Getting aggressive: Overbond uses AI to assess dealer axes

As Vuk Magdelinic explains it, in fixed income, buy-side traders want quicker responses on less liquid bonds and better tools for trading blocks. If this is the case—and the research backs that up—he says traders need more information on dealers’ price aggression to make better decisions. 

This has led Magdelinic’s company, Overbond—of which he is CEO—to develop what it calls AI-adjusted price aggression levels. The tool uses gradient-boosting algorithms and open-source machine-learning libraries to provide three tiers of price aggression for buy-side users: conservative, normal, and aggressive. 

By looking at the aggression level and seeing the real price a trader gets back from a dealer after initiating a request-for-quote (RFQ), traders can then make a judgment call on whether they received a good price or whether a dealer put their foot forward. 

“Aggressiveness is not the same for each buy-side desk,” Magdelinic says. “They’ll be uniquely different based on what they trade, how they trade, what their rebalancing strategy is, and what their portfolio and flow looks like.”

To build these customized tiers, the vendor typically collects 2.5 years of a customer’s trading data to train the model, says Overbond’s product director, Vidal Mehra. This information is used to assess how a trading desk executes on certain kinds of trades, adjusting for things like risk, size, type of bond, or market conditions. 

When there is a high level of volatility in the market, the model will alert the trader that it has a lower degree of confidence in the price output. “It will tell the trader that this isn’t what we call a tier-one price; it’s actually a tier-two or tier-three price and you should tread with caution,” Mehra says.

Whether a dealer offers a good or a bad price is, of course, subjective, but what the buy-side customer wants is consistency. They want to know that the price on the screen is what they are going to get

Vidal Mehra, Overbond

For traders initiating a request for quote (RFQ), their main concerns are primarily information leakage and the price moving against them. On top of that, they’re worried that they’re not getting the best bid or ask from dealers, he says. “We came up with the aggression levels as a way of showing buy-side traders what they can expect back from market-makers based on different factors.”

What has also helped Overbond is a recent partnership with Neptune Networks. Through the pairing, Overbond uses axe data provided by Neptune to enhance its size-adjusted pricing. This information will be included in its analytics tools and trade automation suite, sitting alongside the AI-adjusted price aggression levels.

Vinod Jain, a strategic advisor at research firm Datos Insights, says the buy side wants greater transparency within the fixed-income market. For example, a buy-side firm can interact with five dealers on a trade, but they then have to take in all this information to decide whether they’re getting the best price. Tools that can boil down that information and present it back in an easy-to-visualize format are of great value to a trader.

“Overbond will now use axe data provided by Neptune unique to each client. On the equities side, you have a system on the prime broker, to a large extent, and they will provide that axe data to their clients. In fixed income, axe data is generally provided to buy-side firms in order to indicate at which prices the dealers are willing to buy or sell. Axe data being provided in this case by Neptune … could make this more interesting for a buy-side firm to execute the trade,” Jain says.

The AI-adjusted price aggression levels offering took more than a year to develop. It currently has two live production instances, Magdelinic says. 

The pricing is bracketed based on the number of bonds covered: 5,000 Cusips or Isins is the lowest, then 10,000, then 15,000, and so on. The euro and US dollar are available out of the box, but Overbond can add any extra G-10 currency and tune the model for extra, Magdelinic says. Overbond is also looking to expand into other more niche areas of fixed income, such as USD munis. 

Overbond also has a variety of different ways to deliver this information, including a direct API feed into the user’s order management system; a visualization widget within the OMS; an FDC3-enabled desktop component, done in conjunction with interoperability provider Interop.io; and a FIX API. “Wherever your native tech at the desk is, we’ll drive the integration as efficiently as possible,” Magdelinic says.

While it depends on the size and complexity of a desk’s trading strategy, Mehra says Overbond can ingest a firm’s historical data to build the model in about a week’s time. For new users that aren’t already connected to Overbond, it can take a few weeks more to build out a direct API or other data delivery option. 

“Whether a dealer offers a good or a bad price is, of course, subjective, but what the buy-side customer wants is consistency,” Mehra says. “They want to know that the price on the screen is what they are going to get. That is where—we believe—this could be of great value to them.”