How can Pictet’s Quest AI benefit portfolios

The limitations of traditional quantitative models

Traditional quantitative approaches tended to be re­stricted to analysing a relatively small number of market effects that result in temporary mispricings. This then provides exposure to broader market drivers, known as factors, such as value or momentum. AI offers the possi­bility of hundreds of potential signals/features at higher frequency. These are generated from data that includes company accounts, share prices, analyst notes, press re­ports, investor responses to new information over the short and longer term. And that’s just scratching the sur­face.

AI can generate greater insights into what’s driving stock prices

Unlike traditional machine learning, which identifies linear relationships within the datasets, AI can capture much more complex associations within the data pool, which allows it to have much greater insights into what’s driving stock prices. Being able to identify more complex, non-linear relationships boosts its ability many-fold to find associations between data series.

For example, in a traditional model, an analyst up­grade of a company would suggest its stock would out­perform. However, there are many reasons why such a re­lationship might not hold on a given occasion, or not be captured in time to generate alpha.

A non-linear machine learning model that is trained with historical data can identify the relationships that tell us when the analyst upgrade would most effectively forecast future outperformance. This could be because there’s a broad range of analyst forecasts, or the particu­lar forecast is an outlier or because of timing – say if the company is due to report results soon. If the stock is widely shorted by hedge funds, the model could identify that it is likely to be subject to a short squeeze resulting in a far more dramatic jump in share prices than the ana­lyst upgrade might otherwise warrant. There are poten­tially tens of thousands of these “conditioning” non-line­ar relationships across traditional financial data sets that can generate additional alpha.

AI helps to deliver factor-neutral returns

This much more intricate framework opens the way for the portfolio managers behind the model to isolate the stock-specific effects that influence the stock price. To do so, they strip out a multitude of common factors (market, sector, region, industry, country, styles, econom­ic exposures) from each stock’s performance. In doing so, they can identify and extract pure company-related al­pha.

Over time, the algorithms evolve, understanding changing economic and market dynamics and incorpo­rating new data series.

To be effective, AI-driven models need humans to set investment parameters. But once those are specified, the trained algorithm makes the buy and sell calls on individual stocks.

The power of incremental alpha over time

Because these parameters limit risk, they will also limit the amount of alpha the strategy generates. However, the compounding effect even of incremental alpha is powerful over time. And given that expected returns on equities are forecast to fall to mid-single digits amid high valuations following several banner years, even one to two percentage points of alpha will make a difference. For example, if we assume a hypothetical market return over the next 10 years of 5% and an assumed outperformance of 1.5 percentage points per year net of fees, the wonder of compounding will lead to an additional 24.8% of return for the client over the decade.

How can Pictet’s Quest AI benefit portfolios?

Quest AI harnesses the scale and efficiency of artifi­cial intelligence for stock selection in an enhanced index product with low tracking error.

It was created with a view to produc­ing a strategy that has the same profile and level of riski­ness as the wider market but that produces an addition­al return potential.

But unlike actively managed strategies – which also aim to outperform the market – the fact that our AI ap­proach requires relatively fewer humans means that it has lower management costs. Typically our approach to AI costs little more than a passive strategy notwith­standing the expertise needed to build and maintain the model.

What makes our approach to AI different?

Our proprietary AI models were developed over many years by our experts, who are typically PhDs in physics and mathematics. The models have been trained with some 400 characteristics from multiple data series, in­corporating periods of around 15 years and then repeat­edly tested under different economic backdrops. The rig­orous process is designed not only to maximise forecasting power but also to ensure that the models aren’t overfitted – that’s to say that they don’t only pro­duce accurate results under a narrow set of circumstanc­es but that they work under changing economic environ­ments.