June 30, 2025 (NO COMMENTS)

Towards the end of last year, Vontobel’s quantitative investment arm built an AI stock-picking assistant to help its portfolio managers improve on their personal investment approach.

The Swiss asset manager encoded the investment style of one of its top portfolio managers, and constructed a model to recommend stocks in the same vein.

“Portfolio managers have a style and tend to be fairly consistent,” says Andrea Gentilini, who heads the quantitative investments boutique. But the trap they can fall into is to look “at the same things over and over again”.

AI models don’t have this shortcoming. “Algorithms tend to spit out stocks that a PM would not naturally look for,” says Gentilini.

The PM in the experiment has since added more than 30 stock recommendations to their portfolio, with a hit ratio of roughly 60% to 70%, Gentilini says. Vontobel plans to roll out its tool to other portfolio managers before year end.

“Each portfolio manager has an investment universe of maybe thousands of securities. You can’t have each of them cover the whole target. So, the first screen of investments is very important. And that’s where AI is making an impact.”

The stock-picking bot is one of several experiments Gentilini’s division has undertaken, as it seeks to augment its traditional work with AI.

The unit, which runs both purely systematic and hybrid funds, has created an AI-boosted version of quality factor investing. It is launching its first foreign exchange strategy, based on a model that generates most of its outperformance from knowing when not to trade. Two new specialist teams are using AI, in forecasting and to identify trends in thematic investing.

Better quality

The quantitative investments boutique, which manages more than $20 billion for a mix of institutional clients, hired its first AI specialists five years ago and paired them initially with portfolio managers to scope out areas for research. That original team now manages its own fund, which returned nearly 14% over its benchmark in its first year of live trading.

In mid-2023, the firm set out to explore how AI might improve on elements of factor investing. “Quality investing is pervasive in the way we think about investing,” Gentilini says. Roughly two-fifths of the firm’s investment teams use quality as a core element of their investment style, though all the PMs use it in some form, he says. “Our first foray was to see whether that DNA could be made better.”

Vontobel’s research found about a quarter of stocks in the top and bottom quintiles by quality rose or fell from that quintile each year. Investing in rising-quality stocks could yield up to four times the return of buying quality stocks that stayed stable, Gentilini says.

Vontobel built an AI model to forecast which stocks the risers would be. The AI improved the ability to forecast stock moves between quintiles in 18 of 25 possible transitions. “We now give this to our portfolio managers that invest using a quality style to help them do their job better.” Gentilini says.

Noisy signals

Quant investors aim to parse the signal from the noise. Vontobel’s quants think the noise can work as a kind of signal, too.

The firm’s new FX strategy, which feeds on about 200 types of macro, fundamental and market data, and trades combinations of up to 15 currency pairs, takes positions in line with the strength of its signals versus the noisiness of the noise.

The model borrows from long- and short-term machine learning models and other temporal machine learning models to correctly interpret history, Gentilini says. “Half of the time, the machine says to do nothing.” In backtests, the model’s decisions to not trade increased the Sharpe ratio by about 0.2.

In FX and more broadly, the firm pays close attention to the stability of the signals its AI models generate.

Through iterative retraining of the models – for some weekly, for others monthly – the firm can observe how far the AI achieves consistency of output, Gentilini says.

“We’re looking for robust outperformance as opposed to stellar shots. We don’t need to be right 60% of the time. Fifty-three or fifty-four per cent is enough. The most important thing is that we don’t see large swings in the model. We go for good Sharpe ratios, good information ratios, solid hit rates and win/loss ratios – not unreasonably high ones because we know those are unsustainable. We have noticed that lower values – but still good values – tend to persist more. We go for those.”

Hit ratios measure how often a portfolio manager is right and wrong. Win/loss ratios measure how far a PM is right when they’re right, and how wrong when they’re wrong.

Surprises

Not every experiment the Vontobel team has tried has worked. Tests of enhanced versions of other factors besides quality showed that AI struggled to make predictions that improved on a vanilla approach.

Value investing, for example, depends on stock prices reverting, Gentilini says. Such periods, which are infrequent, prove hard to isolate amid more protracted periods of rising prices. There are fewer chances for the AI to be right.

Some tests have thrown up perhaps unexpected results.

In FX, the firm’s research suggested that despite the size, liquidity and supposed efficiency of FX markets, its AI models could find opportunities to trade.

The macroeconomic environment has become more favourable for FX, Vontobel’s researchers believe. “Interest rates are away from the zero boundary, driven by the return of monetary policy cycles as opposed to the one-way street of monetary easing,” Gentilini says. “We also see deglobalisation, which means that, by definition, currencies are going to move away from each other.”

As for how these experiments are put into practice: in the firm’s hybrid investment funds, where PMs express a view but draw on AI-generated recommendations, an awareness of what AI models can and cannot do has proven crucial, Gentilini says. When Russia invaded Ukraine in 2022, for example, the firm’s portfolio managers knew to override its AI models because the models had no notion from the data that an invasion was imminent.

Here, breaking models into component parts can help. Plane autopilots are often split in this way, Gentilini says, one module to ascend, another to descend, and so on. The fragmented architecture makes it easier for pilots to understand – and trust – each part of the system.

Vontobel has drawn inspiration from the idea, he says, as it has from other industries that have made big technological advances. “Being a pilot was the most dangerous job on Earth back in the 1930s. By experimenting thoughtfully with technology and machine-human interaction, flying became the safest method of transportation.”