Since the emergence of “Modern Portfolio Theory” and the “Capital Asset Pricing Model” in the late 1960s, institutional investors have taken a quantitatively driven approach to portfolio construction, looking to create portfolio diversification and obtain better risk-adjusted returns by balancing their asset-class exposures. This journey has seen several important advancements in thinking about how to optimally achieve desired results.
Funds that want to take the total portfolio approach first need to get the total portfolio view. To do that they not only need data – and lots of it – but a rock-solid understanding of exactly how they’re going to use it.
How Citi wants to use AI to put ‘the cape on Superman’
AI means a lot of different things to different people. To Joe Bonanno, global head of data, digital and innovation at Citi Securities Services, it means the business of asset servicing will be more efficient and effective – capable of “failing faster and then trying something new”.
“I’ve been doing data analytics for 25 years and I honestly haven’t seen anything this explosive since the birth of the internet,” Bonanno tells ISN. “It’s accessible now; ChatGPT made it more accessible, and people can play with it… I don’t look at it as a replacement for people or processes; I look at it as a way of enhancing them and making them go faster. It’s like the cape on Superman or Superwoman.”
To that end, Citi is exploring use cases across a number of different facets of AI, including traditional data science, machine learning, and large language models. The latter is best at replacing “the maker, not the checker”, and their propensity to “hallucinate” makes them unsuitable for making recommendations.
“But you get RFPs, and in different parts of the world we get different questions; do you do custody in Canada? What’s your SLA for this? What’s your disaster resolution recovery program? What do you charge? Today, a person has to take that list of questions and farm it out to multiple people through the organisation – to ops, to sales, etc – then they have to wait for a response.”
“Sometimes the response is good, sometimes the response is incoherent. Sometimes the response is long, sometimes the person is on vacation.”
Trying to solve that problem, Citi turned to the open-source Large Language Model Meta AI (LLaMA) which, with little training, was able to answer 60 per cent of the 600 questions handed to it based off existing documents in the Citi database.
“That’s where the biggest ROI is at the moment, because you can do this on a creaky infrastructure,” Bonanno says. “Firms think they need to get their data right before they can use AI, and that’s just not the reality.”
Around 70 per cent of Citi’s use cases for AI are focused on enhancing productivity in the bank’s internal operations, but there are also opportunities to highlight “anomalies and patterns” in client data.
“I can tell you trades are going to fail because the historical counterparty behaves this way at this time of the year; I can tell you that you have heavy concentrated positions in Russia real time; I can tell you that if you trade these three securities out in this country you’ll be ESG compliant,” Bonanno says.
“You just have to have people to write algorithms for some of these things, and if you set your parameters right it’s pretty easy to do; a lot of this code, quite frankly, is open source on YouTube. Your underlying data is the differentiator.”
Bonanno also oversees Citi’s investments in fintechs and startups, a roster that has grown to include Grow Super in Australia, global proxy voting and shareholder disclosure service Proxymity, and BondbloX, a Singapore-based digital bond exchange. Those investments – and spending plenty of time in the world they come from – also help Citi come up with solutions to its own “inventory of gaps and challenges”.
“It’s important for me to know what’s going on,” Bonanno says. “I need that to get that pulse… I don’t really look at banks as my competition. I look at the Apples and Microsofts and Googles and Amazons.”
