Video transcript
The way that investors are betting on AI is by buying AI stock. But you at Pictet are doing something completely different. You're using AI to decide which stocks to buy. So this is investing with rather than in AI.
And that outperformance is driven not by a model that is developed and tested by a portfolio manager, but a model that is trained on hundreds of thousands of different data points over many, many years to allow us to take the active positions.
It seems to be all anyone is talking about nowadays, especially when it comes to the investing world. AI has arguably been the biggest theme over the past year or two years, but the way that investors are betting on AI is by buying AI stocks. Traditionally, semiconductors, Nvidia, things like that.
But you at Pictet are doing something completely different. You're using AI to decide which stocks to buy. Can you tell us a little bit about how you have an ETF that uses that strategy?
Yeah, you're exactly right. So this is investing with rather than in AI. So PQUS, it's the Pictet AI Enhanced US Equity Fund. It blends the best of passive and active, index-like risk, and index-like performance profile with compounding outperformance on top. And that outperformance is driven not by a model that is developed and tested by a portfolio manager, but a model that is trained on hundreds and hundreds of thousands of different data points over many, many years to allow us to take the active positions.
Interesting. Now, when investors hear the word "model," they think about factors or quantitative investing more broadly. Yeah. How is this ETF different than, say, a factor fund? So we're explicitly not trying to take any factor risk versus its benchmark. And we're able to do that because we do three things. So firstly, when we train the model, we train to forecast just the residual part of return, the piece that does not have any element of factor-driven performance within it, whether that's styles, whether that's industry performance, whether that's economic exposures.
Now, the risk sometimes is that the AI will still learn something like momentum. It's really hard to kill the momentum trade. But then when we build the portfolio, we make sure that it's tightly constrained versus the benchmark on dimensions like momentum and value. So it will be uncorrelated. It will be different to traditional factor-driven performance. Got you.
Now, when people hear AI, they think about the chatbots that they're used to using every day, ChatGPT, Claude. Yeah. What kind of AI is this fund using? Is it large language models or something else?
So again, it's explicitly not a large language model. So language models, they do have usability in quant investing, but generally, it's going to be to capture maybe sentiment from certain data sources that traditional approaches would struggle with. What we're using is a tree-based approach, thousands and thousands of decision trees.
Now, how do they help us better than a large language model? Well, the model is trained explicitly to do one thing, and that's forecast the next 20-day residual returns. A language model is more of a generalized approach. It's not built to do specifically what we need it for. So we train on 15 years of data, thousands and thousands of decision trees. They produce great forecasts. They're stable, they're interpretable, and they don't take as much compute power as language models would take to train them.
Got you. Now, when people think about AI models, they're constantly being updated and improved. How often is your model being changed and improved?
So firstly, every three months, we retrain the model. So we use a 15-year window that we train on. Every three months, we roll that by three. So we drop off the oldest period. We add on three new months. It allows it to learn some of the new evolving dynamics within the market, the relationships, and hidden patterns that a traditional quant model would struggle to fully capture. Within that retraining as well, we're constantly adding new features, new characteristics about companies. So for example, when we first launched this approach in Europe a couple of years ago, we used 200 different ways of assessing a company. We now use 400 ways. So yes, we both retrain the model regularly, and we evolve the types of data that we're using as well.
And what kind of performance can investors expect with this type of fund? Is it going to be swinging for the fences, or is it going to be closer to the broad market?
No. Again, this is something that is trying to give the best of passive and active. So we want to give a beta one to the market. We want to have a similar risk profile to the index, and then we want to clip away 1% to 2% above the benchmark. If I look at what we've done in similar strategies in other vehicles, we've actually been delivering somewhere between 2% and 2.5% above the benchmark. But again, we'll do that on a very consistent basis. Now, obviously, a really interesting strategy.
What type of investor should consider adding this to their portfolio?
So investors that are maybe thinking about how they can evolve the course of their portfolio. So we know a lot of advisors are using passive at their core. Now, that's worked for them pretty well over the last 10 to 15 years. But as equity returns likely reduce in the coming years, if they are looking to try and get a higher return from their core, then adding something like PQUS to provide a core holding and deliver some alpha on top within that core piece, that's the type of investor that should be using this.
Fantastic. PQUS, we're going to be keeping a close eye on it. David, thanks so much for your time. Great. Thank you very much.