Accurately assessing liquidity with explainable machine learning models

Being able to explain how machine learning models work has been a point of contention since the technology’s inception. Bloomberg is set to release further empirical metrics, at the end of this year, to bolster its liquidity models’ explainability. The metrics will initially be available via Bloomberg data feeds and featured more prominently on the Bloomberg Terminal liquidity screens later in 2024.

This is the most recent step in a long-running effort to get clients to trust so-called ‘black-box’ models in a field where those models enjoy a convincing edge over more traditional methods. But the journey has not always been easy.

Bloomberg has long proven to its own satisfaction through backtests that its machine learning liquidity models are more accurate than other more conventional approaches. And yet, in the early days, the firm struggled to convince clients to trust the new models in the same way.

“In early days, I would often dance around my words a little bit to not use the term ‘machine learning’, but instead would explain in detail exactly what the model does,” says Zane Van Dusen, who is the global head of Bloomberg’s risk and investment analytics data business.

Nor was Bloomberg’s experience unique. In the field of liquidity risk other firms, too – notably BlackRock – shelved machine learning models in the past, precisely because they could not explain the model’s inner workings. Today, regulators are pressing buy-siders to track liquidity risk more closely, though. Last September’s UK gilt crisis and the turmoil in US Treasuries in March 2020 have shown the damage such episodes can inflict. Bloomberg’s forthcoming metrics are intended to win over any doubters that still are unsure whether to trust Bloomberg’s black boxes.

“Optimal massage”

What Bloomberg calls its Liquidity Assessment Solution – in fact a collection of models for a range of asset classes – launched in 2016 with the first rudimentary use of machine learning to estimate liquidity risk.

The LQA engine seems simple. Users input metrics such as how much they wish to trade and how fast, and LQA tells them what doing so will cost. “The heart of what the LQA does is it allows you to understand the relationship between volume, cost and time for a security. You plug two of those variables into the model and it solves for the third,” Van Dusen explains.

Under LQA’s hood, the suite of models calculates an individual liquidity profile tailored to an individual security, learning how to weigh as many as 150 features that could contribute to an asset’s liquidity profile.

The features are straightforward, Van Dusen is at pains to emphasise. For bonds, say, they would include the amount outstanding, maturity and coupon, the number of trades in the past month, number of executable bids, number of liquidity contributors, price, volatility, and the bid/ask spread. “Nothing too esoteric,” he says. “We’re not putting any bizarre alt data in there.”

Van Dusen’s team does, though, refashion data into forms its machine learning models can make better use of. “As part of the machine learning process, we’ve figured out how to optimally massage the data so that we get better accuracy.” For corporate bonds, for example, the amount outstanding is converted to US dollars for all bonds globally, then the logarithm of that value is calculated to produce a final feature that is fed into the model.

“The machine learning algorithm is really good at figuring out, of those initial 150 features, which ones are the top 30 to 50 features that really make a difference in terms of how much it costs to liquidate,” Van Dusen says. “It tries to figure out the optimal weighting of each of those features.”

The team trains the model quarterly using historical data and calibrates it daily to current market data.

Naysayers

Early iterations met with scepticism, Van Dusen says. “In some ways, we were a little too early to machine learning. Because it was such a new concept in the industry, even though we had built something with great results, it was a tough conversation with clients.”

His approach to winning confidence, then, was to focus on the data and results before getting into the methodology. And the toil of going through the model validation processes of Bloomberg’s early clients helped the firm understand and explain its own models better.

“We had methodology papers that we shared with our clients that would be 25 pages long, but at the end of the model validation exercise, we ended up with a 200-page document,” Van Dusen says. Now market participants have become more familiar and comfortable with machine learning. But even so, many still welcome assurances that the models are limited in how far their ‘thinking’ can go.

“Explainability is about exposing as many raw metrics as possible,” says Van Dusen. Bloomberg’s forthcoming release of empirical data aims to show clients where output is driven by what the model has learned and where it’s primarily influenced by observable data.

The liquidity squeeze caused by the onset of the Covid-19 pandemic in March 2020 proved the value of machine learning in liquidity risk models, Van Dusen says. Before then, asset managers had thought mostly about a repeat of the global financial crisis. “Up until 2020, I would say the entire focus of the industry was Lehman, Lehman, Lehman – that was the scenario that everyone tried to simulate.” The gumming up of the world’s most liquid markets showed how far liquidity could become a problem in unexpected places. Suddenly asset managers were asking for models that could tell them how trading conditions might change in a far wider range of scenarios, Van Dusen says.

Regressing to the future

The modelling techniques have come a long way since 2016 when machine learning meant little more than linear regressions. Bloomberg first made use of deep neural networks to model nonlinear effects of liquidity in 2017 when it launched its model for US municipal bonds. It used a similar approach to model securitised products in 2018.

Most recently, it upgraded its original model for government and corporate bonds from a linear regression model to a deep neural network model in 2022.

Backtests, especially for sudden periods of volatility like March 2020, reveal the advantage of these newer dynamic models. Traditional techniques only offer a view on liquidity extrapolated from current observable data, Van Dusen says. Machine learning models are able to estimate accurate liquidity even when such data is limited.

The more conventional classification approach defines broad categories like ‘US investment grade corporate bonds’, ‘European high grade government bonds’, and ‘Asia-Pacific high yield corporate bonds’ and then assigns generic liquidity measures to each category. The key challenges with this approach are that it is not granular enough because liquidity is instrument specific, and it does not react to changing market conditions and company-specific events. For the rules-based/waterfall approach, a firm will work its way down narrow to broader data repositories to assess liquidity. Starting with recent trades for a specific bond, if that is not available, to trades for bonds from the same issuer, and if that is not available trying sector level and so on. Granularity is an issue with this approach too. Liquidity changes over time, new issues often trade at higher volumes, while long-duration bonds generally have wider bid/ask spreads. “Further, you risk underestimating liquidity if you only account for observed trades in the fixed income market. Estimating the gap between actual and available volume is essential,” Van Dusen says.

Machine learning models are especially useful for understanding the dynamic relationship between features of liquidity. The technique comes into its own, Van Dusen says, for thinly traded assets where liquidity risk is a “small-data, large-universe problem”.

“Asset classes like municipal bonds and mortgages get even more complicated because the data is even more sparse, and more complex. That’s where we had to evolve into things like deep learning models and neural networks to help us get better accuracy.”

Machine learning models have also proven more accurate in dynamically modelling extreme events – the times when ordinary assumptions no longer hold. Markets are full of anomalies
that can surface at such times, Van Dusen says – high yield names that trade daily and are more liquid than certain investment grade names that trade rarely, for example. “That is the type of thing that we want to capture. That’s real liquidity risk management.”

Above all, machine learning is better than conventional models at expecting the unexpected, Van Dusen argues.

Looking back at the March madness of recent years – the US Treasury selloff in 2020, the market volatility sparked by Russia’s invasion of Ukraine in 2022 and the banking turmoil in 2023: “These are all things that took the market by surprise. If you try to model liquidity based on what you actually see, you will severely underestimate meeting liquidity in many sectors.”

This article was written by Celeste Tamers and is reproduced from Risk.net.