Quant funds look to AI to master correlations

Machine learning shows promise in grouping assets better, predicting regime shifts

AI correlation

Relationships – whether between financial assets or people – are a complicated business.

Investors pore over asset correlations for clues as to how the behaviour of one security is likely to affect the behaviour of another. Now, buy-siders are seeking to harness artificial intelligence to control the sorts of changes in asset relationships that can unbalance portfolios, water down diversification – and, ultimately, crimp returns.

The work involves reappraising how securities should be grouped together; as well as peering into the future to foresee unexpected changes in correlation.

“If we can predict future regimes better, that’s helpful. If we can predict how the components of the portfolio will behave together from a risk perspective, that’s also helpful,” says Andrew Chin, chief risk officer and head of quant at AllianceBernstein.

AllianceBernstein is one of a number of firms, including Acadian Asset Management, Data Capital Management, Lazard Asset Management, Neuberger Berman and XAI Asset Management, using machine learning to try to improve their portfolio mix.

Traditionally, firms such as these rely on assumptions about the extent to which positions offset each other based on historical covariances – a statistical measure of how asset prices move in step. And sometimes those correlations break, reverse, or change, before reverting back – or never revert back at all.

That can mean diversification vanishes, jacking up risk without boosting returns. The effect can occur at asset-class level, as between stocks and bonds in early 2018, or between factors, sectors, strategies or instruments. Alternative risk premia funds were caught out last year when usually complementary trading programmes synchronised unexpectedly.

To combat this, firms are paying closer attention to “cross-effects” – the ways that apparent linkages between assets form and develop – as an alternative to the standard approach of using static correlations.

“Conventional linear models miss the cross-effects between factors,” explains Jack Kim, chief risk officer at Data Capital Management. “When you think of a linear factor model it can’t even capture something like value strategies working well for high cap versus low cap.”

One new approach the firms are taking is to regroup securities by risk profile rather than by traditional asset class buckets, using so-called clustering algorithms. The idea is to reflect more accurately how assets move together and so help determine allocations more precisely. A second is to build regime forecasting models that might see a correlation switch coming, enabling managers to adjust portfolios ahead of time. Here machine learning is helping create more accurate models than before.

A grip on unstable correlations could boost risk-adjusted returns dramatically. “When you look at the cross-effects in portfolios, that’s where there’s a lot of performance differences [between funds],” Kim says.

That said, the use of machine learning in portfolio allocation, as elsewhere, faces the challenge of explaining opaque models to clients and training the algorithms on often small amounts of data.

“Portfolio construction can become very computationally complex depending on the types of risk measures you apply, particularly as the number of assets grows,” says Seth Weingram, director of client advisory at Acadian Asset Management. “We can use machine learning to help process and illuminate patterns in complex or unstructured data. It can help in inferring the nature of non-linear predictive relationships. But applications must be tailored to the specific context; there is no one-size-fits-all approach.”

Birds of a feather

Data Capital Management is looking to create a next gen of clustering models using machine learning technology. The firm feeds data into the algorithm and allows it to work out clusters of its own, based on the risk and return profile of the asset. This gets to the heart of the allocation problem – knowing how the different holdings in a portfolio behave relative to one another in different conditions.

DCM tried several approaches. One preset the criteria for clustering, such as volatility, beta to US equities, or the yield curve. Another grouped assets using returns over different time periods, from a day to a year. A third looked at the response of assets to market shocks.

The exercise revealed unexpected linkages. For example: “During certain regimes, both gold and biotech stocks can exhibit heightened interest rate sensitivity,” Kim says. High yield bonds could be categorised as equity like. Low volatility, high dividend stocks could be designated as more bond-like. “These relationships may be missed in a traditional clustering scheme based on sectors and asset buckets,” he says.

Interestingly, when DCM tasked a machine learning algorithm to sort its holdings by risk profile rather than asset class, 30% of the securities changed classification. The firm’s high level of reclassifications lays bare the risk in sticking to conventional asset labels.

“If you are running a risk parity strategy and you’re investing 60% in equities and 40% in bonds, up to 80% of your 40% may be intrinsic equity risk depending on the credit quality distribution of your bond portfolio,” Kim says. Such a portfolio would be far harder to unwind in a period of heightened equity risk or a high yield crisis.

Lazard is using clustering techniques in similar ways, even asking itself whether high yield should be thought of as a standalone cluster at all, says Jai Jacob, managing director leading the company’s multi-asset investment team. The pressure to adjust asset allocation through time is reduced as a result of more precise categorisation at the outset, he says.

XAI Asset Management is incorporating machine learning into the firm’s portfolio optimisation process. The system takes in XAI’s inputs and decides portfolio weights for its various strategies, adapting as it goes. “The system is not really a big data-driven system because we have relatively small datasets coming in,” says Aric Whitewood, co-founder of XAI, and previously head of data science at Credit Suisse. “What we have under the hood is a neural network. But it’s not a very deep one.”

Other firms are trying to spot when regime shifts loom that could foul up correlations. Chin at AllianceBernstein says regime forecasting models have long been common but newer, better ones employing machine learning are in development.

Portfolio construction is an area where transparency is particularly important to us, because we have to be able to explain to our investors how the things that we look for in our investment process are actually getting into the portfolio

Seth Weingram, Acadian Asset Management

Predicting what a coming regime will look like gives buy-siders a better sense of how correlations are likely to develop, since these linkages are often regime-dependent.

“In the past, we’ve tried to do that. But our models were too simple, so we didn’t spend a lot of time on it,” Chin says. “Now, at least on paper, machine learning models seem to be doing a better job.”

Firms struggled before to incorporate central bank sentiment into these types of models, but now they are attempting to parse central bank statements for predictive signals using AI algorithms. The same goes for tweets from politicians. “Much of the new research centres on pulling unstructured data to put into these models,” Chin adds.

Neuberger Berman Breton Hill, meanwhile, is using machine learning to figure out the optimal mix of factors that would reduce risk in its portfolios without curbing returns. The aim is to make short-term adjustments to factor exposures, says Ray Carroll, chief investment officer of the fund manager’s quant group.

“One of the big ways to control risk is to make sure you have the right mix of underlying factors across the portfolio. We’ll have a longer-term strategic mix that we think is the right way to risk-control the portfolio in the long run. But we also take tilts around that, which are data-driven tilts.”

The firm’s machine learning models might indicate to tilt towards value, say, if they determine the assumptions behind Neuberger’s long-run factor mix have broken down.

Whitewood concedes machine learning doesn’t help all the time. But the technology can find patterns not easily detected by simpler linear models. And often those patterns crystalise at critical moments. “In our experience [those times] more likely correspond to periods with higher volatility, where you can either make or lose more money,” he says.

Some explaining to do

Of course, machine learning brings risks of its own. While machine learning models are better at capturing relationships that change over time, it is often difficult to explain their inner workings. That leaves quants balancing the power of some of these frameworks with the ability to interpret them.

“I can use a machine learning model that uses 50 variables, some economic, some technical, and the machine will somehow optimise those variables to create a better forecast of the regime,” says Chin at AllianceBernstein. The word “somehow” is telling; it suggests that the alchemy at the heart of many AI techniques is, as yet, part mystery to many practitioners.

Acadian this year erected guardrails around the use of AI to avoid getting hurt if the algorithms malfunctioned. The firm also points to the glaring risk of using unexplainable models specifically in the area of portfolio allocation.

“Portfolio construction is an area where transparency is particularly important to us, because we have to be able to explain to our investors how the things that we look for in our investment process are actually getting into the portfolio,” says Weingram. “If you do something with portfolio construction that disrupts the ability to have that conversation, that’s a problem.”

Another difficulty, specifically for regime-switch models, is gathering data for the algorithm to train on. “Where you really get paid is where you have a shift in the data, and the market determines that something that wasn’t risky yesterday, is risky today,” says Paul Moghtader, a portfolio manager at Lazard.

Those types of events are difficult to predict for machine learning models simply because there are so few shifts like that in history to learn from. “It’s difficult for the algorithms to have enough data. It is crucial to have a solid fundamental understanding and to interpret any output alongside other analysis,” Moghtader says.

Equally, algorithms that focus on pattern recognition can struggle to extrapolate beyond the short term. Jacob at Lazard compares this to weather forecasting. “It’s quite easy to predict what the weather will be in five minutes, much harder to predict what the weather will be in five years,” he says. “There are broad distributions that you can make assumptions about, but in terms of specifics, it gets hard.”

Some managers, like Jeroen van Oerle, a portfolio manager at Robeco, have found it hard to forecast regime switches or big moves in factors, often called style rotations.

“Style rotations such as we’ve seen in recent months and the severity of those – mostly driven by ETFs and high-frequency trading – I don’t think you can predict them with AI. These kinds of things are so different from the past,” Oerle says.

That said, Robeco isn’t giving up entirely on the approach. “We’re not convinced that the quality of the algorithm is robust enough to trust in it. But we do use it as an overlay to help in the decision-making process,” he says.

Back at XAI and looking to the future, Whitewood says clustering techniques can be applied to do more than reshuffle portfolios. Buy-side firms could create baskets of securities that respond in a targeted way to movements in given markets, indicators or factors, he argues.

Already buy-siders are investigating the creation of clusters with enhanced exposure to factors such as momentum, he says.

AllianceBernstein’s Chin also sees a further potential for using machine learning in portfolio allocation.

“You can imagine using machine learning to go through different combinations of assets for the portfolio and play out different scenarios to see which portfolios are most robust,” he says.

Editing by Alex Krohn

  • LinkedIn  
  • Save this article
  • Print this page  

Only users who have a paid subscription or are part of a corporate subscription are able to print or copy content.

To access these options, along with all other subscription benefits, please contact info@risk.net or view our subscription options here: http://subscriptions.risk.net/subscribe

You are currently unable to copy this content. Please contact info@risk.net to find out more.

You need to sign in to use this feature. If you don’t have a Risk.net account, please register for a trial.

Sign in
You are currently on corporate access.

To use this feature you will need an individual account. If you have one already please sign in.

Sign in.

Alternatively you can request an individual account here: