Ben Polidore, Managing Director, Algorithmic Product Management, ITG
Traders seeking electronic execution tools in the equity markets have no shortage of choices. At last count, there were more than 1,600 broker algorithms and strategies, each with its own unique fingerprint. Despite this vast landscape of trading strategies, traders don’t necessarily see variety as a good thing. A study earlier this year by Greenwich Associates found that only 7% of buy-side traders feel completely satisfied with the standard algorithms offered by their brokers. It is this discontent, along with advances in research and technology that present an opportunity to reinvent algorithms. At ITG, we’ve been developing technology for self-directed trading algorithms since 1998, and are beginning to incorporate Artificial Intelligence (AI) into our algo suite following two years of extensive research. Given the growing popularity of the topic, the following is a brief overview of the evolving AI landscape and its growing application to the financial industry, particularly equity execution.
Gamers, Techies and Quants
The explosive growth of the video game market, which currently stands at $100B+ in global revenues, created millions of customers for graphics processing units (GPUs) and continues to drive rapid improvements in GPU capacity and functionality. Today, a commercial GPU with 9 TFLOPS of computing power (i.e. able to perform 9 trillion floating-point operations/second) is as fast as the 500th best supercomputer in 2008 and retails for just $5,000. Computationally intensive calculations, once requiring expensive and complex server farms, are now run on a significantly smaller footprint at much lower cost.
As advances in computer engineering quickly catch up with academic research in computer science, what was once considered theoretical is now achievable. Simultaneously, widely available research on AI and machine learning is leading to an explosion of interest across many industries in adapting this technology to solve real world problems. In addition to research in academic journals, vast amounts of information are freely available on the internet. Anyone seeking an introduction to advanced reinforcement learning concepts need look no further than YouTube to get started. Since much of this research is publicly available, this results in fewer patent entanglements further reducing the barriers to entry and the costs to employing AI.
The Challenge of Equity Execution
Despite these technological and intellectual advances, speed remains a challenge in applying AI to equity execution. In a market operating at microsecond latency, AI-driven execution models operating on a millisecond scale are too slow for certain tasks such as limit order placement or pre-market risk checks. Major advances in training models via horizontal scaling (both on-chip and across chips) haven’t done enough to improve latency at runtime for real-time, in-band trading decisions. Therefore, we have focused our efforts mainly on discretionary order planning, which isn’t as constrained by latency and may even have more upside for performance improvement.
The need for transparency in the execution process is another big hurdle to AI adoption. In short, AI-based strategies are less deterministic than traditional trading algorithms. On the one hand, this behavior may prove valuable to a trader as less predictive strategies may reduce information leakage by minimizing the signaling risk caused by repetitive behavior. On the other hand, when you ask a model to find hidden relationships between many underlying fundamental factors, it may be difficult to understand and explain what motivates a specific action. Although the motivation may lack transparency, the governing logic is easy to understand, since the goal and the context to achieve that goal are explicitly defined in the AI model definition. To effectively understand and explain AI-based strategy behavior requires traders to effectively reframe the problem and educate their stakeholders (i.e., investors and portfolio managers) making it necessary to equip the trading desk with the appropriate analytics required to facilitate this task.
Another challenge to adapting AI for execution products is that posed by the paperclip maximizer in AI researcher, Nick Bostrom’s, now-famous thought experiment: AI algorithms seek to fulfill their goals with maximum efficiency, and with no regard for consequences outside of their objectives. In Bostrom’s thought experiment, a super-intelligent AI tool tasked with maximizing paperclip production might try to turn all available atoms in the universe into paperclips. In an execution context, this means an AI algorithm set to a particular benchmark, say the opening price of a stock, might behave too aggressively in hitting that goal, essentially rediscovering practices already tried and discarded by human traders. Carefully considering the objective function and its second-order effects, is crucial when employing discretionary AI in the execution process.
Rise of the Robots?
To some, the term artificial intelligence elicits feelings of deep despair at the rise of the machines and the inevitable decline of human work. It is beyond the scope of this article (and its author) to speculate on the potential socio-economic impact of artificial intelligence on the financial industry. However, it is worth recalling that similar fears arose with the introduction of trading algorithms and direct market access. In time, they simply became weapons in a trader’s arsenal and a way to manage an increasingly complex market. We view AI in the context of equity trading in two ways: First and foremost, as an opportunity to improve trading performance and second, as a pathway to simplify a landscape of trading strategies which has grown in complexity over many years. Used appropriately, AI can enable traders to effectively manage trading risks and allow them to focus on the tail events that adversely impact trading performance and, ultimately, investor returns.