Investment strategies used by hedge funds have evolved over the years, although the biggest changes have come in the use of computers to develop portfolios. Rosetta Analytics is a woman-founded and woman-led CTA that’s pioneering the use of artificial intelligence and deep reinforcement learning to build and manage alternative investment strategies for institutional and private investors.
Julia Bonafede, co-founder of Rosetta Analytics, served as president of Wilshire Consulting before co-founding Rosetta Analytics. She sat down with ValueWalk for an interview about how her fund is changing the way investment strategies are developed.
Gates Capital Management's ECF Value Funds have a fantastic track record. The funds (full-name Excess Cash Flow Value Funds), which invest in an event-driven equity and credit strategy, have produced a 12.6% annualised return over the past 26 years. The funds added 7.7% overall in the second half of 2022, outperforming the 3.4% return for Read More
During her time at Wilshire, Bonafede found that active managers have a difficult time having consistency in delivering alpha, which is what investors are paying for. She explained that much of that difficulty has to do with using the same quantitative investment models that have been available for decades. Bonafede added that the compute power needed to identify patterns in data in a way that traditional models couldn't do didn't exist until 2014.
"It was intriguing to me to see neural networks deployed in other industries outside asset management. We had this idea we thought could transform active management," Bonafede said. "What's very interesting about deep reinforcement learning… is our models are designed to maximize return and, specifically we're maximizing both the current portfolio return but also longer-term or compound return."
She explained that the term "reinforcement learning" means that the model has trained in a way that reinforces positive outcomes or decisions and penalizes when decisions are less optimal or incorrect.
"So over time, the model learns how to optimize market exposure. The resulting model is trained to mitigate loss while maximizing return," Bonafede said.
She said Rosetta Analytics' RL One strategy trades S&P 500 E-mini futures contracts.
"The actual performance of [RL One] shows that the model has created a portfolio through time that diversifies other portfolio asset returns," Bonafede said. "And then that if you think about it from a market value standpoint, portfolio assets have grown more efficiently. The model is trained to not lose ground by defending portfolio gains"
The firm's RL One model is up 18%. Bonafede added that the returns they have seen come at a third of the volatility of the underlying markets with low correlation to the markets. They also add alpha.
Partnering With Verger Capital Management
Rosetta Analytics was started with seed capital from the firm's partner, Verger Capital Management. Bonafede said they were looking for a way to have asset management move into the realm where latest technology improvements could help inform the investment process. Verger also invested directly in the strategy. After the Rosetta team decided which markets they wanted to target, they worked with Verger’s own portfolio to build an investment strategy that enabled it to produce a better outcomes.
"It was an interesting opportunity because most of the time when somebody creates a new investment management firm, they're spinning off from an old investment management firm, taking clients and strategies with them," Bonafede said. "We started with the idea that we could build an entirely new firm. We really could potentially bring something meaningful to the industry in terms of how to improve active management."
Building What's Possible
During the early years of her career at Wilshire, Bonafede worked in the firm's analytics group and started building business analytic portfolio software. She saw the limitations of common factor risk models, most of which were built off the same academic framework. Thus, she saw how difficult it was for a firm to differentiate itself. However, Bonafede added that deep reinforcement learning now provides an information edge because it uses neural networks.
"It's a more powerful in it’s ability to identify relationships in data. Reinforcement learning, is a holistic model framework that learns the best path forward to compounding returns," she said.
How Rosetta Analytics' Strategies Work
Bonafede said for their RL One strategy scales the portfolio exposure to the S&P 500 in and out of the market as it identifies patterns in data. Their strategies are long/ short, so the model scales in and out between boundaries of 100% long and 100% short. The strategy trades off between having exposure to the market or being in cash. She added that the portfolio has learned that natural risk mitigation is what allows them to more efficiently trade that market.
Bonafede explained that their portfolios have more upside capture than downside capture and diversifies most absolute return managers which is where their strategy fits within an allocation. She added that from an asset allocation standpoint, it does provide portfolio protection from downside risk.
Experimenting With Other Strategies
She said they are looking at expanding their strategies and are building new strategies using their deep reinforcement learning platform.
"We're an emerging manager, so we're working to grow ourselves to a level where we can expand our strategies," Bonafede said. "Our platform actually works with a range of markets and even individual securities. We have experiments that we are trading in simulation, a stock-level portfolio and a strategy that trades stocks and bonds… So a 60/40 strategy. We see similar return behavior for those strategies as well, so we know this is an inherent property. The model understands relationships in data and allocates market exposure in a more optimal way than what traditional managers are doing."
Bonafede and her team have created experiments looking at an underlying manager's active exchange-traded funds and then overlaying their models onto those strategies. She explained that in simulation, they add 300 to 400 basis points of alpha on top of what the manager is earning.
Through reinforcement learning, they have learned how to add value to those portfolios, and Bonafede said it's because their models are superior to portfolio construction approaches that managers are currently using. She noted that this technology is used outside the asset management industry where models have surpassed human intelligence. She used search engines and Netflix as examples of this technology.
Bonafede said they are the only ones using their model. Some investors are using similar technology, but as an add-on to their existing technology rather than the foundation of their models.
This post first appeared on ValueWalk Premium