Existing screening technology is unable to keep pace with the demand for instant payments and is fraught with challenges around high volumes of false positives.
To discuss solutions, Napier’s Chief Strategy Officer, Luca Primerano, and Chief Data Scientist, Dr Janet Bastiman, were invited to share their expertise in the Data Protection World Forum webinar: 'Instant payments need instant screening.'
You can watch the replay or read the summary of their key points below:
The biggest sanctions screening challenges
Luca began by explaining most financial institutions using legacy technology are facing two main screening challenges:
1. Regulatory compliance challenges:
- Increasing transactional volumes are increasing instant payments and the need to check transactional data in real-time.
- Opaque algorithms based on thresholds and legacy technology are not providing the explainability regulators need to operate effectively
- Inflexible rules and workflows make it difficult to reconfigure the system to keep pace with frequently changing sanction regimes.
2. Operational processing challenges
- High levels of false positives generated by common entities consume significant amounts of resources. As a result, valuable time is spent investigating these false matches rather than genuinely high-risk payments.
- Suboptimal workflows with significant handovers between departments create even more inefficiencies.
- Manual testing of scenarios makes it difficult to manage and adapt to different scenarios.
These screening challenges are compounded by the complexity of name screening. Luca provided several examples where it isn’t clear whether the name should generate a match against a sanctioned entity.
There are lots of challenges that the screening system must be sophisticated enough to deal with:
- Multiple types of error, such as phonetic similarity e.g. Luca vs Luka
- Missing words and names
- Multiple last names
- Partial name matches
- Name variations (e.g. Nick, Nicholas)
- Spelling errors
To add to these complexities, many criminals intentionally set out to manipulate screening systems. The screening system therefore needs to be sophisticated enough to limit the frequency of false positives under the pressure of facilitating instant payments, without discounting any true matches. This is why financial institutions need machine learning and artificial intelligence (AI).
The screening solution for instant payments
The basic requirement for any screening solution is to be able to search, match and act:
Search: search the database you have access to - is the originator of the payment in a criminal database or the target of a sanction?
Match: how similar are they to the criminal?
Act: should the payment be stopped and investigated? This decision should be aided by the support of your screening system.
Janet explained that, with this basic requirement in mind, any screening engine should meet all of the following requirements:
1. Risk appetite and risk sensitivity – The screening system must allow you to set the match level appropriately, and you need to be confident you’re getting all the hits you should be.
2. Rules and scenarios based on type of payments – You may need to make adjustments based on the type and value of payments, which can change risk appetite.
3. Search against thousands of names in sanction lists – There are thousands of names in sanctions lists that are updated daily. It’s important the system uses the latest lists.
4. Learn from user feedback – If an analyst decides a hit is a false positive then the system needs to learn from this.
5. Use all available data to discount or escalate – This will make matching quicker and more effective.
6. Trigger actions in workflow – At what point do you stop a payment? This will depend on the type of transactions and your rules. Ideally the system should be able to make decisions up until a particular point, but certain match levels may be held to go to further review.
7. Allow auditability – If a transaction is stopped/someone complains/you to need to escalate to an authority, the system must transparently explain why you got that match or why something was discounted.
Name similarity in practice
It’s important to apply multiple similarity algorithms and models to get an accurate match, including:
- Phonetic similarity – For example, Gemma Payne vs Gemma Pain.
- Cultural similarity – You need to understand what could or could not be more important, such as choosing to be known by second name rather than a first name in western countries.
- Words – Are any words missing?
- Order – How are you looking at the matches of the whole name, the individual parts of that string and the order of those parts?
By feeding in risk policies and information from historical reviews into screening AI, you can achieve a good determination of whether to discount or review with an easy to understand explanation. For example:
- 'John Taylor' vs 'Mark John Taylor'
The AI system would discount this match because of the dissimilar first name component.
- 'Bak Real Estate' vs 'Real Estate Bank'
The AI system would review this match. Bak could be a spelling error but there may also be an intentional movement of words to bypass the system.
- 'John Snow' vs 'Snow'
While most legacy systems would flag this match, the AI system would discount because data records show John Snow is a person and Snow is a vessel. The system has incorrectly matched different types of data.
- 'Paul Mirano' vs 'Iran'
The AI system would discount this match because the context of ‘iran’ in a person’s surname is not related to the sanctioned country. Most legacy systems would however flag this for review.
Napier’s screening solution reduces false positives and increases the effectiveness of screening
Napier’s AI-powered API can be used on top of existing screening solutions. It passes all hits through to artificial intelligence for further review, where obvious false positives can be discounted, and the more complex hits referred for human investigation. Importantly, Napier’s system will provide a suggestion, score and explanation to feed into the workflow.
Performance can also be optimised by tuning the system with your rules, risk appetite, models and feedback data. The feedback loop is important to ensure constant learning and therefore false positive and false negative reduction.
Benefits of Napier’s transaction screening system:
- Efficiently reviews potential hits – It takes a long time to manually review false positives, and this comes at an opportunity cost. AI provides full explainability, so you can be confident it is reviewing hits intelligently Reduces risk – The screening system gives you the time and resources to focus on the hits that really do need to be reviewed.
- Improves efficiency – Enhanced matching reduces false positives.
- Easy performance validation – Historical data can be tested using sandbox environments. Here you can use a challenger score or challenger model to validate the robustness of your decision when comparing existing models.
In the past screening challenges were seen as a compliance issue. Now with the increasing demand for instant payments, they are becoming customer problems that are having a negative impact on customer experience.
Artificial intelligence holds huge potential for overcoming these challenges by adding another layer of review. It provides additional confidence and focuses human resources onto the hits deemed to present the greatest risk.
That said, artificial intelligence is not a silver bullet; it is just one piece of the puzzle. You need to get your basics right. This starts with a thorough and accurate definition of your risk appetite and policies.