As sweeping new rules on best execution kick in across Europe, vendor Tora has announced the debut of its AlgoWheel product, a machine-learning-powered pre-trade analysis tool that it claims can weigh historical data to determine the appropriate execution method for client orders.
The platform takes information from Tora’s existing post-trade transaction-cost analysis (TCA) tool, feeding it into its new pre-trade TCA engine. From there, it uses artificial intelligence (AI) functionality built into the system to evaluate the appropriate method of execution using similar historical activity as a benchmark.
“We have a dataset of historical transactions that Tora collects and processes via the execution management system functionality that we have, then we have a TCA engine that conducts the calculations,” says Oren Blonstein, head of product at Tora. “So you can do certain things where, if you have an order that’s going to be X percent of today’s volume, you can understand that this instrument in this sector has a medium average spread size but high volatility, and then we can return the expected slippage based on the historical algo performance. Then we have this AI that basically sits on top of that, and will recommend the appropriate execution strategy that lines up with a given order.”
With the revised Markets in Financial Instruments Directive (Mifid II) coming into force on January 3, brokers are increasingly being forced to demonstrate not only that they achieved the best possible execution for client orders, but how they did it as well.
The engine can work in a high- or low-touch mode, automatically selecting the appropriate strategy and executing an order in the former, or recommending it to a human trader in the latter, where they want to be involved. Tora says the decisions made by the AI engine, which it calls the Strategy Server, are fully auditable for regulatory purposes.
Built on a convolutional neural network developed in-house by Tora, the outputs of the AI can follow essentially two paths—either a firm can receive recommendations based on its own historical trading activity, or it can agree to have data correlated with peers, and the engine can turn out recommendations that are benchmarked against a wider historical pool of information. As such, the AI can be run in multiple instances depending on client preference.
“There are a few modes for how this can work,” Blonstein explains. “You can either leverage all of the trade data that the system has access to, which means data from other clients, or it can be only your own data. If a customer says that they will contribute their data to the AI then they can benefit from the [wider group TCA]. But it’s an opt-in model, and the AI is only trained on data that has been approved to share, otherwise, it uses your own.”
To read the full article, click here.