A Canadian startup applies machine-learning to corporate bond issuance
WITH the exception of a few governments big enough to run their own auctions, anyone wishing to issue bonds must seek bankers’ help. A hefty fee will buy assistance in calibrating the size, structure and timing of a bond issue, as well as connections to lots of buyers. And once a bank has agreed to underwrite an issue, it bears the risk of failing to get a good price for the bonds. But the process is old-fashioned and inefficient (the head of bond origination at one American bank jokes that “not a lot has changed since 1933”), and the accuracy of the advice is hard to gauge. Overbond, a financial-technology startup in Toronto, wants to change all that.
Investment bankers responsible for bond issuance still operate largely by feel, calling up asset managers to get a sense of demand, rather than by crunching numbers. Rules against insider trading mean they cannot talk directly with their trader colleagues. Data on existing bonds are more abundant. In America, for instance, information on the price, timing, yield and volume of all bond transactions must be reported publicly within 15 minutes. But so far, comparing primary and secondary markets has been difficult. By crunching a wide array of public data, Overbond seeks to provide a link between the two.
Its main offering is a set of machine-learning algorithms powered by neural networks, a type of artificial intelligence, that predict the timing and pricing of new bond issues. The service is already fully in place for the Canadian corporate-bond market, and partly so for the American one. The algorithms crunch through credit ratings and real-time data on secondary trading for a firm and its peers, among other things. Recent predictions for the yield on new bond issues have been, on average, off by less than 0.02 percentage points.
A subscription buys tailored estimates of demand for new bonds, including the interest rate the market is willing to bear. This helps corporate treasurers gauge market conditions and decide when to issue bonds and in what maturity. Of the 200 or so Canadian corporations that issue debt frequently, 81 are signed up.
Investors can use a basic version of the service without charge, partly because the firm collects data from them that then feed into the algorithms. They can, for instance, get estimates of the timing of the next bond issue to hit the market, using data on the timing of previous issues, issues by similar companies and balance-sheet data. Around half of Canada’s institutional bond investors use it in some way.
Canada’s corporate-bond market is a relative tiddler, with a total of 604 new bond issues in the past two years. Its investment-banking community is small, too; Overbond reckons that every new bond issue passes through one of just seven individuals. But the firm now hopes to break into America, the world’s largest corporate-bond market with around 3,000 new issues annually. There, issuance is much more fragmented. Around 40 banks are active in bond origination, and no firm has more than a 12.5% market share, according to Thomson Reuters, a financial-data firm.
Vuk Magdelinic, Overbond’s founder and chief executive, says that starting small in Canada gave the firm the chance to perfect its algorithms. It has refined its timing-prediction algorithm for the American market (see chart for an example on Microsoft). Some actively managed bond funds have already expressed interest. It has opened a New York office and is seeking funding from American investors.
Bankers, perhaps unsurprisingly, proclaim themselves sceptical that something as sophisticated as bond origination could be pried from their grasp by a fintech challenger. Instead, they think they spy an opportunity. Some have expressed interest in using Overbond’s timing algorithm to help spot firms in need of financing before they come asking for it. In finance, as elsewhere, machines and humans may be more powerful together than either is alone.