Overbond applies machine learning and AI analytics to assess bond market liquidity

COPYING AND DISTRIBUTING ARE PROHIBITED WITHOUT PERMISSION OF THE PUBLISHER: SContreras@Euromoney.com

By:
Peter Lee
Published on:

The deadline looms for SEC-regulated investors to report on the liquidity of individual bond positions, but the more pressing question is the accuracy of fund valuations.

tech-city-digital-AI-780

June brings the deadline for US asset managers to comply with SEC rules first proposed in 2017 – and amended in 2018 – to implement and disclose their liquidity risk-management programmes.

The aim is to reduce the danger that customers of mutual funds, who may have assumed they could cash out their investments in all circumstances, find their money trapped behind redemption gates if funds are unable to dispose of positions in a market panic.

The rule requires every fund to classify each individual portfolio investment into one of four liquidity buckets: highly liquid, moderately liquid, less liquid and illiquid investments.

These buckets should take into account relevant market-, trading-, and investment-specific considerations, as well as market depth and whether sales of an investment would notably change its market value.

Funds must review these liquidity classifications at least monthly, if not more frequently.

Tough challenge

The rules also impose on open-ended funds a cap of no more than 15% of portfolios to be in highly illiquid investments, and require disclosure of the proportion to be held in highly liquid assets and so, in theory, be easily convertible into cash to meet redemptions.

Complying with this is going to be a tough challenge for investors in bond markets.

Liquidity or illiquidity is not a static feature defined by an issuer’s size or credit rating or by an individual bond’s remaining maturity or issue size. Liquidity can improve or degrade from one week to the next, from one day to the next.

However, the almost universal opinion among bond market participants Euromoney speaks to is that only a small portion of recently issued US treasuries are highly liquid: almost everything else in the rates market and in investment grade and high-yield credit is less liquid, and many corporate bonds are only infrequently traded.


COBI-Pricing depends on layers of models working together, some still with an element of supervision, but our aim is to develop towards full automation 
 - Vuk Magdelinic, Overbond

Regulators may expect periods of price change and volatility to tease out two-way flows in equities, with different investors taking opposing views on fundamentals, so allowing stocks to turn over in size except in an outright market crash.

However, that’s not how it works in fixed income.

For example, amid the bear markets rolling through risks assets at the end of 2018, secondary equity volumes picked up as prices fell. By contrast, secondary volumes declined markedly in high-yield bonds even as prices fell – which might have been expected to attract new money.

The fund management industry pressed the SEC to delay implementation of the portfolio classification requirement, originally set for the end of 2018, and the regulator has acknowledged the inherent difficulties in fixed income, given the lack of market data in the absence of stock-exchange equivalents for bonds.

Some market participants see an ulterior motive in pressing bond funds to classify the liquidity of individual positions.

Vuk Magdelinic-160x186

Vuk Magdelinic,
Overbond

Vuk Magdelinic, chief executive of Overbond, a provider of fixed income analytics, tells Euromoney: “Being required by regulators to report on whether each individual portfolio position is highly, moderately, less liquid or illiquid brings with it – quite deliberately on the regulator’s part in our view – a whole new level of heightened sensitivity around the valuation put on each investment and in aggregate on whole portfolios.

“Almost by definition, the less liquid a bond is the less confidence attaches to the pricing any holder puts on it in the absence of observed actual turnover in an active two-way market.”

Overbond was established in 2015 in Toronto, with offices also in New York, concentrating initially on enabling price discovery in primary markets through dissemination of target terms between bond investors and potential issuers. It also overlaid this with predictive analytics on new issue pricing.

In the past couple of years, what looked like an add-on has developed into a core offering.

Overbond has developed corporate bond intelligence (COBI), a collection of analytics that includes COBI-Pricing, a three-phase artificial-intelligence (AI) algorithm engineered to supply indicated pricing for secondary markets and new issues in varying market conditions for specific bond issuers, even in the absence of observed hard pricing data.

Magdelinic believes this may help bond investors comply with the new liquidity management rules.

“For the buy side, producing these portfolio classification reports frequently makes manual reconciliation quite unfeasible,” he says.

“That might be plausible if they only had to report once a year, but liquidity changes constantly and requires constant monitoring to report monthly. Compliance will require some level of automation.”

The SEC understands that, given the market data gaps in fixed income, virtually all fund managers will rely on technology from outside providers to assess liquidity of portfolio positions.


How COBI-Pricing algorithm worksCOBI-Pricing-Graph-780
Source: Overbond


Magdelinic says: “COBI-Pricing deploys machine-learning algorithms across publicly available secondary and primary bond market data, issuer specific balance-sheet data and proprietary data derived from the buy and sell sides.

“It aims to report a pricing output and a liquidity flag for a given issuer at a given tenor, which is the combination that regulators are asking for.”

Overbond’s growing customer base includes buy-side institutions with more than $2 trillion under management and the treasuries of over 80 bond issuers of different types and ratings.

Magdelinic explains: “To help us calibrate our pricing algorithms, we have asked treasurers for the indications they get from investment bankers proposing new issues of different structures and tenors which come anonymized and aggregated.

“Similarly, from the buy side, we request non-executed colour: given the valuations COBI-Pricing puts on their current portfolio, what secondary switches might investors be inclined to make or what primary investments would they like to put on, at what price and in which tenors?”

These are the kinds of feedback loops ­that classic AI algorithms can learn from and begin to identify and model correlations, despite the vast data gaps that characterize bond markets.

COBI-Pricing was not conceived as a regulatory reporting mechanism, neither as a fixed income trade execution algo. Rather, it is a relative value pricing tool designed to enhance market monitoring and decision-making for both issuers of bonds and discretionary portfolio managers.

Magdelinic says: “When we started, COBI-Pricing was at 60% to 70% precision in forecasting likely issuance activity and pricing on a three-day forward view. After we went to the buy side and the treasurers for their proprietary aggregated and anonymized data, we have improved that to just over 80% precision.”

He suggests there is a natural limit.

“It’s a remarkable level of precision, says Magdelinic. “It’s unlikely to proceed now from 83% to 93%; rather it might inch up incrementally, from 83% to 83.1%.

“COBI-Pricing depends on layers of models working together, some still with an element of supervision, but our aim is to develop towards full automation.”