What to consider when using artificial intelligence in capital markets

Admin
Fonetic Team
Tuesday, 14 August 2018 / Published in machine learning, voice surveillance, Trading

What to consider when using artificial intelligence in capital markets

 

AI_Linkedin-2

Artificial intelligence (AI) has come a long way since the 1950s when scientists first started exploring problem solving using computers. From the early beginnings of simple neural networks to the growth of Machine Learning, AI technology has touched nearly every industry, the Ffnancial markets are no exception.

Breakthroughs in deep learning and self-teaching algorithms are driving the AI boom we’ve seen in recent years and this has single-handedly revolutionised the financial technology (fintech) and regulatory technology (regtech) markets. Although it's important to embrace technological change, we need to understand how it works and how to use it properly.

First, let's take a step back. According to Forbes magazine, there are many definitions of AI and different industry leaders focus their research in various ways. The English Oxford Living Dictionary describes artificial intelligence as “the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages.”

Artificial intelligence is essentially the science of training machines to perform human tasks. It allows machines to carry out repetitive high-volume tasks without getting tired. This is incredibly powerful both for optimizing workloads and reducing human error in monotonous tasks. However, the human touch is still essential to ensure its success. Manual inputs both before to set up the system, ask the right questions and after, to analyse the results, to ensure absolute precision.

Nowadays, we put a lot of faith in technology. Is it all good news? We need to understand how far it's come but also its limitations and how to manage them using a variety of methods and techniques.

 

What are the difficulties when using Artificial Intelligence?

  • It's data-dependent

The main weakness of AI dependant technologies is that it learns from the data it collects. Therefore, if the data it learns from is inaccurate, the results will also be inaccurate. In other words, bad data breeds bad results. In order to achieve the highest possible accuracy in your results, you first need to start off with good quality and accurate data.

When looking at AI systems used to analyse trading floor communications, detecting and categorising that accurate piece of data could be the difference between winning or losing a multi-million-dollar lawsuit. Therefore, it's essential that any surveillance system can ensure a high accuracy rate before analysing your data.

  • Only good for one job and one job only

Another area where there's limitations, for now at least, is that AI systems are trained to perform tasks within a very narrow context. This means that an AI system trained to give you legal advice could not drive a car or detect fraud. In other words, each AI system is very specialised and useless in a context it was not built for.

  • It needs a data-rich diet

Typical solutions based on AI models require a lot of data to learn from. They need immense amounts of data to train their algorithms and deliver relevant conclusions. This is not always practical or feasible. Again, there is also no way of knowing if it’s being trained on inconsistent or incorrect data. If a machine interprets bad behaviour as the “norm”, then all the results after that will show good behaviour as outside of the norm. That doesn't seem right somehow, does it? This is potentially very dangerous. External rules must be created to ensure that data is correctly classified.

How does Fonetic use Artificial Intelligence? 

Fonetic has a crack team of computer scientists and linguists that have developed some impressive software and machine learning technology. There are a couple areas where Fonetic uses AI techniques to enhance the usability and effectiveness of its solutions.

  • Accurate results call for clean data

Employee communications data is vast, amounting to millions of gigabytes and thousands of hours for compliance teams to analyse. And as you can imagine, there is a lot of “noise” generated around the pieces of data that really matter. Communications including silences or disclaimers are eliminated to ensure our surveillance solutions are working with accurate and optimal data.

  • We know trading, we were built for trading floors

Our surveillance model was built for financial fnstitutions to detect misconduct or intent to commit fraud. Our tool can detect tactics such as language switching or the use of unauthorised channels. Financial native algorithms pre-trained with data from trading floors give a powerful base to build any market-specific policy, requiring less data in the beginning stages. 

  • Machine learning methods where it really counts

At Fonetic we have made sure that our analytics engines work with cleaned and accurate data. Communications are linked to their context in order to generate the most accurate alerts.

  • Keyword clusters can be generated from one initial word, including slang or non-standard language, to fine tune policies. There're options to manually adapt the policy afterwards using an easy-to-use interface and to include non-lexicon-based policies.

  • Intent detection technology such as language switching or detecting if employees are using unauthorised channels are essential indicators for evaluating the intent to commit fraud or misconduct.

  • Advanced natural language processing (NLP) for entity extraction. Entities can be detected based on the context of conversations, even when they are spelt wrong. This is why we reach up to 90% accuracy among 50 entities including prices, counterparties and organizations.

  • Explainability. You always know how the alert was created and what data was used to create it. This allows your institution to provide transparent reports to any auditor or regulator and comply with MiFID II and MAR regulations, where accountability is also regulated.

  • Enhanced trade reconstruction, which can reduce 20 hours of manual reconstruction into an automatic process available at the click of a button.

Does your compliance tool produce up to 90% accuracy rates when classifying data? Find out more about our technology and what it can do by downloading our product datasheet here:

Here´s your Datasheet  - Holistic

 

If you liked this post, you might enjoy:

 

RECENT POSTS