Replacing current market risk models with something that has been derived through artificial intelligence (AI), especially unsupervised machine learning, “we are a long way from seeing that,” says Daniel Percy-Hughes, Consultant at Synechron.
“There are a couple of reasons for that. One is that they [AI and machine learning] while you can test your model and you know what your input data is, and in the case of risk managers they definitely know what their output data should look like, an unsupervised machine learning algorithm doesn’t offer that very transparent explanation as to how it has derived the outcome, and the model that is used for outcome.
“I don’t think risk managers currently would be comfortable working on that basis, and I don’t think that regulators would approve that type of model,” says Percy-Hughes.
“The amount of data you would need to train such a model would be way more than the ten years that banks, or market risk managers tend to use in their back testing. So, you’ve possibly got limits there on who might be able to adopt that type of technology were they to look at the biggest or most well-funded organisation that have access to that depth of history of data,” says Percy-Hughes.
There is a reluctance to adopt that technology until it is tried and tested, according to Stuart Brock, director of Seal Software, but there are clear advantages in the future for those that do adopt the technology now.
“You can’t try it until someone actually uses it, so I think there is a hesitation to be the first person to use it to do that.
“It’s also that complication piece, that bringing in something that is very complicated that has taken people years to develop. But if you bring in the AI now and you start using it for the parts that it can support, and then you start teaching it for the other parts, over time that AI can do the same thing that is being developed today.
“And more importantly as those requirements change year to year, as the committees come out with the different requirements then you see that AI can be adopted more readily than the current system of people,” says Brock.
With the sheer volume of regulation expected in the next couple of years, the possibility of banks being able to invest the time and money into adoption of AI in market risk calculations is unlikely, says Percy-Hughes.
“There is also Basel III coming up, there is also changes to benchmark regulations across the EU, and potential replacements for Libor, all that sort of falling together in and around 2019 to 2020 into 2022 for market risk methodology.
“You’ve got a very big concentration of very high-profile regulatory change that will have very real impacts on banks. The market risk methodology will change how much capital banks have to hold and how much they are charged for not meeting aspects of the regulation,” says Percy-Hughes.
“With that amount of regulation in the risk space, the likelihood of banks being able to invest heavily in research and development for something such as machine learning at the same time is probably quite remote,” says Percy-Hughes.
This is something that Mayra Rodríguez Valladares, managing principal of MRV Associates sympathises with.
“To be able to use these AI solutions you really need highly technical physics doctorate types, you need to be willing to invest the money, and not just in what I’ve mentioned but in auditors, and compliance people who have a high level of math. Often people forget the compliance and the auditors, you can’t just have the regular ones because they haven’t been trained in that level of math,” says Valladares.
Potential for AI in data analytics
In a discussion paper published in December, the French Prudential Supervision and Resolution Authority (ACPR) found that “the development of artificial intelligence in the banking and insurance sectors should be accompanied by practical reflection on the minimum criteria for governance and control of these new technologies.”
According to the discussion paper, the performance of AI is largely dependent on data quality and lack of bias in processing.
A first step in an evolutionary approach to AI adoption in market risk analysis is to enable banks to have consistent views of their data, according to Percy-Hughes.
“The change to the Fundamental Review of the Trading Book (FRTB) rules is asking for consistent use of what may currently be a very disparate data set. The number of risk factors that a bank might be looking to observe and the market data that supports them might be quite fragmented across an organisation. It gives banks an opportunity to how they use their data, and how they source even on a trade level, data that goes into those models,” says Percy-Hughes.
“A risk manager or trader is dependent on the quality of data that goes into their models, so I think as something that can identify anomalies or improve quality of data that using machine learning or the data science approach that allows anomaly, or outlier detection might be beneficial to them,” he says.
For Brock, there is a recognition by both regulators and the market that it is impossible to continue to delve into the levels and amounts of data at the current pace without some sort of automation.
“We just cannot sustain this curve that we are on,” says Brock.
“Let’s just say that AI in its current state cannot perform the calculations themselves. They certainly can use the platform to go on the data, because they have to pump the data into those algorithms and so they are getting it from somewhere, so AI can help to automate that and plug it into whatever algorithms or system of record they are using to calculate that,” he says.