How AI can Detect Trader Fraud Issues on the Trading Floor

Nigel Cannings, CTO at Intelligent Voice will be speaking on this topic at NVIDIA GTC, taking place online April 12 – 16, 2021. Register free and join us for his session 14th April 5pm BST SS31452: ‘Detecting Deception and Tackling Insurance Fraud using GPU-powered Conversational AI’ plus so much more. GTC offers free access to 1400+ sessions covering the latest breakthroughs in AI, data center, accelerated computing, healthcare, intelligent networking, and more. Click here to learn more. 

Voice is the centre of communication for the financial markets. It always has been, and with advances in technology voice remains an imperative part of the trading process. Whether using desk phones, cell phones, video conferencing platforms, software-based voice trading systems or physical trading turrets, voice continues to be the favoured communication channel for trading. 

While IM chat has been making significant inroads into voice communications over the years, the COVID-19 pandemic changed that.  COVID has affected economies and financial markets in unprecedented ways, radically disrupting work environments around the world. This paired with trading increasingly becoming more electronic and algorithm-driven, could lead you to assume that one to one conversation would be overtaken by trading platforms. However according to research by Greenwich voice/chat was the preferred channel at every stage of the trading process during the crisis and is expected to be going forward. Video conferencing usage for work has risen by X according to Y during the pandemic.

As compliance monitoring requirements become stricter, impacting more processes, with attention on comprehensive and accurate data, financial institutions need to ensure their voice recording infrastructure can deliver these requirements. In the UK, the FCA has recently reinforced the need for firms to record all of their voice and video interaction in Market Watch 66.

Capturing and analysing voice can be a challenging in a trading environment. Traders speak fast, often speaking in jargon or using specific terminology. It is also common for traders to speak in multiple languages, sometimes switching languages in the same sentence. Call times can last for minutes, or can go on for hours with very little conversation, and with numerous people leaving and entering the same call. These issues produce difficult conversational data that is full of irrelevant content and false positives.

Using artificial intelligence (AI) and machine learning puts unstructured data into a structured format removing the manual effort of surveillance teams. Conversational AI and speech analytics engines can spot suspicious activities and signs of fraud.

How can AI detect trader fraud?

Alert Terms

Advanced speech recognition and the ability to continually perfect a language model with financial services and trading-specific area lexicons and acoustic models make it simpler and faster to extract accurate voice data.

Alert terms are a useful tool to recognise when corruption is taking place, examples of these in the finance market could be “higher fixing” or “on my home number”. These phases can be identified and flagged as problematic, however traders communicating in codewords with hidden meanings will not appear on search term lists and can lead to fraud going undetected.

Using AI to complement human searching can also identify fraud and decrease risk by recognizing themes that may have slipped past a reviewer or not been caught by human designed alert terms. AI analyses and learns from concepts, so you can identify those innocuous terms, flagging them to surveillance teams. Offering increased capabilities to transform an organization’s risk mitigation strategy.

Behavioural Analytics

With conversational AI, semantic parsing allows the meaning to be extracted from each and every word, immediately.

We are able, at scale, to infer how positive an interaction is, how the participants are feeling in relation to the subject or subjects of the conversation, and how strong the relationship is between the participants.  We are also able to track changes in these measures over time.

Using a trained human to do this properly could take several hours for each phone call.  Using Conversational AI, the same job can be done in seconds. Here is an example of :

5 December 2007, Manager D stated (in an internal telephone conversation with Manager E) that he had “touched on [the] topic” of LIBOR submissions at a meeting with the FSA.

Manager D stated “we didn’t say anything along the lines of, you know, we’re not posting where we think we should […] because of. I just talked about dislocations, LIBORs […] and kept it […] simple, shall we say”.

This short two sentence answer is full of red flags. So much so that when given to our law-enforcement trained Forensic Linguistic Analyst they came back with two pages of explanation as to why. Here is just a snippet of that:

Manager D starts by saying in the negative what wasn’t said. This is sensitive to Manager D. The expected is for him to either simply say what was said or to deny any allegation. Also, of note is the pronoun “we.” Whatever wasn’t said relates to more than one. Manager D doesn’t want to be alone at this point in the statement. The shortest answer is best. His answer breaks the principle of brevity, thus giving the impression of a need to convince as opposed to simply convey information.

Markers of negation, explainers, hedging, all would be flagged by AI and machine learning. Intelligent Voice’s solution LexiQal identifies all these ‘markers’ triggering the system to raise an alert. To find out more about LexiQal and conversational AI join Intelligent Voice’s GTC session, SS31452: ‘Detecting Deception and Tackling Insurance Fraud using GPU-powered Conversational AI’ 14th April 5pm BST.

Language Switching

The finance market is global, it is common for traders to speak in multiple languages. Sometimes switching languages in the same sentence, allowing fraud to slip through undetected. In an increasingly global world, we cannot expect every conversation to take place in English. Which is why it is important that voice data capture can identify different languages and dialects.

AI and machine learning allows for speech recognition engines to be trained in different languages and dialects, and even lets them run simultaneously. Allowing the capture of every nuance of what is said, no matter what language.  


Knowing what was said and by whom, can be important evidence in fraudulent activity. In a call about a specific transaction, there could be multiple parties joining or leaving the conversation at any time, it is an important requirement to identify and associate all the necessary metadata with the media file to retain, retrieve and make sense of the call at a later stage.

Often the counterparty metadata is limited to just an institution name:  We need to know whether a person has called before, and whether they are speaking to the person they usually speak to.  Or has someone misused a turret, logging in as someone they are not to evade identification.  Biometrics helps us identify these behaviours.


Financial services organizations are required to record communications to be used as evidence in case of disputes or regulatory requests. Across Europe, companies are obliged to not just record but also monitor the recordings of trades, “periodically monitor the records of transactions and orders subject to these requirements, including relevant conversations.” [Article 16(7)] MiFID II.

As stated above, the COVID-19 pandemic has also led to authorities becoming tougher on recording work from home communications. The UK Financial Conduct Authority (FCA) recently published information on their expectations of communication recording during remote working. More can be read about this here.

This requires full transcription and constant monitoring of conservations rather than selective sampling. Which is a continual burden on compliance officers, operational and technical teams to regularly monitor and verify the state and performance of the recordings especially in a tough environment such as trading.

This is where AI can relieve much of this burden. Using the above examples, biometrics, behavioural analytics, metadata and search help give the ability to detect or identify information that is similar or identical. All this data can be used in an explainable AI solution permitting data into a readable format for humans such as a SmartTranscript or dashboard. This data can then be used as evidence to create defensible claims.

Platform Based

Taking all this data and putting it into a simple-to-implement platform allows fraud to be detected quickly and with reduced false positives. Using clever GPU technology to build a solution that is robust and scalable, we create systems that can process hundreds or thousands of conversations, giving you access to timely compliance monitoring.  In most environments, this voice data can be seamlessly added to existing compliance workflows without expensive retooling.

Businesses should consider using speech recognition technology that is based on APIs designed to integrate with a multiple platforms and existing solutions such as regulatory reporting engines, e-discovery, and analytics. 

AI, automation, and machine learning still has a way to go but using conversational AI alongside the human element has the potential to provide vital information that could not otherwise be accessed, or missed. The finance market now has access to previously unseen data that can be used as proof of evidence for regulators but also will have the intelligence to anticipate and avoid fraud.