AI-Driven investment: Inside Pictet's Quest AI strategy

Alexandra Nagy’s investment career owes more than a little to artificial intelligence. And Pictet Asset Management’s quantitative artificial intelligence strategy, QuestAI, owes more than a little to Alexandra Nagy.

Now an investment manager, Nagy was a graduate trainee in Pictet’s Group Risk function, in 2021, when she was seconded to the quantitative strategy team under Stéphane Daul, a senior investment manager. It was a natural match. The quant team was in the early stages of developing an AI investment model and Nagy was a physicist with an academic background in machine learning.

The model they were working on used AI to sift through millions of market-related data points to find ways of generating additional returns from a portfolio for a given level of risk. The team’s early work resulted in an academic paper that formed the theoretical cornerstone of the QuestAI strategy:  “Performance Attribution of Machine Learning Methods for Stock Returns Prediction,” which Nagy co-wrote with Daul and another Pictet Asset Management colleague, Thibault Jaisson.

The paper was crucial because the Pictet Asset Management quantitative strategy team understood early on that it wasn’t enough to create a black box machine learning model with given inputs that generated accurate forecasts of stock performance.

The team also knew that for AI to be applied successfully to an investment strategy, they needed to be able to identify where the forecast performance was coming from. In other words, they needed to know in detail how the model was generating its stock recommendations.

Which is where Nagy came in.

Her doctoral work was on speeding up machine learning training processes, making it possible to train ever more complex models – which had become essential at a time that Pictet Asset Management’s quantitative experts were building up the complexity of their own models.

Nagy's PhD thesis at the Ecole Polytechnique Federale de Lausanne (EPFL) was on machine learning and quantum physics, and she undertook it after working as a researcher at CERN, the largest particle physics laboratory in the world, based just outside of Geneva. At CERN, she developed large-scale simulations, handling vast datasets, and developing mathematical solutions for highly complex, uncertain systems.

But the work was purely theoretical. She was looking for real world applications, which led her to Pictet.

Her experience was perfect for the sort of research Pictet Asset Management’s quantitative team was doing: distilling reams of data into accurate forecasts of future stock performance.

“For instance, we use IBES data which captures analysts’ earnings forecasts and sentiment about companies,” she says. “It helps our models to identify subtle signals and trends that might not immediately be visible in price or volume data alone.”

The modelling approach the team eventually settled on is known as "boosted trees". This involves generating multiple decision trees – models that are structured like flow charts – and then identifying and improving the trees in areas where their predictive powers are weakest through repeated training.

“It’s like going to the optometrist,” Nagy says. “With each round of lenses, they try to sharpen the vision a little more until they get the right prescription.”

Although Nagy’s work with the quantitative team was important enough to earn her co-authorship of QuestAI’s foundational research paper, an opportunity at another investment firm’s quantitative research team then saw her leave Pictet.

“But I stayed in close contact with the team. I would have lunch with Stephane every few weeks,” she says. “I wanted to see how the model was developing – I felt it was a little bit my child.”

So when QuestAI went live the team asked her to come back. For Nagy, returning to Pictet was an obvious next step. From the start, the strategy has been a big success, raising USD1.7bn in assets over 18 months. Nagy explains why in simple terms: “The model’s results are broadly in line with what we anticipated.”

The performance has been clear, but the model is anything but simple. For one thing, it’s trained on roughly 400 of what the team call features – characteristics, drawn from numerous data series, such as earnings per share and price-to-book ratios, as well as analyst sentiment data.

The model captures both straightforward and more complex relationships between features and stock returns. Some of these are linear effects, where a feature has a direct proportional impact on predicted returns. Others are non-linear, where the effect is more complicated but can often still be captured by traditional models. The most complex are the interactions between features, where the combined influence of two or more features can only be identified with advanced AI techniques.

The team spends considerable effort breaking down the model’s predictions to estimate how much each of these sources contributes to the final forecast. By carefully attributing portions of the prediction to different effects, they can better interpret the model’s signals and have greater confidence in its recommendations.

But humans – intelligent and experienced humans – are still necessary for the continued development and refinement of the model. For instance, during the portfolio construction stage, managers have the option of adjusting positions if they believe that a company’s current situation is too specific to trust the model’s prediction – say during mergers and acquisitions. Then there's oversight of the model: validation by investment managers is needed before any of the model’s recommended trades are executed. And they are necessary to specifying a portfolio’s constraints where they exist, such as ensuring the strategy is factor neutral, sector neutral, and geographically neutral.

Development is rigorous and continuous.

“The model is retrained and updated every three months, while new developments are introduced every six to 10 months. We take a lot of time to understand how it works within the portfolio,” Nagy says. “It’s an exacting process.”

In a manner of speaking, Nagy’s investment management career developed alongside the QuestAI model. Both continue to evolve together, with AI increasingly establishing itself as a third way in investment, alongside active and passive management.