Bloomberg Adds New NLP Capabilities to TOMS

An eternal struggle for technology companies is preparing for long-term disruption while answering the day-to-day needs of customers. For Bloomberg’s sell-side solutions business, this meant carving out a group of engineers whose sole focus is building solutions that don’t necessarily meet the demands of clients today, but anticipates their needs in a few years as new technologies become democratized, regulations hit, and the world becomes more complex. At the same time, it’s the kind of forward-thinking move that can help future-proof the data and tech giant against cutting-edge startups.

As an example of how this unit tries to stay ahead of the curve, it developed a new search functionality that’s embedded into Bloomberg’s Trade Order Management Solutions (TOMS) Trade Analyzer.

This new question-answer interface is, at its core, an engine that heavily leverages natural language processing (NLP) and machine learning algorithms to deliver answers to questions that are unique to the needs of traders. It aims to reduce the number of clicks needed to search for and within certain datasets, such as querying trade histories with a buy-side customer or looking up missed trades during a given time period, Robert Simek, head of sell-side analytics at Bloomberg, tells WatersTechnology.

As soon as a trader or front-office user types in a question and gets the wrong answer, they’ll never use it again. The level of accuracy and understanding that we need for the product is a really high bar.

Robert Simek

Those arduous searches typically require going into a blotter and performing a host of manual tasks, such as adding and deleting columns, changing date ranges, and applying different filters. Alternatively, clients would often just export the data, or ask someone else on the desk to curate a report with the answers they sought. As a result, it could take hours or days for a trader to get an answer to what should be a simple question.

As much as an NLP model takes loads of training time and data to understand standard human language—as a simple example, how a month like August may be spelled out or abbreviated, but mean the same thing—the team spent just as long training it to understand phrases specific to the financial domain.

“There’s slang that’s used: What does a ‘completed trade’ mean? What does a ‘done trade’ mean? What does a ‘missed trade’ mean?” Simek says. “As soon as a trader or front-office user types in a question and gets the wrong answer, they’ll never use it again. The level of accuracy and understanding that we need for the product is a really high bar.”

From ideation to full production, the project was about a year in the making. The question-answer format is only the foundation, Simek emphasizes, as clients also want to use the interface to ask more analytical types of questions, such as wanting to compare the number of trades they do on platform A versus platform B. 

Meeting that demand is where the team will take the project next. Because the machine-learning model learns over time, it can start to get a sense of patterns in questions asked, such as emerging trends, top bonds, or other topics that might require aggregation from multiple databases to get to an answer. As of today, the traditional database structure the tool currently utilizes requires too much time to answer more analytical questions.

Part of the remaining strategy, Simek says, is moving that data to more modern database structures that allow for more complex queries to yield an instant answer. Traditional relational databases are indexed on a single value—for example, the transaction ID. If a trader wants to ask a question of another column in the database, perhaps looking at counterparties in a particular sector, they’d ask, “Who are our top 5 accounts in the tech sector?,” then they would have to scan through every single record to find all the counterparties, group them, and then sum them. That’s why it takes too long, Simek says.

It’s similar to an Excel grid, he adds. To solve this problem you can use a non-relational database that can group together any data point that is represented in your dataset and make the data available much quicker then scanning all the records.

Today, as regular consumers, traders can easily use their mobile devices to answer questions by using a search engine like Google, but also access data related specifically to themselves. As apps become more interoperable, a cell phone can provide real-time info pertaining to an individual’s health and daily physical activity, for example.

At the enterprise level, traders are increasingly expecting that the same kind of functionality embedded into the tools they use every day are available in their professional tools. Simek says that although enterprise software has traditionally been built using a waterfall development process that has the software following specific written instructions, enterprise solution vendors increasingly have the engineering capacity to implement cutting-edge technologies like machine learning and more sophisticated NLP models. By working with end users, the adoption rate can more readily align with or even outpace the “diffusion of innovations” curve, Simek says.

“I think that gap today between the two [consumer tech and trading tech] is quite wide,” he says. “You’re on the train coming into work, using your phone, and you’re leveraging some of these modern technologies to search and find information,” he says. “Then you get to sit down and do your day job, and the software hasn’t caught up entirely. There’s various reasons for it. … But when you start to see this change, I think people get really excited about it.”