Start-up uses ‘Magic’ to democratize access to AI for hedge funds

SigTech, a Software-as-a-Service platform that provides portfolio analytics services to hedge funds and asset managers, is using large language models and generative AI tools to improve firms’ data retrieval times and increase efficiency.

Bin Ren, SigTech’s founder, is the former chief investment officer at Brevan Howard Asset Management’s Systematic Investment Group, and he originally developed the platform as an in-house system for the fund. But the hedge fund spun it out in 2019, following a playbook written by other large asset managers, such as BlackRock and its decision to externalize its Aladdin order management system, Ren tells WatersTechnology.

“SigTech builds tech infrastructure that has a much bigger TAM (total addressable market) than mere internal use by BH,” he says. “Aladdin is the perfect example to illustrate the strategic importance of such a move: to become [the] industry standard.” Ren explains that Brevan Howard has not developed an in-house successor to SigTech’s platform, and the hedge fund builds on top of SigTech’s proprietary infrastructure.

Since 2019, SigTech has refined its business model and looks to compete with other portfolio analytics platforms at small to medium-sized investment firms. Typically, firms attempting to build technology stacks in-house tend to run into high costs, and firms outside the rarefied group of household names like Citadel and Millennium often don’t have the human capital in terms of quants and analysts to compete.

Krishna Nadella, vice president of commercial and corporate development at SigTech, hopes that the company’s newest project, the Multi-Agent Generative Investment Co-Pilots, or Magic for short, can help. Before taking on his current role, Nadella was head of corporate development and market strategy at Symphony, and before that, he worked for 10 years within the enterprise data division at Bloomberg.

Nadella reckons that the combination of GenAI and large language models will save time for quants and researchers, as Magic is able to retrieve relevant data and insights in minutes, while human analysts on their own can take hours. He’s confident, too, that Magic will see uptake once it’s released in September.

“We don’t want adoption, we want addiction,” Nadella says. “We want these guys to be so reliant because it’s so easy to use, because it’s natural language, that they don’t want to get off it.”

Inside the spellbook

Magic leverages the collective powers of a corpus of LLMs, each trained on a specific subsection of data, to answer common questions in a few minutes. Sharing his screen over a video call in early June, Nadella uses his cursor to indicate each model. Each proprietary model, portrayed visually as an inch-high box with a differently colored outline, is described as an individual “agent” within Magic. The various names for these agents describe both its specific function and which datasets they have been trained on.

For example, there’s a SigTech Quant Agent, trained on quantitative data, and a SigTech CFA Agent, trained on the questions and answers required for a human to become a chartered finance analyst, a process which typically takes three or more years to complete. Both these agents sit alongside the Browser Agent, which is responsible for sourcing information beyond the immediate expertise of the other agents.

One quirk of LLMs is that asking the same question of two models can produce different answers, based on how they are trained and constructed. To counteract this, Nadella’s team have ensured Magic can swap between different AI chatbots by adding a drop-down menu where users can select their preferred option, including OpenAI GPT, Grok, Mixtral, Google Gemini (formerly known as Bard), and Anthropic.

When a user types a question into the interface, one of the Magic agents, the Planner Agent, selects the most relevant model to answer the question, which avoids agents being forced to give answers outside their scope. If asked a question about the current price of a stock, the Planner Agent sends signals to the Browser Agent to access relevant data online. This is SigTech using Retrieval-Augmented Generation, which overcomes the LLM limitation of answering questions related only to its training data by accessing real-time data via the internet.

In his example demonstration, Nadella asks Magic what the major market trends are today. The process involves three of the agents ‘talking’ to each other and takes about 45 seconds to flash an answer onto the screen. Initially, it produces a few lines about the stock performance of chipmaker Nvidia, but after a couple more seconds, a longer, detailed series of paragraphs comes up about possible macroeconomic factors to be aware of when investing. Nadella says this is a result of the Macrobond Agent—a specialized agent trained on data from financial data provider Macrobond.

Nadella says that the purpose of specialized agents like the Macrobond Agent and the CFA Agent is to democratize the buy-side space by allowing smaller funds to access top-level data and CFA charter-holder experience “in the palm of [their] hands”.

“Our CEO took the repository of all the questions available for the CFA level one, two and three exams, and was able to apply large language models to them to answer the questions,” Nadella says. “When we started this process, I think [the agent] was at a 70% success rate. Now it’s at a 90% success rate, and we expect it to get to a 95% success rate by the end of the year. So that effectively allows you to think of having a CFA charter-holder in the palm of your hands, because they effectively can answer any of those questions as well as, if not better, than a lot of people who take that same exam.”

Success rates are an interesting thing to bring up when it comes to GenAI, especially given the occupational pitfalls that come with the technology. Hallucinations—robo-generated falsehoods that can occur when there’s not enough training or input data to answer correctly—undermine the credibility of GenAI applications, and making trades on incorrect information can be costly.

Nadella says that while the team notes that hallucination risk is never zero, they have taken steps to mitigate that risk within Magic through the Browser Agent, which makes API calls to SigTech’s data offering.

“We’re using large language models to ask the question; we’re not using them to answer it,” Nadella explains. “SigTech has disaggregated its offering down to the API level, so every time a question is asked, it’s making an API call back to a very succinct and distinct source that allows you to have control and know where the data is coming from. We’re making API calls back to our flagship product—our data.”

Any volunteers?

While the zeal of GenAI has attracted a number of ardent believers, there are skeptics out there, too. Even companies which develop their own co-pilot solutions do not think that AI-assisted products are always necessary. The founder of a Nordic hedge fund, which is a client of SigTech, says that in the investing space, GenAI products are more bark than they are bite.

“In general, I would say it’s more of a gimmick,” they say. “I’ve used ChatGPT for some coding questions and so on, and it seems to be doing quite well, but as a co-pilot? I’m not sure.”

As a longtime user of SigTech’s SaaS services, the fund founder explains that while Magic is interesting, it would not change how the fund already uses SigTech’s offerings.

“The way SigTech has built it with several different models talking to each other, I think, is quite impressive, but I think [Magic] won’t have a major impact on how we use SigTech,” they say. “For us, it’s a nice extra and quite useful in finding information.”

Gimmick or not, as GenAI hype continues to build, companies are looking to capitalize, with some viewing what constitutes AI more liberally than others. Nadella disapproves of companies claiming that they are building AI solutions with no use obvious use case, and he wants to make it clear that SigTech will not be caught up in the same hype cycle.

“We are not an AI firm. We’re not evolving to AI. We’re not, all of the sudden, going to slap an ‘AI’ on our name,” he says. “Over the last two years, the number of firms that have either conveyed that they are an AI firm or they’re building out from AI has really led to a lot of questions about the validity of AI from a practical and pragmatic standpoint. Show me the money. Show me the beef. Leveraging AI [and] leveraging large language models to apply to an existing product offering—that’s the future.”

A portfolio manager at an alternatives asset manager that recently started using SigTech believes that as GenAI technology matures, it’ll become a staple of the financial services space, offering “unparalleled insights” that traditional methods “cannot match”. They explain that one of the more important aspects of SigTech’s offering is helping to democratize access to powerful AI models on the buy side.

“SigTech empowers smaller buy-side firms by providing access to sophisticated AI models that were previously the domain of larger firms with substantial resources,” they say. “By leveraging these advanced technologies, smaller firms can enhance their data analysis capabilities, make more informed decisions, and compete more effectively in the market.”

Nadella says that SigTech has spent the past month demoing Magic to existing customers, and while the official launch is in September, select customers will get advance access to the product in August, so they can provide feedback. He’s expecting a lot of it.

“We’re going to say, ‘Hey, look, we’re still fine-tuning, so caveat emptor—buyer beware. But at the same time, you’re going to give us so much feedback because you’re going to be hammering this thing,’” he says.