Inside RBC’s Aiden project: 5 years of deep learning

Aiden, a trading platform launched last year, is the product of five years of experimentation with deep learning by RBC Capital Markets on top of an additional five years of hypothesizing about what best execution would one day require.

When RBC Capital Markets launched its artificial intelli­gence-powered trading plat­form, Aiden, in the fall of 2020, it was the culmination of 10 years of active de­velopment work, as the firm sought to address common client issues such as slip­page and alpha erosion while executing trades, and periods of volatility that could upend the best historical models.

Machine-learning and AI projects on Wall Street become outdated quickly as market conditions and technology evolve, keeping the makers of such tech­nologies on their toes, and end-users in the ongoing throes of complex, onerous integrations, compliance checks, and training processes.

So when RBC Capital Markets set out to build a trading platform that could withstand the test of time and evolve with its surroundings, they turned to deep reinforcement learning, an advanced form of machine learning under the AI umbrella that remains underutilized in the capital markets.

Shary Mudassir, RBC’s co-head of global electronic trading, joined the firm in 2009 as an associate in the capital markets rotational program, testing his software engineering skills on different desks and gaining an affinity for elec­tronic trading. He watched the space change drastically over nearly 12 years, as clients went from phoning brokers, to scouting sellers or buyers for their orders, to executing their own trades directly in the marketplace, which itself has become a complex network of lit venues, dark pools, off-exchange venues, alternative trading systems and dozens of other liquidity pools.

Ten years ago, the electronic trading desk zeroed in on crafting high-quality execution algorithms, building them inter­nally from scratch. But it wasn’t until five years ago, when RBC Capital Markets partnered with Borealis AI, RBC’s research institute, that the concept of the Aiden algorithmic suite was conceived.

One of these algos, volume-weighted average price (VWap), was the first Aiden solution to go live last fall. With more than 150 of RBC’s largest clients now using it, the algo is meant to reduce slippage—the difference between the expected price of a trade and the price at which the trade is executed—against the VWap benchmark, which gives the average price a security has traded throughout the day, based on both volume and price. Slippage can happen fairly often, but especially so during peri­ods of heightened volatility, he says.

“As we built those solutions over the past 10 years, there was significant growth in the complexity of the code base. And when you marry the complexity of the code base with the sheer dynamic nature of the market, you really need a very robust review and enhancement process to ensure that your execution algos will always be at a high-performing level,” Mudassir says. “Because the moment you roll out your execution algos and the intelligence within them you’ve coded in, market dynamics may change, market participants may change, others might pick up on your approach to executing.”

In past periods of market volatility, RBC would have to re-code its algo­rithms to adjust for new conditions, and continually review and tweak the algos for as long as the volatility lasted. And when that happened, the firm had to hope that those changes worked. It wasn’t an ideal system, Mudassir says.

So RBC and Borealis began devel­oping their own deep reinforcement learning models to account for all the possible scenarios that can upend a trade—and even a market—so that the algos could become self-learning and autonomous.

The way Mudassir thinks about deep reinforcement learning is that it must work toward an end goal, but how the algo gets to a decision matters the most. Essentially, it isn’t about taking the best single action, but about taking the best possible series of actions—much like a chess game. In chess, “you make a lot of moves, each move can have points, but it’s the cumulative best moves by a player that lead to that person winning,” Mudassir says.

Though the rules are constant, every game of chess is slightly different. Like an algorithm, a player must retain some memory of previous games, while both adapting to changed or changing vari­ables and forecasting for future variables yet unknown. It is exploitation (sureness of what is already known) and explora­tion (lower levels of confidence in the unknown) working in tandem.

With those components in mind, a skilled chess player or a sophisticated algorithm must consider that “a move right now may seem extremely great, but it may not be the best move because I’m working toward a future end goal,” Mudassir says. That’s the principle that guides Aiden.

The first real test of Aiden’s adapt­ability and durability arrived with the onset of the Covid-19 pandemic, which spurred the failure of countless historical models. Before the October launch, it had been in a beta trial with a number of large clients for months, who were able to preserve their performance despite overwhelming market uncertainty, with­out any manual intervention from RBC, Mudassir says.

But all the promise of deep learning comes with a catch. The more complex a model is, the bigger the issues of transpar­ency and explainability.

So in tandem with Aiden, RBC built a secondary platform alongside it called Aiden Insights, a client-facing portal available via web or mobile devices, which offers users real-time visibility into the decisions Aiden makes on how it executes on their order flows. Mudassir hopes that Aiden Insights can be example of trust and transparency for the rest of the industry, especially those who are hesitant to use the technology due its opaqueness.

RBC has begun building off of its VWap algo, with a new release pegged for the fourth quarter of this year that is currently in testing with clients. The second algo targets another important benchmark—arrival price, which is the midpoint between the bid/ask prices at the time an order is placed. It can be difficult to hit that midpoint because as a trader starts buying up certain shares, sellers can get a sense of that movement, causing the price to rise and effectively minimizing the alpha that a portfolio manager wants to capture for their own clients.

Though RBC’s live or in-testing implementations of deep reinforce¬ment learning are currently limited to execution, the firm is considering how it can be used to navigate the fragmented network of liquidity pools and how to best deliver timely and consumable information to clients.