The next generation of smart order routers is upon Wall Street. Unlike the rules-based first generation of SORs or the second generation that incorporate liquidity-seeking algorithms, the third generation of SORs leverage artificial intelligence to boost their performance.

However, routers that rely on AI to decide where to route and order on an order-by-order basis are still a ways away, according to experts.

“Currently we do not feel that the technology is mature enough to make changes in real-time although in practical terms if you define the correct scope, there is no reason it could not,” Medan Gabbay, chief revenue officer of Quod Financial, told IntelAlley.

Similar to other implementations of AI within the capital markets, firms turn to disciplines like machine learning to augment existing performance by making simple changes to static systems.

“Machine learning methods are mostly making subtle improvements to capture spreads in a market where margins are being squeezed tighter and tighter,” said Gabbay. “It is able when used responsibly to smooth out errors and provide contextual recommendations to traders who move into the role of ‘operators’ and ‘exceptions managers’ as opposed to micro-managers of trading decisions.”

However, machine learning cannot run in a vacuum, and firms need to implement it with low-latency risk controls, data visualization tools, as well as monitoring and reporting capabilities.

Quod Financial has applied machine learning to areas that enhance SOR performance, such as predictive failures, parameter optimization, price revision, and pre-trade strategy selection.

“Each of them is operating as a part of a larger system to improve the overall execution, but this is a far cry from unleashing a pure-AI onto the financial markets,” he said.

It will help with simple predictive decisions, such as when to peg an order, become aggressive, tweak one of the more than 300 SOR parameters, or select a broker’s trading algorithm.

“These are all decisions that have a huge impact on P&L and can be easily derived from backtesting and pattern recognition while leaving human traders to focus on illiquid and complex trading without having to understand the nuances of a 6.5 or a 7.5 out of 10 settings on an aggressivity setting,” said Gabbay.