We are extremely pleased about the work done in collaboration with Alexis Stenfors on cross-market manipulation and we look forward to bringing this to the next stage. Congrats to Alexis and Features Analytics team for the results achieved!

Cristina Soviany, CEO Features Analytics

This paper presents the challenges institutions encounter when implementing or deploying market surveillance solutions. A guidance framework is provided to design and deploy solutions that can overcome the current challenges. The support technologies that should be used include artificial intelligence (AI) techniques, advanced data science and statistics. The next-generation surveillance solutions should make use of these technologies that provide a robust framework for the detection of any market abuse, both known and new emerging patterns, while reducing the operational cost.

Summary

The paper presents an overview of financial markets surveillance solutions for the detection of various abusive behaviors and application of high-performance artificial intelligence (AI)-based solutions. It features eyeDES, a cutting-edge AI-based technology and platform whose functional components provide market intelligence and unbiased detection of market abuse. eyeDES allows the detection of both previously known and completely new abusive behaviors in real time, effectively combining the use of advanced data analytics to enrich the original data space with new Features and anomaly detection to find inconsistent cases. Each case is provided with a score that measures how different that market participant’s activity is from the others, and a number of possible explanations for this. eyeDES is based on a solid and robust reasoning process, and it is an explainable AI technology, because it provides explanations of the rationale behind the decisions.

_{Cristina Soviany PhD, CEO Features Analytics}

Abstract

This paper presents a case study on the use of advanced artificial intelligence (AI) for the detection of payments fraud. The process applies AI within a typical online payment environment to detect fraudulent transactions in real time. The design focuses on an effective supervised learning engine with a data analytics component to support high-performance fraud detection, improving the predictive value of the original data. The design exploits the discriminant properties of customer data by finding hidden patterns.

This feature significantly improves fraud detection rate and performance stability compared with a rulebased solution. The developed solution, based on an advanced AI-based technology and platform increased fraud detection rate from 85 per cent to 90 per cent (in terms of number of transaction records) and to 95 per cent in related amount volume (in terms of transaction value), while the alert rate (the percentage of daily transactions investigated manually) was reduced from 40 per cent to 10 per cent. The solution falls under the category of explainable AI because it can explain the rationale behind the decisions.

Journal of Payments Strategy & Systems is only available by annual paid subscription.

Volume 12 costs £355 (Europe) $520 (US/Can) £370 (rest of the world) and includes four 100-page print issues with a free online version.

Subscribe online here.

Payments Cards & Mobile - a leading magazine for global payments news, has asked Cristina Soviany - CEO and VP R&D of Features Analytics, to talk about the applications of machine learning and big data to the financial industry. The article is entitled "Behind the Big Data hype" and it appeared in the Jan/Feb 2016 issue. In the article it is discussed how the data rich payment industry can use now machine learning to derive intelligent actions for the business.

In the article section entitled: Does the size matter? (pages 20-22), Soviany answers the question: "When it comes to data, is it really a case of bigger is better, though? How much data is enough?"

For Cristina Soviany, CEO, Features Analytics, a Belgian firm specializing in machine learning technologies, huge quantities of data alone is not enough to build accurate predictive models. Soviany notes that sometimes one can get by quite well with limited or small amounts of data. "The answer lies in the quality of data combined with the ability to enhance it with the right features, or sets of variables, able to detect hidden patterns but also evolve with the data," she says. When Features Analytics builds models to detect payment fraud, having the right type and amount of historical data is important, ideally 12-18 consecutive months of transaction data. The data also needs enough statistical coverage. "In the case of payment fraud solutions, we are used to building models where the fraud class size is 0.005 percent to 1 percent of total transaction numbers," explains Soviany. "If the volumes of data are large enough - more than hundreds of millions of transactions or samples, for example - and if the above percentages hold, then we have enough statistical data," Good data quality enables Features Analytics to apply algorithms to select the best variables, and to design new features to ensure the models learn and evolve. So, it is not always a case of bigger is better. Both quantity and quality of data matter - as well as the underlying modeling technology, which drives insights.