Amid the extreme volatility in today’s markets, quant investors are leveraging artificial intelligence ( AI ) to enhance decision-making at the margins, rather than fully replacing traditional techniques.

Despite the buzz around AI, Ramkumar Rasaratnam, chief investment officer, equity quant, at AXA Investment Managers ( AXA IM ), believes that the new technology could better serve its purpose in investing by focusing on safety, explainability, and incremental edge, rather than wholesale automation.

In an interview with The Asset, Rasaratnam notes that AXA IM combines the use of AI, such as machine learning ( ML ), neural networks, and natural language processing ( NLP ), with non-AI tools such as “optimization” to manage market volatility in portfolios.

For example, AXA IM’s quant team uses AI to identify stocks with greater short-term volatility or those with a risk of rare but severe losses. Such stocks, also known as “tail-risk” stocks, are identified by using a “neural network”, a type of ML model inspired by the human brain, which identifies complex patterns in data to improve investment decisions, especially in areas like stock selection, risk assessment, and price prediction.

Unlike traditional models, which combine signals linearly, neural networks allow asset managers to capture nonlinear relationships between these signals. such as dividend cut probability, default risk, and asset impairment.

“This model improved performance by 20% to 30% in backtests and flagged issues like the collapse of Silicon Valley Bank in Q1 2023 before it happened, helping us avoid exposure,” says Rasaratnam.

A change of tone matters

Another way AXA IM uses AI for managing market volatility is to enhance “sentiment analysis”, a method of using NLP and other techniques to gauge the overall attitude or emotional tone of investors towards a specific asset or the broader market.

In this case, asset managers use NLP to transcribe and sift through hours of recordings of shareholder meetings or telephone calls with analysts to get more information and insights that will add incremental alpha to traditional stock selection.

“This is especially useful when managing broad portfolios without direct company engagement. The precise language ( of company officials in a meeting ) can indicate confidence and correlate with positive future fundamentals. Shifts in tone or vocabulary from previous calls can signal underlying operational issues or reorganizations,” Rasaratnam says.

But while AI tools enhance stock selection, AXA IM also uses optimization techniques ( not neural networks ) for actual portfolio construction to maximize risk-adjusted returns. Optimization techniques are employed to adjust a model's parameters to minimize a loss function and improve its performance.

“We use an optimizer to build our portfolios. An optimizer is a way of solving a non-linear problem. And for us, we've been through a process of improving the calculation engine of the optimizer and improving that aspect of it, but we feel that that's still the most appropriate way of actually building a portfolio,” he says.

Stock selection process

Since its founding 40 years ago, AXA IM has been using an optimization process to maximize the risk-adjusted return of a portfolio. But now it uses AI tools – neural networks and NLP – to enhance stock selection, and then uses an optimizer to build the portfolio based on the stock selection generated by AI.

“What we're trying to do when we build a portfolio is create our stock selection view, where we use neural networks and natural language processing. Fundamentals are the main thing that we look at to build our return expectation for a stock.  AI helps forecast returns and risks, and this feeds into the optimizer, which balances exposure across names, sectors, and geographies under nonlinear constraints,” Rasaratnam explains.

The asset manager also utilizes AI tools to discover new investment trends. Recently, it started using AI to seek out innovation trends, leveraging machine learning to analyze patent data and uncover promising innovators or firms.

“We did a trial last year where we read the patents that companies file over 30 years, and we found lots of useful information that we could get from the fundamentals of a company,” he shares. “And one signal that we found lots of value in was around identifying unique innovators, identifying companies that were filing very unique patents relative to their peers. We found that companies that were uniquely innovating had good fundamental outcomes in the future and we were able to capitalize on that.”

One challenge, though, is the limited datasets currently available. “One of the datasets used to train ChatGPT, for example, was BookCorpus 2, which contains 11,038 books – approximately 74 million sentences – by unpublished authors. Wikipedia, another training dataset, contains approximately 300 million sentences, according to our estimates. In contrast, a global equity dataset containing monthly data going back 30 years will contain only 3.6 million stock observations, a large dataset but several orders of magnitude less than the dataset available to LLM models,” Rasaratnam says, citing a study he conducted and published by AXA IM in January 2025.

Also, financial markets are non-stationary, so past data may not reflect today’s environment. For example, tobacco companies, which dominated stock market indices in the 1980s to 1990, had heavy balance sheets, which are unlike technology giants of today, which are IP ( intellectual property ) focused, Rasaratnam says.