How many clicks does it take for you to find what you’re looking for online? Five? 10? More? What if you could find exactly what you wanted with no clicks at all?

This was the idea behind FactSet’s voice-activated analytics, developed in collaboration with Microsoft. Trading firms have stockpiles of data at their disposal, but finding what they need quickly remains a challenge, says Chris Ellis, head of strategic initiatives at FactSet.

In a demo for WatersTechnology, the voice-enabled technology processed the search request⁠: “BlackRock ETFs (exchange-traded funds) launched in the last two years with an expense ratio between 50 and 100 basis points with a weight in Germany greater than 2%.” It served up matching results in roughly eight seconds.

“What we’re doing with FactSet smart search is allowing the user to just say in English what they want, give them the answer, and then give them the link to where there’s more information. So, it’s one part an answer to their question, and another part [is] providing them the navigation tools to go down the rabbit hole and do a much deeper dive into the analytics,” Ellis says.

In September, the data vendor will roll out an integrated version of the voice-controlled technology across its ETFs and funds datasets, designed to help traders, portfolio managers, and analysts verbally query the data within the application. In February, FactSet announced its integration with Microsoft Teams, whereby Teams users can access and share the data vendor’s financial news and content within the Microsoft communications and collaboration platform. As an extension of that partnership, FactSet is leveraging Microsoft technology to help build out its voice-activated search tools.

Microsoft provides the underlying natural-language processing technology that recognizes the human voice and translates it to text via a speech API. This function is available to FactSet users across its data and analytics platform via the vendor’s web or desktop application.

For the ETF and fund screening datasets, FactSet has developed a more integrated and advanced voice-activated capability. On top of the Microsoft speech API, FactSet has built a layer of machine-learning technology that applies meaning to industry lingo (for example, “expense ratios” and “ETFs”), then formulates how search requests are understood and how results are generated.

“What we’ve done is not only taught the engine how to interpret specific requests, but we use our knowledge of finance and of the fund and ETF ecosystem to say, ‘If you put in a qualitative request, instead of a purely quantitative request, we’re going to interpret that in an intelligent way, but then still give you the keys to the car to go in and change it to what you want,’” Ellis says.

The model can recognize specific and generic verbal search requests. For instance, a specific request might include different components like the name of an issuer, such as in the case of the BlackRock request; the fund’s launch date; the expense ratio; and the weighted average market cap.

A qualitative or generic search request might include general adjectives like “new,⁠” “low,” or “large.” An example Ellis provides is “new ETFs with low fees that invest in large companies.” In this instance, FactSet’s model will allocate default rules for what “low fees” might equate to, such as fees lower than 15 basis points; or what a “large company” might be, such as one with a weighted cap of $10 billion or more. These default settings, however, can be adjusted and customized by users.

Schooling the bot

Building machine-learning models for understanding speech and financial language is not without its challenges, says Mike Marley, product manager for search and lookup technologies at FactSet. The most difficult part, he says, is training the model to recognize and comprehend the different ways in which a person might phrase a question or search a request. This involves teaching the machine to accurately interpret multiple sentence formations and syntax.

FactSet has been building the model for more than two years and Ellis says a huge portion of that time was used to train the technology with sentences from beta users for the ETF and fund screening application.

“It’s a solvable problem to say, ‘Here are 50 different questions; we can answer those.’ But you have to be able to answer the questions based on how it’s being asked, and there are so many different variations and different ways to ask those questions,” he says.

When building a voice-activated solution, developers also must account for human error. Take, for instance, a user who requests the price of Facebook shares on July 10, a Saturday. The system should be trained to recognize the market was closed on that day and should generate the results for Facebook on July 9, the Friday before.

So far, FactSet’s model, which is built on top of Microsoft’s speech API, works only for its ETF and fund screening datasets. However, the vendor has plans to extend this functionality across its analytics suite. As part of the broader integration roadmap, FactSet also hopes to build connections to digital voice assistants like Amazon’s Alexa, Apple’s Siri, and the Google Assistant in the future.

“We have started down that path or are looking down that path. We have not done that integration into the voice assistants themselves yet, but that is something where a lot of the work that we’re doing now would apply as we go forward,” says Marley.