In a downtown Toronto skyscraper one block away from the Hockey Hall of Fame, a small hedge fund is hoping it has found an edge in financial markets. Castle Ridge Asset Management is betting on Wallace, a purpose-built supercomputer behind the hedge fund’s trading strategies driven by artificial intelligence.
For years, hedge-fund players have wondered if AI could help them beat the market, but AI-trading efforts they launched often led to disappointment and amounted to little more than marketing schemes to pull in client money. With the launch of ChatGPT in November 2022, however, this new breed of AI-driven hedge fund players has been reinvigorated.
Castle Ridge, which was founded by current CEO Adrian de Valois Franklin in 2015, is a relatively tiny player in the world of hedge funds, with around $190 million of assets under management, and operating in a town that is not known for producing market-beating hedge funds. Still, Valois-Franklin believes the investment fund’s approach to predicting movements in financial markets using AI could make it a serious player in the multi-trillion-dollar hedge fund industry.
A former investment banker with little previous quantitative-trading experience, Valois-Franklin claimed that Wallace’s main selling point over rival AI-powered hedge funds is its ability to constantly refine its own models using evolutionary processes that have been likened to selective breeding. Speaking to MarketWatch, Valois-Franklin described Wallace like it was a multi-manager hedge fund in which virtual portfolio managers are constantly “battling each other to see who is the most fit in this environment.” But the hedge fund chief notes that unlike human portfolio managers, Wallace never needs to sleep and never needs a pep talk.
In simple terms, Wallace’s evolutionary process sees the machine create thousands of differently-weighted virtual investment portfolios each day, which are tested and ranked according to their suitability to current market conditions, Valois-Franklin said. In a recurring eight-hour cycle, Wallace picks out its top-performing portfolios and gives them priority to “breed.”
“On a daily basis, Wallace will make thousands of copies of itself, each a virtual portfolio manager with different characteristics. It will turn certain weightings and patterns up and down, or on or off, and then determine whether each portfolio manager is better or worse suited in this market environment we’re in today… If it’s better, it increases the probability that it breeds” Valois-Franklin said.
‘Like a flock of birds’
Castle Ridge has had some success in generating investment returns. Since the inception of Wallace in 2017, the investment fund has generated annualized net returns of 12.4%, compared to the S&P 500’s
SPX
12.1% returns over the same period, documents seen by MarketWatch show. It’s going up against massively resourced quantitative hedge funds, like Two Sigma and D.E. Shaw, that are working to make inroads into machine learning and AI.
Speaking to MarketWatch, Castle Ridge’s chief scientific officer, Alex Bogdan, argued Wallace’s evolutionary approach allows for a deeper level of understanding compared to the neural networks used by systems like ChatGPT, which are modeled on the human brain.
In Bogdan’s view, these evolutionary processes represent the future of AI, in allowing machines to go beyond simple mimicry. Bogdan explained that neural networks, which are most prominent in the form of large language models (LLMs) like OpenAI’s Chat GPT, simply “mimic the responses that a human would make given the same input.”
In contrast, Wallace’s “genetic algorithms” work to combine the individual bits of knowledge it has, to build on its own understanding and become “incrementally smarter.” “What GPTs are, are clever algorithms,” Bogdan said. “We have enough mimicking. We need understanding, not cleverness.”
Early research into AI first started in the middle of the 20th century, on the back of developments in computer science made during World War II. In one 1961 experiment, British scientist Donald Michie successfully developed a machine made out of matchboxes that was able to solve the game of ‘noughts and crosses’ and play against human opponents.
Michie’s machine, which was called the Matchbox Educable Nought and Crosses Engine — or MENACE for short — used matchboxes to represent all 304 states of play in the game of tic-tac-toe, with each small box containing beads to mark the relative advantages of each position.
The matchbox machine would in turn make moves based on the number of beads in each box, in a system that saw it rewarded with beads for each winning move and punished with the removal of beads for moves that saw it lose, until it eventually solved the simple pen-and-paper game. The systems used by Wallace are based on a subfield of AI called “evolutionary computing,” which seeks to solve complex problems using continuously adapting algorithms.
Like Michie’s machine, Wallace maps out scenarios to pick out those that are most successful and then reinforce those winning strategies. But unlike Michie’s matchbox engine, Wallace operates in the complex world of financial markets, where the parameters are always shifting.
Castle Ridge’s success is based on its AI machine’s ability to adapt to shifting market conditions — meaning that unlike Michie’s matchbox machine, which quickly solved the game it was designed to play, Wallace must constantly be adapting its strategies.
“The system isn’t trying to determine what’s going to happen in the market. It’s trying to anticipate how the players in the market are reacting, to news as it happens. From that perspective, the system is less interested in the fundamentals of the cards that have been dealt in this poker game, and more interested in the tells of the other players at the table,” Valois-Franklin said.
Castle Ridge says that as a byproduct of this strategy, Wallace has successfully predicted a series of market events ahead of official announcements, based on signals in the data it analyzes.
Valois Franklin explains that Wallace looks at markets “like a flock of birds, that’s constantly shifting and morphing,” to pick up on signs of early movements driven by insider knowledge.
““The system isn’t trying to determine what’s going to happen in the market. It’s trying to anticipate how the players in the market are reacting, in real time, to news.”
Those predictions include Wallace’s bet on Gilead Sciences
GILD,
ahead of the company’s push to acquire New Jersey biotech Immunomedics in September 2020, before shares in the cancer-treatment company surged by more than 100% after the takeover deal was made public.
“As soon as individual securities start to fly away from the flock, that’s one signal to Wallace that says, ‘Zero in on this, why is this security behaving more independently versus its peers.’ And often independence of behavior is indicative of knowledge that’s imprinted on the security. Often, when people don’t really know anything, they tend to act in lockstep with others.”
The hedge fund’s staff now spend their time trying to “break” Wallace’s system, by throwing in “unknown unknowns” and offering the AI new data. In one case, the team fed Wallace satellite images of Walmart
WMT,
parking lots, to see whether the information could help the machine predict consumer behavior.
In Valois-Franklin’s view, this sort of work may soon occupy the majority of the working day for those working in the world’s top hedge funds. “It will replace some types of jobs but it will open up capacity in certain areas. We’re not sitting around reading research reports but we are doing other things to help increase the productivity of the system.”
