Paperjam attended a presentation on Pictet’s Quest AI fund by David Wright, head of quantitative investments at Pictet Asset Management on 6 May 2025.   Photo: Pictet Asset Management

Paperjam attended a presentation on Pictet’s Quest AI fund by David Wright, head of quantitative investments at Pictet Asset Management on 6 May 2025.   Photo: Pictet Asset Management

Pictet's AI-powered investment strategy uses machine learning to anticipate short-term market movements. This strategy aims to generate better performance than a benchmark index without straying too far from it. At every stage--from model creation to daily portfolio management--human oversight ensures the rigor and relevance of the AI's decisions. Dan Wright, head of quantitative investments, explains.

“AI is not really a revolution. It’s part of a broader revolution in what has always been an evolving demand for more data, more technology and faster processes, so it really does fit into the broader themes of the way the quant industry has worked,” said David Wright, head of quantitative investments at Pictet Asset Management on 6 May 2025. Starting in early 2010, he explained, machine learning, a subset of the artificial intelligence field, was one of the areas that Pictet started to look at alongside alternative data and natural language processing.

“One benefit of quant is that you can remove the potential risks of that bias and subjectivity… once you’ve codified processes, you can make very fast decisions in a disciplined way,” he stated.

An AI fund ticking all the desired KPIs for a fund selector for now

Wright explained that Pictet’s Quest AI fund is an AI driven global equity investment approach that has been structured as a scalable “enhanced index strategy”, designed to provide index-like risk characteristics while aiming for active outperformance.

“Machine learning… allows us to make more accurate short term forecasts… for stock selection

David Wright, head of quantitative investments, Pictet Asset Management

Since its inception in March 2024, the fund has gathered over $1bn in assets under management, and Pictet’s marketing material reports that it delivered, net of fees, an outperformance of 3.21% as of April 2025 compared to its benchmark, the MSCI World. This outperformance has been achieved while maintaining a low tracking error (currently 1%, against a target of up to 2%). The asset manager targets an outperformance of 1.5% per annum, gross of fees, which range between 0.90% and 1.60% plus performance fees.

Wright claimed that the fund will “always have the risk characteristics of the benchmark” as it holds 400 stocks in the portfolio while “only tilting slightly away” from the benchmark weights (country, sector and stock) and “the style exposures (momentum, value, quality)… are exactly the same as the benchmark.”

Looking under the hood

“Machine learning… allows us to make more accurate short-term forecasts… for stock selection,” affirmed Wright. He argued that ML enables Pictet to process a huge amount of data and use many more data sources, signals or features from companies under scrutiny. He added that it allows Pictet to use higher frequency information to understand the complex relationships between different data sources, such as “the interdependencies that are not easy to identify, even with traditional models,” and test the model over decades.

Pictet’s AI stock forecasting model runs daily and forecasts the 20-day (one month) forward specific return for stocks in its broad investable universe (around 1,600 companies). Wright commented that the stock forecasting model utilises 300 features as input information, derived from sources like sell-side forecasts, market prices/volumes, investor positioning data (mutual funds, short interest) and calendar effects from “many data sources.” It’s much more about the supply and demand in the marketplace than focusing on company fundamentals. “Who owns a stock? Who’s buying it? Who’s selling it? Who’s shorting it? How will that change, and how is that going to evolve over the coming months?” As a sign of its time, Wright noted that the model has “an element of ESG integration within the strategy.”

Interestingly, Wright stressed that the goal is “to ignore any of the common drivers of return,” such as a company’s exposure to the industry, its country of residence, the country it sells to, its economic exposures and the style (momentum or value). These elements can be obtained from factor models.

He explained that the model is then trained on data going back to the 1980s with the live version based on the last 12 years of training, which is retrained quarterly. The primary machine learning technique used is “gradient boosting,” built upon sequential decision trees. This involves training thousands of trees, each learning from the errors of previous ones, combined for accuracy. Retraining the model on the Amazon Web Services space cost Pictet €20,000 to €30,000, according to Wright.

Once the stock forecast and the ESG scores are established, Pictet constructs a diversified portfolio “every single week” by applying “tight constraints on portfolio construction versus the benchmark.”

Pictet “penalises certain style risks and turnover to make sure that we’re not trading too much and spending too much on trading costs.” Yet the turnover is 150% to 200% per year, according to Wright. Then the “final portfolio is reviewed by PMs for any potential late-breaking news before we go and trade.”

Humans are not that far after all

Wright specified that the strategy’s development process took about four years before launch, emphasising human-led work at every stage, from selecting techniques (like testing gradient boosting vs. neural networks) to choosing data and refining the model. Pictet’s preference is for traditional financial data due to its long history and breadth, which supports robust model training, although the asset manager is exploring the use of alternative data such as internet search or social media activity.

Despite claims of minimising the human factor, the latter continues to play an important role when deploying the machine learning quantitative strategy. Wright emphasised that its dedicated portfolio managers are responsible for unfolding the strategy daily. The same people built and trained the model.

“It gives us an additional element of risk management and oversight and ownership that you need to have when you’re running client money.” Wright explained that the portfolio managers review the output of the model and question, for instance, whether the distribution of the forecast is sensible compared to what is normally expected or if the largest changes are of the expected magnitude. “We will only trade on the model once we’re comfortable with it… every human intervention is about risk mitigation rather than trying to add additional alpha.”

Unveiling the black box: “We’re not flying blind in running the strategy”

“Before we ever built the model, we were already focused on how we would be able to crack open that model and understand the positioning that came out of it, and therefore the position drivers in the strategy,” stated Wright. He explained the contribution of the various factors for an overweight/underweight position in a stock.

The performance attribution tools enable Pictet to specify whether the contribution comes, for instance, from longer-term sell-side views, short interest price momentum, fundamental changes to a company, etc. He thinks that various relationships between that different data are “only capturable from the machine learning development of the approach.”