AI in Asset Allocation

Hi, welcome to another episode of JPMorgan's All Into Account podcast.

I'm your host, Thomas Salopek from Cross-Asset Systematic Strategy and Research. And today I'm joined by my colleague, Shizuka Suga, also from research on the QuantEquity team. Welcome to the market's podcast, Shizuka.

Thank you, Tom. We recently attended the Central Bank seminar. For listeners who weren't there, can you briefly describe what the event is and what audience it brings together?

Yeah, this is a great event in New York. We're honored to be a part of it. It's the 51st annual Central Bank Seminar, bringing together 64 participants from over 35 countries. And we're talking about central banks, multilateral development banks, official institutions representing over $5.6 trillion in reserves and hosted by the public sector team. And it's a fun thing for us to be invited year after year and talk about AI and asset allocation. So back to you, Shizuka. One of the things you covered at this event was how are central banks using AI today and what are their main constraints, whether it's data, governance, talent, security. Why don't you give us an update on that?

Sure. Central banks are using AI in three main areas. First, macro and financial analysis, especially now casting to track productivity and the economy in near real time. Second, payment system oversight, including anomaly and fraud detection, and identifying risks like bank failures, cyber events, or money laundering. And third, supervision and financial stability, that is, using AI to process large sets of documents for information extraction and translation at scale. And for example, classifying borrowers with higher than expected credit risk. Beyond this, AI creates at least three major risks for central banks. One is cyber, and we've all seen the recent headlines around Mythos and Project Glasswing. It really turned AI-driven cyber risk into a financial stability issue, not just a technical one, which now requires board-level attention and international coordination. Two is third-party dependency and algorithmic herding. With only a small number of credible LLM providers, institutions can end up with the same model behaviors, same model, same output, same hallucinations, and this may amplify volatility through herding, liquidity hoarding, and potentially elusive outcomes. And lastly, bias and discrimination and the broader credibility risk that comes with black box decision making. So Tom, let's bring this to asset allocation. How are you using AI to solve asset allocation problems?

Sure. So, you know, I'm much more of an asset allocation person than an AI person. But that being said, you know, I fully appreciate that this toolkit has really made our lives a lot better on the asset allocation side. But, you know, if you look at the things that are in common in asset allocation, you know, whether you're a researcher or portfolio manager or reserve manager at a Central Bank, the idea is that we're thinking about a lot of the same things, meaning that expected returns, how do you come up with them? What are the different drivers of expected returns? How do you come up with your risk inputs? How do you use the