Banks Embrace the Use of Synthetic Data

Synthetic data is playing an increasingly important part in testing solutions, particularly as pressure mounts on banks to harness newer, more sophisticated technologies while complying with a raft of regulation.

Although auto-generated data is not a new tool for many banks, it has proven particularly useful in recent cases where firms look to adopt new capabilities, similar to those of Netflix, Amazon or Google, that require vast amounts of data.

According to Giuseppe Nuti, head of algorithmic trading at UBS investment bank, replicating functionality that can, for example, recommend investment products to clients and understand user preferences involves the processing of huge volumes of data, a capability that big tech firms have in orders of magnitude more than investment banks.

In these cases, the synthetic data is used to compensate for the lack of testing data required to validate the technology to a comparable standard as that of Silicon Valley applications.

“We are talking tens if not hundreds of millions of users actively buying stuff or watching movies, versus a few hundred clients for a bank,” Nuti says. “Even in the biggest of investment banks and for the most successful of its desks, we are talking 200 to 300 active clients. The statistical difference is substantial—hence the need for synthetic data.”

Synthetic datasets are often based on vast amounts of reproduced historical data that includes insights or patterns that have already been identified. It is used to validate algorithms and AI-driven models, but only tests against predictable outcomes or previously determined answers. Yet, Nuti says, the data is a crucial component in evaluating a broad set of functionality and sits at the core of UBS’s development process.

“It is a necessary step that you need to take,” he says. “It doesn’t guarantee a solution because the world may not behave the way you thought it would, but it certainly ensures that if it does, you are able to pick that up,”

As one example, synthetic data has been incorporated into the development and testing of UBS’s recommendations engine, which is used to suggest trades to its asset managers and hedge funds and to identify potential clients. Algorithms are trained to analyze user behavior and provide automated suggestions based on their activity. Synthetic data comes into play when the technologies are used to test patterns and offer analysis before using the algorithms on real client data.

Data Protection

In many cases, the reason for using synthetic data is tied to compliance and the need to avoid using client data for testing solutions. As financial institutions are under increasing pressure to comply with global data protection and privacy laws, such as the General Data Protection Regulation (GDPR), they are having to take specific measures to adhere to cross-border data sharing and prevent client data getting into the hands of unauthorized users.

“When you have large reams of corporate client information such as we do, we have an obligation to respect the client around cross-border data sharing and there are very strict controls around that,” a senior data executive at a tier 1 bank says. “Instead, we are exploring how we can use synthetic data. So we can generate artificial data such as credit payments or whatever it may be and then use that for development use cases.”

The bank is currently in the early stages of research and development on testing synthetic data compliance use cases. According to the executive, the adoption of cloud technology or cloud environments will prove very useful in generating and storing vast amounts of synthetic data in a more on-demand and efficient way in the future.

However, they add, while synthetic can be very effective in testing technologies, it is not applicable everywhere. They say that for unpredictable anomaly detection, for instance, synthetic data can create its own issues.

“It doesn’t work for all types of use cases,” the data executive says. “It does help us when we are looking at historical pattern analysis, time series analysis and that sort of thing. If you are trying to do anomaly detection you don’t necessarily want to artificially generate data [for that use case] because it is almost as if you are planting an anomaly.”