Voice data could transform AI-driven bond trading algorithms

This image has an empty alt attribute; its file name is logo-cropped-11-news-300x66-1.png

“The larger the trade, the more likely it is to be done by voice. The deeper into high yield you go, the more voice, and the deeper into emerging markets, the more voice,” says Vuk Magdelinic, chief executive of Overbond. That makes it harder to apply artificial intelligence to the process behind pricing a significant portion of the global bond markets. 

Source: EuroMoney

In July, just nine US high-yield issuers with $19 billion of bonds outstanding were  downgraded, the lowest number in any month since December 2018. July was the seventh  consecutive month in which upgrades, which accounted for $33.4 billion of bonds from 30  issuers, outpaced downgrades. And while 111 US corporations with $132 of high-yield  bonds outstanding have been downgraded so far this year, almost two and half times as  many have gone the other way.  

The agencies have improved ratings on 256 US companies with $344 billion of high-yield  bonds outstanding.  

Chris Holman, portfolio manager at TwentyFour Asset Management, points out that the US  high-yield default rate has declined by 559bp this year to just 1.17% and is on track to rest  below 1% by the end of 2021, while distressed bonds today account for just 0.18% of the US  high-yield market, the lowest level since 2011. 

When spreads backed up by 40bp recently on fears of the Delta variant, some investors saw  that as an invitation to put risk on. Holman suggests: ”The outlook for the asset class is  compelling.” 

Overbond, the AI-driven data and analytics and trade automation solutions provider for the  global fixed income markets, notes that in the euro high-yield market monthly new issuance  volume, which hovered at around €16 billion in the first three months of 2021, picked up to  €29 billion in the three months from May to July. 

However, the higher the risk, the harder it becomes to capture the efficiency now coming to  liquid government bonds and high-grade bonds where algorithms suck data from all the lit  trading venues and also from T+2 settlement data, accurately predict price before a trade  and benchmark best execution.  

In these lower risk and more liquid bonds, trading is increasingly automated. But where  much of the key action is, trading remains analogue. 

“The larger the trade, the more likely it is to be done by voice. The deeper into high yield  you go, the more voice, and the deeper into emerging markets, the more voice,” says Vuk 

Magdelinic, chief executive of Overbond. “That’s just the way the market currently  functions.” 

That makes it harder to apply artificial intelligence to the process behind pricing a significant  portion of the global bond markets. 

Magdelinic says: “On any given day one quarter to one third of bond trading is by voice and  on more volatile, risk-off days it can be half by number and certainly by value.” 

And this is why Overbond thinks it latest step could be transformational. It will apply AI to  point of trade voice transaction data, in partnership with IPC, the lead provider of  communications turrets with dedicated phone lines to traders. 

“Natural language processing capability has been around for a while,” Magdelinic says. “But  traders speak to each other in their own jargon, almost in code. They’ll discuss doing so  many yards of such and such an on-the-run or off-the-run and they all know what they  mean. But for anyone else reading a transcript of the conversation, the first question is:  ‘what bond are they even talking about?’ The key with this innovation is that it translates  trade jargon and tags a specific Cusip number or Isin code onto the conversation.” 

IPC converts unstructured voice trade data into a searchable, exportable structured data  format in real-time. Overbond ingests this and feeds it to AI-driven pricing algorithms  already working on live price feeds from trading venues, settlement layer data and  individual clients’ own order management histories. 

Magdelinic says: “This translation at source, at the point of execution, adds to the speed,  precision and perhaps above all to the breadth of coverage, of our pricing algorithms. It is  incremental data on trades that would not otherwise even have been visible until T plus  two. And we believe it could increase coverage by 10-15%.” 

Overbond and IPC are testing the new approach with a select group of buy-side and sell-side  clients, presumably that do large volumes of voice trades, and could be ready to open it up  to all customers before the end of the year.  

“Our base business model is to aggregate data, now including voice, for single clients,” says  Magdelinic. “Now adding in voice data brings us closer almost to creating a consolidated  tape one client at a time. They all want to be the most powerful trading desk on the street.” 

The venture has been some time in the making. A year ago, Robert Santella, chief executive  of IPC, hinted at it when telling Euromoney about how many so-called soft turrets the  company had sold to bank traders suddenly working at home that they could download  even onto their iPads. 

“Banks are also looking more at natural language processing to deliver voice-to-text reports  from traders working remotely straight to their compliance staff. That might also be a vast  store of potentially valuable market insight,” he told us then.

Now Santella says: “Our partnership with Overbond continues the digital transformation of  fixed income trading by fully harnessing the power of voice data. As a company, IPC  embraces an open platform approach to reimagine how financial institutions everywhere  trade, share information, and optimize workflows.” 

IPC and Overbond are not yet pooling anonymized and aggregated voice data, though that  would be a tantalizing prospect. Both sides on any call already know it will be recorded as a  regulatory fairness requirement. 

Most trade conversations are initiated about and remain focused on a specific security and  the technological breakthrough here was tagging the conversation workflow to a particular  bond. 

“If conversations shift to an associated trade, dealers could be asked to click on screen to re tag the conversation to a new bond,” says Magdelinic.