identify more complex types of market abuse that are currently difficult to detect, such as cross-market manipulation.
Nine teams from across the globe including Belgium, Australia, Ukraine, USA and the UK worked together over the 3-month period. At a Showcase Demo Day on 20 July 2024, teams presented solutions focused on eliminating harm and ensuring a fair and competitive market.
Over 200 representatives from RegTech and AI vendors, public sector agencies, academia, major banks and senior representatives from the FCA attended on the day.
You can see video recordings of each team’s presentation below.
Problem statements
The TechSprint aimed to tackle 3 set problem statements:
- How can AI and advanced analytics improve accuracy and frequency of post-trade market surveillance alerts and signals.
- How can AI and advanced analytics help identify instances of complex types of market abuse that traditional rules-based surveillance tools struggle to identify.
- How can AI and advanced analytics help identify market and trading anomalies indicative of market abuse, manipulative strategies, and disruptive trading practices.
Participants were provided with access to 1TB of pseudonymised markets data, including transaction reports, order books, news feeds, and price feeds, facilitating the development and testing of advanced AI/ML models in the data-rich environment of the Digital Sandbox.
Showcase Demo Day
The solutions developed over the 3 months and presented during the Showcase Demo Day included:
- AI/ML models such as isolation forests to reduce false positives and improve the accuracy of alerts, ensuring genuine cases of market abuse are identified more reliably.
- Anomaly detection techniques such as Bayesian Network Analysis were employed to identify unusual trading patterns indicative of manipulative strategies and disruptive trading practices based on historical data.
- The application of econophysics theory and kinetic energy models to identify subtle price changes in orderbook data and large language models to contextualise outliers and filter out false positives.
Showcase demo day presentations
Problem 1: accuracy and frequency of post-trade market surveillance alerts and signals
Features Analytics
A parameter-free solution with machine learning algorithms that automatically calibrate to market conditions. An eyeDES platform generates alerts with contextual information, designed to function during high volatility. A feedback loop tool incorporates analyst decisions to train a model that categorises alerts by relevance, with the capability for automatic retraining.
Video: watch Features Analytics' presentation
eflow
A case management tool that identifies potential market abuse using flexible parameterisation of market and non-market data. The system incorporates machine learning to dynamically update parameters and improve detection accuracy over time.
Video: watch eflow's presentation
Problem 2: identifying complex market abuse
Pytilia
A surveillance platform leveraging a multi-layer, multi-dimensional approach using volume and time metrics. The platform is designed for rapid reconfiguration to accommodate new models, inputs, and rules, and features a flexible architecture supporting integration with various transaction reporting systems and formats.
Video: watch Pytilia's presentation
Spring Autumn (KCL & UCL)
The proposal introduces an econophysics model that analyses limit order book dynamics by treating orders as particles in a physical system. This approach aims to identify subtle market manipulation patterns that conventional systems might miss.
Video: watch Spring Autumn's presentation
Steel Eye
The solution correlates trade orders, communications, and market data using AI to identify coordinated manipulative actions affecting prices. This approach combines efficient screening of trade and order data across multiple asset classes with LLM analysis to assess potential violations, aiming to improve the accuracy and efficiency of market abuse detection.
Video: watch Steel Eye's presentation
Problem 3: identifying market and trading anomalies
Conatix
Uses AI to classify time-series data within asset and entity classes for anomaly detection. The model analyses real-time trading logs to flag anomalous transactions, aiming to enhance the accuracy of identifying suspicious activities while providing explainable results for fraud analysts.
Video: watch Conatix's presentation
b-next
Analyses non-linear time-series data using AI/ML for normalisation, classification, and anomaly detection.
Video: watch b-next's presentation
Kaizen
The solution introduces a multi-modal approach to anomaly detection in trading data. The combination of techniques aims to enhance the accuracy of identifying anomalous trades in financial scenarios.
Video: watch Kaizen's presentation
DHI
The solution utilises Bayesian Networks to develop an anomaly prediction and ranking model. This approach is particularly effective in scenarios with limited data or minimal anomaly events.
