BlackRock to use machine learning to gauge liquidity risk

Firm close to rolling out new models for redemption risk and market liquidity

BlackRock is turning to machine learning to better understand liquidity risk.

Over the next two months, the asset manager will incorporate internal trade data into its existing market liquidity model, and apply machine-learning techniques to more accurately calculate the cost of liquidating fund positions in the case of redemptions.

“Liquidity is multi-dimensional and impacted by so many features. It is highly non-linear. So this is a typical [use case] for neural networks,” said Stefano Pasquali, head of liquidity research at BlackRock, speaking at Risk.net’s Quant Summit in New York earlier this month.

To date, the industry has mostly used the bid/ask spread to calculate the likely cost of liquidating positions, but BlackRock’s new module takes into account time to liquidation, transaction cost and volume.

Vendors such as Bloomberg, MSCI, State Street and StatPro are developing similar models.

The fund manager is also aiming to use machine learning to better assess the probability of large net flows out of its funds. This is being tested on 65 funds within BlackRock in the municipals, fixed-income multi-factor, high-yield corporate and total return sectors.

In a recent backtest on a high-yield bond portfolio, the model did a good job of identifying five out of six extreme redemption events over the course of 250 days and identified two further redemption events that were false positives, although BlackRock did experience redemptions in other high-yield bond portfolios on those occasions.

We moved to neural networks because, when it comes to the dynamics of fund flows, there are two very different embedded regimes – normal flows and gigantic flows – and that’s what we’re trying to forecast

Stefano Pasquali, BlackRock

Existing regression models failed to accurately model tail-risk events and had poor out-of-sample testing, Pasquali said. The new model uses a non-parametric neural network, which allows the fund manager to incorporate more than 300 factors into the model, such as fund returns, high-yield sector flows and total sector returns. The firm is researching whether use of a news sentiment index might improve calculations of the probability of redemptions.

“We already have a regression model trying to forecast outflows but it doesn’t work very well out of sample. We moved to neural networks because, when it comes to the dynamics of fund flows, there are two very different embedded regimes – normal flows and gigantic flows – and that’s what we’re trying to forecast,” Pasquali said.

BlackRock aims to build a unified liquidity risk management framework that combines market and fund liquidity, which can be used to support the fund manager’s risk management, trading support, portfolio construction and regulatory reporting activities.

The framework will also help BlackRock comply with the new liquidity risk management rules from the Securities and Exchange Commission (SEC), which are due to come into effect on December 1, 2018, and aim to reduce the instances of funds being unable to meet redemptions due to illiquidity.

The SEC’s rules require US mutual funds to classify investments into four liquidity buckets ranging from highly liquid to illiquid. One of the reasons BlackRock is turning to machine learning is the difficulty of complying with the new rules using existing techniques, such as time-series analysis of historic fund flow data.

As well as using the framework internally, BlackRock also plans to make it available to other funds via its Aladdin Risk Management tool. The fund is conducting calls with clients to update them on its research.

A chief risk officer at a US mutual fund that uses Aladdin said: “To the extent BlackRock gets this solution up and running in time – and if after our evaluation we’re comfortable – I suspect we will endeavour to switch from our internally built system to their system and then use our system as back up.”