1. Stéphane Daul, how does your academic background in physics apply to your work as lead manager of Pictet Asset Management’s Quest AI-driven strategy?
I was awarded my PhD in theoretical physics in 1997. After publishing around 15 research articles in the field while in academia, I joined the financial industry in 2001, specialising in risk management for over ten years. I obtained the CFA in 2004 and joined Pictet AM in 2011, and now work in the quantitative equity strategies team.
During my career in finance, I’ve adapted all the techniques I’ve learned from my years in academia, as well as shifted my perspective. In theoretical physics, you need 99.99% accuracy, while in risk management you only need 95%. But in asset management, it’s really difficult to beat the odds, so you’re aiming for 55%. The interesting thing with active quant investing is that it allows you to take advantage of large numbers, meaning that it’s easier to beat the odds with a portfolio of 300 stocks than a more concentred one.
Now, with artificial intelligence and machine learning (AI/ML), this can be further expanded to even bigger numbers.
2. What is the genesis of AI/ML at Pictet Asset Management?
When I joined Pictet AM in 2011, I was one of the first to talk about statistical learning, the ancestor of machine learning. We tried some strategies that didn’t work at the time, but about four years ago things started to shifting. Advancements in computing power and the emergence of new techniques made the previously unattainable possible. For example, the development of LightGBM1 in 2017 allowed for faster training of decision trees. We’ve been researching and testing this approach – you can read about it in the recent publication of a work related to our new Quest AI-driven strategy.2
It’s not just about constructing a model that works, but about finding the correct features and implementing a strong discipline to avoid traps such as overfitting or bias errors.
The final objective is to deliver a signal of over- or under-performance, and this takes a robust platform, as well as a seamless collaboration between engineers, data scientists and portfolio managers within our investment team.
3. How do you use AI/ML in your strategy?
Among other methods, we’re using classification and regression trees, or more specifically gradient boosted trees,3 to predict stock returns.
Our model has been trained on decades of data, with about 250+ features such as stock fundamentals, analyst sentiments, or price and market activities.
This is backed by our belief that over short time horizons, investor behaviour is the main driver of stock returns. By enabling the combination of a wide range of features, while capturing relationships and interaction effects between data, AI/ML is a powerful tool to quickly and efficiently assess those changes in behaviour and thus make more accurate forecasts.
We also think that it’s essential to decompose our model to really understand its predictions. Which features are the most effective? Are there any patterns? We don’t want to create a black box with a prediction power that can disappear without us knowing.
What’s really interesting with this field is that it never ends. There’s always something more to discover, or even test. For example, we’re currently looking at using deep neural networks in our strategy, and we’re testing other features to see if we can extract signals. We’re also researching ways of implementing AI/ML in our optimiser, which is the tool we use to construct and appropriately size stocks in our portfolios.
