JP Morgan Asset Management is using an artificial intelligence tool to help its internal portfolio managers and analysts build thematic funds. 

The tool, called ThemeBot, was created to help JPMAM’s portfolio managers generate a list of stocks associated with a theme. For example, it was used to construct JPMAM’s Genetic Therapies Fund, which aims to provide diversified exposure to companies in developed and emerging markets working on genetic treatments to address the underlying cause of diseases. 

Since the fund’s inception in October 2019, it has grown about 58% to a market value of $1.3 billion. 

Yazann Romahi, managing director and CIO for quantitative beta solutions at JPMAM, says one of the issues with using AI in finance is that a lot of the models are essentially black boxes. “A crucial part of understanding what a model is giving you is actually being able to look under the hood, so to speak,” he tells WatersTechnology.

According to Romahi, this is where ThemeBot’s strengths come in. It first creates an initial seed query—for example, “gene,” “cell,” and/or “therapy.” That is enough for the AI engine to analyze articles and other data sources and create an initial portfolio. 

Mind mapping

ThemeBot screens more than 10,000 stocks globally using natural language processing (NLP) to analyze hundreds of millions of primary and secondary data sources like company profiles, sell-side research from JP Morgan, and external research, regulatory filings, and news articles. 

Then, it creates a mind map of all the words and phrases that co-occur with “gene,” “cell,” and “therapy.”

“Imagine you were actually asked to create a portfolio on renewable energy,” Romahi says. “Your brain is going to think, ‘That’s solar, wind, hydro, etcetera.’ You’re immediately going to create a mind map in your head naturally, of words and phrases that are related to renewable energy before you start analyzing a company. So that’s essentially what ThemeBot tries to do.” 

Once ThemeBot generates the mind map of words and phrases related to the seed query, the analyst comes in. The larger the bubble on the mind map, the more relevant it is to the initial search. 

Fundamental research analysts can then refine and augment the mind map results by adding words to an “Allow” and “Deny” list, essentially narrowing down the related words and phrases. 

Romahi, whose team is focused on building factor and systematic solutions, says ThemeBot is “smart enough” to classify words like “and” and “or” as irrelevant to the seed query. The Allow and Deny list comprises words and phrases that the analyst wants to “lean into.” For example, the Deny list will represent words and phrases that the analysts think might have co-occurred with those terms but are not relevant. 

From those phrases, ThemeBot then generates a list of companies from the selected investment universe. For JPMAM’s Genetic Therapies Fund, the universe used is S&P Broad Market Index—which spans 50 developed and emerging market countries, and more than 11,000 companies. 

“It looks at every stock in the universe that we defined. In this case, we used the S&P Broad Market Index universe, but we can make the universe narrower—we can make it region-specific—and it will rank every stock in that universe based on two things, textual relevance and revenue attribution,” he says.

Textual relevance and revenue attribution

Textual relevance calculates how often terms and phrases in the mind map and the Allow list appear in the company’s regulatory filings, earnings transcripts, and news articles. Analysts can see where exactly those words and phrases appear in those data sources. 

ThemeBot scores a company’s textual relevance between 0 and 1, with 1 representing a high relevance of the terms and phrases appearing in the company’s filings and news articles.  

The second score is revenue attribution. The tool looks at the revenues of all the companies in the universe and determines the extent that those companies can link their revenue back to gene therapies. 

For example, take uniQure, a Netherlands-based company focused on gene therapies to treat patients with hemophilia, Huntington’s disease, and other severe genetic diseases. Using ThemeBot, uniQure has a score of 1 under revenue attribution, as 100% of its revenue is related to gene therapy. 

“In this case, it’s helped by the fact that the company itself breaks its revenue down on that basis, since they are a pure play,” Romahi says. 

Additionally, ThemeBot takes an aggregate score of the textual relevance, and revenue attribution and ranks how relevant the stock is.

“The beauty of that is because of this transparency, when we work, all our portfolios are done in collaboration with our active equity team, so we provide [a] safeguard. The idea is that with our research analysts, they’re able to look at every single name, look at the output and verify that yes, this is indeed a gene therapy stock, and so on,” he says. 

He says an interesting point about textual relevance is that a company may generate a smaller proportion of relevant words and phrases in their filings and news, but they may have a much higher dollar value—for example, if the theme is cloud computing and analysts are comparing Amazon and Dropbox. “Amazon is massive in cloud computing. But Dropbox is 100% cloud computing because that’s all they do, whereas Amazon does lots of other things. Would you really rank Dropbox higher than Amazon? Probably not. So, you do need to take into account the size, and not just the percentage. So that’s essentially the difference there,” he says. 

ThemeBot is a portfolio analysis tool, but the analyst has to assist the AI (to varying degrees) to create the initial mind map. The amount of assistance can depend on how focused the searches are. Romahi says creating a query on something like renewable energy probably wouldn’t need much curating, as ThemeBot will produce a focused mind map. A broader theme or topic such as “circular economy” or “future city,” meanwhile, might require more interpretation and work from active analysts. 

In a live demonstration for WatersTechnology, where “quantum computing” was the seed query, ThemeBot curated words like qubit, entangle, supercomputer, and quantum compute—all terms related to quantum computing.  

The companies—which included the likes of IBM and Barclays, both of which have active ongoing quantum computing projects and experimentations—showed revenue attribution of 0, as ThemeBot struggled to find companies reporting quantum computing-related revenues, despite having textual relevance to quantum computing. For example, both IBM and Barclays have a textual relevance of 1 in relation to the quantum computing query.  

“In this case, we would probably say that actually, this is probably too small a theme to create a portfolio on it, because it’s unable to attribute revenue,” he says. 

Theme beta

Over the last year, JPMAM analysts built almost 100 themes, not all progressing to active portfolios. Romahi says some of these themes were created through client conversations or internal interest from JPMAM’s analysts. 

“We typically want to hold a diversified portfolio. So, we have exposure to the broad theme rather than any idiosyncratic stock-specific risk. We will typically hold the top 100 names of a theme, and this is very different from when you think about an active thematic strategy, for example, where they often have maybe 30 to 40 names,” Romahi says. “This is much more about capturing the theme beta. So long as its aggregate score is high enough to be in the top 100 stocks, it will be in the portfolio.”

Particularly for a theme that’s quite early on in the development cycle, like genetic and cell therapies, there will be a lot of companies where revenues are still small. Although its textual relevance might be high, its revenue relevance could be low. 

If a company is still early-stage, it will be ranked lower and JPMAM’s weight on it will be lower. But as it starts to get its drugs through trials, its textual relevance will naturally increase. And, once the drugs go to the market, revenue relevance will also increase. 

“What you’ll find is we’ll capture it early on, but at a small weight, and then over the lifecycle of the drug development and then to market, its weight will steadily increase in the portfolio,” says Romahi. 

Of course, any portfolio is not without risks. Some of the hazards particular to genetic therapies include large upfront costs, technology or drug failures, and proving efficacy. JPMAM caters to that by incorporating fundamental metrics as part of the portfolio weighting scheme. It uses 10 different metrics measuring profitability, financial risk, and earnings quality to ensure it captures companies with more robust earnings that do not take on too much leverage.

ThemeBot runs daily to capture any news events or mergers and acquisitions activity, but the formal rerunning of the output is done monthly. However, that doesn’t necessarily mean there will be massive changes month-on-month. “The top names are probably going to stay the top names,” he says. “It’s more about capturing any new entrants or developments, whether the companies managed to get their drugs into phase one, two, and so on, so it may be more for sizing the portfolio weights. It’s rebalanced on a monthly basis, but it’s a relatively low-turnover strategy.” 

ThemeBot is similar to JPMAM’s Textual Analytics tool. Romahi says ThemeBot is based on related technology and shares the underlying core. 

“The intention is [for] all the [tools] that we all develop, we all share with each other,” he says. “So that’s why, for example, you have active analysts who are actually just using ThemeBot as part of their analysis of their own portfolios as well. Wherever there’s innovation in terms of machine learning and big data, it’s built into Spectrum (JPMAM’s proprietary technology platform) through the technology and then it’s available across the organization.”