Fighting Financial Crime with GPT

The early 2010s saw a global crackdown on money laundering in the wake of several high-profile scandals worldwide. This included the threat of significant fines by international regulatory bodies and even the potential of sanctioning repeat offender banks by removing their banking license. According to the UN, around $800 billion to $2 trillion are laundered globally annually, or 2% - 5% of the world’s GDP.

Money laundering is the process of turning money earned from illicit activities, such as drug dealing, arms trafficking and corruption, into money that appears legitimate. Money laundering is a persistent challenge as criminals consistently evolve in how they conceal their identity, mask the trustworthy source of their wealth and funnel money into legal businesses through inconspicuous channels. After the money laundering incident, financial institutions strengthened their anti-money laundering processes, and regulators implemented stricter legislation.

It is now common for a bank to perform thorough financial crime (FinCrime) investigations into a significant portion of their current and prospective customers. This process involves analysing their transactions, cash deposits, sources of wealth, business activity and customers’ business partners. This was not common ten years ago when the US Department of Justice started going after money laundering. Technology has been a massive driver of this behavioural change. Data and Artificial Intelligence advances have made it possible to automate swathes of the analytical activities experienced professionals previously performed. This article will focus on fighting money laundering using Natural language Processing (NLP) and Large Language Models, such as GPT-3 (ChatGPT’s driving technology).

Harnessing the power of Natural Language Processing

NLP is the branch of AI that analyses sentences to make predictions. For example, if a sentence is describing a giraffe, NLP would be able to predict that the subject of that sentence is a giraffe. In the real world, NLP tools process customer feedback, identify complaints, and use their insights to deliver better customer service.

In FinCrime, we can use NLP to predict the business nature of a company using information from the internet. A company’s nature of business is essential because a company’s business line enables us to contextualise its activity. For example, a casino handles a lot of cash and large sums of money. However, no checks are involved in tracing where the money has come from, placing a casino at high risk of being involved in money laundering. Another scenario is having large cash deposits of £100,000 weekly at a fish and chip shop. Even though takeaways are a cash-intensive business, £100,000 weekly revenue through legal means is highly unlikely.

Public information on what a business does for a living, such as the UK company register (Companies House), does exist. However, it is often incomplete or inaccurate, as the business completes this information and is not independently verified. NLP can plug this information gap.

NLP: anti money laundering weapon for the future?

KPMG has developed the technology to predict the business nature of enterprises. The solution works by taking in the company name and location and then pulling information from the company’s website. This data is then passed through an NLP model, which predicts the company's actions. It returns the likeliest SIC codes (UK system of 5-digit business nature classification codes) by comparing them against a reference list of SIC code descriptions.

KPMG’s solution was able to uncover a few unexpected results. It identified a company that claimed to be processing metal sheets when, in fact, it was an aerospace engine manufacturing and repair company that was active in the defence industry. Defence is a high-risk sector because of its record in busting sanctions, involvement of Politically Exposed Persons (PEP), and a track record of corruption and bribery.

Another exciting find was an LED signs manufacturer that builds and installs non-illuminated road signs. A closer look revealed that the manufacturer also produced electronic equipment that was ‘dual use’ — components with both civilian and military applications. Companies exporting dual use technology are prime vectors for sanctions-busting activity.

The solution's effectiveness stems from its ability to identify additional SIC codes, potentially high-risk ones, that were overlooked by Companies House or other data sources.

How GPT changed our approach

GPT belongs to the family of Large Language Models; KPMG’s solution uses an NLP bi-encoder drawn from the Hugging Face Transformers family. Both these technologies are based on Transformers NLP technology; the T in GPT stands for ‘Transformers’.

A Transformer is a Machine Learning module that converts text into vectors. This, alongside the ‘attention mechanism’, is the core component of the current generation of NLP mechanics. Transformers can be deployed as ‘encoders’ — where the word vectors are compared to a library of terms to find the closest similarities — or ‘decoders’, where the input vectors are used as a basis to generate new text. IBC’s technology uses encoders; GPT uses decoders.

For you

Be part of something bigger, join BCS, The Chartered Institute for IT.

We examined how effective GPT-3 Davinci was at predicting what a business does for a living. Information from the company websites obtained by our tool was fed into GTP-3 to assess how accurate GPT was at predicting their SIC codes compared to the KPMG tool.

At the broad classification level, the KPMG Transformers driven solution, IBC, and GPT-3 DaVinci were both able to predict an industrial division (e.g. ‘Civil Engineering’ or ‘Food and beverage service activities’) with 61% and 63% accuracy, respectively. (For context, there are 86 industrial divisions.)

However, when we drilled down to the detailed business classification level, IBC could pick a correct business classification (i.e. a 5-digit SIC code) for a business 52% of the time. GPT, on the other hand, was only successful 33% of the time. (For context, there are over 750 business classifications — the ‘random’ result would have a 0.14% chance of being right.) IBC was over 50% more accurate than GPT DaVinci at business classification.

However, chaining the two solutions gave the best overall outcomes. We first use GPT to predict the high level sector and use the outcome to narrow the possibilities; we then use that output to drive more tailored predictions from KPMG’s solution. This ensemble solution was able to get the proper business classification (5 digit SIC) 63% of the time (a 20% improvement on our previous best), and the industrial division (the 2 digit SIC code) was accurate 79% of the time (a 12% performance improvement).

The natural endpoint for the combined solution is to deploy our ensemble in a Know Your Customer (KYC) support role in the fight against FinCrime. This model is particularly applicable to customer journey points such as onboarding, transaction counterparty analysis, and periodic customer reviews, where it can bring to light additional context that can sway a risk decision in the right direction.

Following the events of the 2010s, 13 years later we now have the technology and computing power to fight against FinCrime. By employing NLP technologies, we can extract the necessary information from a sea of online data noise, using publicly available information to quantify risk accurately. The work above illustrates that when using machine learning to quantify risk, ensemble models —combinations of algorithms — outperform individual star performers in getting to the truth with precision.