Artificial Intelligence at QB

It is hard to escape the seemingly incessant talk of Artificial Intelligence these days. The media has endless coverage about how machines are taking over our lives, automating even more tasks and far outsmarting humans. Without doubt there are some significant advancements happening in this area, but there is also a lot of hype. So, what does Artificial Intelligence mean at QB?

A definition of AI is “the capability of a machine to imitate intelligent human behavior”. Since we launched Bolt, we have effectively been offering an artificially intelligent execution algorithm. Not in an “inter-order” sense (over the course of many orders over time) but “intra-order” (while a specific order is being worked). As such Bolt is capable of making complex decisions in real-time based on its reading of market data. Bolt’s intra-order artificial intelligence is significant – it has no pre-defined schedule, it understands the historical context of how the instrument typically trades, combined with reading market conditions in real time. Every decision it makes is for the purpose of doing what is optimal for that order at the time. This replicates how the best human trader would read the market and handle the execution, but at a speed, scale and complexity that no human can undertake. The trade off between capturing spread, minimising market impact and price risk is optimally managed, completely systematically by Bolt. 

To improve performance in the long run, our algorithmic engineers and researchers carefully review how Bolt performs. Execution quality is reviewed, outliers are analyzed, and we recalibrate based on what we learn. To that end, Bolt’s evolution in the years since it was launched has been an iterative and ongoing process, making it better able to handle all types of orders, large and small, across wide ranging instruments and market conditions. This “inter-order” learning is undertaken scientifically but manually by our team. The next frontier is ongoing machine learning – artificial intelligence whereby an execution algorithm is able to continually educate itself from each order and market experience, to learn what it could do better next time. We are, however, a long way from this being a reality. While this kind of machine learning technology is developing, and we welcome these new approaches here at QB, the main reality check for this kind of AI is testing. We undertake extensive reviews and testing for all code changes, including full releases in our highly realistic Simulation environment long before going live in the real market. Allowing the algorithm to effectively change its own code is something we would need to spend a lot of time getting comfortable with, and unlikely anytime soon under the industry’s well justified scrutiny of algorithms. There is no doubt clients, executing brokers, exchanges and regulators would all have concerns. Where AI advances will most likely help us is in identifying potential improvements but allowing for thorough checking and testing by our engineers before any changes are rolled out. 

For these reasons, although our area within financial technology will likely be a late adopter of full machine learning applied inter-order, in other respects we are already very advanced, perhaps more than most people would imagine. “Artificial Intelligence” is an appropriate description for Bolt, QB’s leading arrival price execution algorithm for futures and cash treasury markets. 

For more on the AI topic, QB recently participated in a video discussion and an article

Quantitative Brokers
New York
July 24, 2017