Remember me

Register  |   Lost password?

The Trading Mesh

Round table report - Embedding AI & Machine Learning in Financial Services

Mon, 10 Jul 2017 09:04:25 GMT           

Artificial intelligence (AI) is already embedded in tools, including anti-fraud systems, chatbots and security software. At some point in the future it will become so commonplace it will fade into the background, according to attendees at the roundtable ‘Naturally Intelligent’ held by Realization Group and Software AG on 7 June 2017. However, with regard to financial services, many impediments to operationalising AI must be addressed before ubiquity can be achieved.

Firstly, firms must remain realistic. Possessing an AI capability does not unpack complex problems on its own. A starting point will be to actively engage in deployment of the technology, identifying appropriate use cases and building out from them, particularly in capital markets. That can be achieved by augmenting the statistical models that are already in use with more contemporary machine learning technologies to make them even stronger as decision-making tools.

This augmentation has gradually become easier. Whereas much of the mathematical and statistical basis of machine learning techniques were established as long as 20-30 years ago, the tools to access and apply this knowledge have until quite recently been sparse and limited. Great progress has been made in this regard over the last 5 or so years, with a number of mainstream tools becoming available both from the open source community and commercial vendors. 

 

Early growth and restrictions

The democratising of the tools and technology that underpins AI has allowed a critical mass of people to develop these tools and make them genuinely useful for financial services firms. 

It also serves as a draw for enterprises engaged in machine learning, to attract and retain the best people, as they often want to contribute to open source projects. However, for a some enterprises to contribute to an open source project is challenging, due to contractual stipulations on intellectual property development for employees. That can lend support to working with third parties.

The counterpoint to the easy accessibility of systems is the risk of ‘legacy spaghetti’, with no strategic oversight or plan for the interaction between a numerous combinations of development tools, systems and data connectivity. Data is fundamental to the successful deployment of AI, as systems need to be trained to run on specific data sets. For that reason, a centralised management of data is crucial to enabling trust in AI systems, through quality assurance and maintenance. 

Cloud environments can be very useful in providing the elasticity to handle very large data sets and provision of compute resources. Yet it too has challenges and some enterprises have had to resort to building private cloud instances within public cloud offerings in order to overcome concerns about data security. The risk of falling foul of regulators makes firms err on the side of caution.

Regulation, specifically transparency, can also be a barrier to adoption of AI in certain areas of the business. Academics have noted unease when looking at deep learning techniques because even experienced researchers could not explain how the subsequent decision-making process worked. If a system were to suffer from reduced accuracy, identifying that change without full transparency over the processing could be difficult.

 

Clear benefits

While there are challenges that firms need to overcome, the potential upsides of AI are enormous, picking up a whole new ranges of activity that were never conceived before.

When new technologies are applied, flaws are quickly highlighted, for example with self-driving cars. However, the statistical comparison with human drivers is not made. There will be a tipping point where AI will be acknowledged as less flawed than humans. 

Talent, buy-in and culture must all be addressed to support successful adoption and embedding of AI within a business. Data scientists are needed to manage the data, and figure out what is not useful – often assisted by machine learning – early on in the process, so the right data can be collected. 

Such people are highly sought after, and firms will need to work out how to create an environment that attracts data scientists, or find partners with that talent pool. Culture is an important part of that. Certain methodologies, such as ‘fail fast’ are alien to the culture in some large enterprises making it hard to kill the ideas that are not going to work. 

There is no way to get projects into production without going through governance, which can be very bureaucratic and the cost cutting culture over the last 15 years has made it harder to support IT development. Moreover, delivering marginal gains that are valuable and valid will often not be seen as enough to get support for a project. In contrast, projects that are unable to succeed due to their ambition will be cleared for approval. 

When setting up an AI project, keeping one’s feet on the ground has to be balanced with the excitement needed for getting funding and the reality of delivery.

Regulatory challenges can be overcome to some extent by engaging with the regulators who offer a sandbox facility, although these tend to be retail focused and regulators can find it challenging to answer questions. 

 

Planning and preparation

A lot of the future AI world is still far away and firms need to plan a route to get there, with awareness of the current challenges, and a strategy for addressing them. Simple example AI applications, such as password resets, are good early stage developments that develop expertise, make challenges apparent and allow solutions to be found. The underlying architecture needs to be put in place at this stage, so that analytics can be run across data sets and data can be made accessible. 

Statistical analysis that has been conducted in the past should be used in operational environments to improve the development of artificial intelligence and where possible teams ought to have AI capabilities in their skill sets so opportunities are not missed. Despite the temptation for ‘moonshot’ projects, a well-developed case for more prosaic technology applications at lower levels of sophistication can build up comfort within senior management for its use. 

Notably, AI has already arrived and yet often not been noticed, suggesting its value can be realised without controversy.