Machine learning – technology that creates relationships between known data points and uses those relationships to make predictions on new data – enables computers to learn like a human, operating with pre-defined rules but adapting as more empirical evidence is gathered and interpreted.

BlackRock’s multi-asset analytics engine Aladdin uses machine learning to identify patterns in portfolios consisting of stocks, bonds, mutual funds, ETFs, options and structured products and the use of the technology within the asset management industry continues to expand.

In March, Acadian Asset Management announced that it would use the Bing Predicts macroeconomic indicators offered by Microsoft to augment its investment forecasting frameworks. Bing Predicts is a prediction engine that uses machine learning from data on trending social media topics, as well as sentiment towards those topics, along with trending searches on Bing.

Richard Bateson, founder of Bateson Asset Management and former head of MAN AHL’s Dimension fund describes his BAM Lexicon fund as being dedicated to applying artificial intelligence techniques to predicting and adapting to market changes in real time, initially in futures markets although there are plans to extend it to equities.

Rather than using a fixed formula or equation to calculate a trading signal (such as trends from prices), machine learning can combine diverse sources of data to provide a prediction of future price movements, he explains.

“It is the next generation of systematic trading. Since as it is not based on a fixed formula it can evolve with time to cope better with changing market conditions. The increased computing power available through big data can be leveraged to provide predictions simultaneously across many markets.”

What differentiates machine learning from more traditional quantitative approaches is that when new or additional data becomes available, the computer algorithms adjust to this information without human intervention, says 3Edge Asset Management chief investment strategist, DeFred Folts.

What differentiates machine learning from more traditional quantitative approaches is that when new or additional data becomes available, the computer algorithms adjust to this information without human intervention, says 3Edge Asset Management chief investment strategist, DeFred Folts.

“These models typically use parameters built into their programmes to form patterns for the investment decision-making process,” he says.

“Traders may construct systems that utilise machine learning in order to identify very short-term market trading opportunities. In addition, computer modelling is not encumbered by the human biases that can compromise investment decision making.”

Patrick Flannery, co-founder & CEO of capital markets software developer MayStreet, suggests that use of machine learning differs little from more conventional research methods in that the output has to be tested against a wide range of assumptions to ensure data isn’t missing and that it a valid output.

AUTOMATED PROCESSES

In the long term the biggest impact from machine learning on finance will be automating business processes, according to Flannery. “You can come up with a long list of functions currently performed by humans that can be improved and if not entirely automated, at least partially automated. For example, parsing and getting the terms for bond deals completed automatically where the fixed income reference data is often still generated by humans.”

As social media became ubiquitous, quant funds started to apply natural language processing to associate crowd sentiment with trend formation and data providers now offer a range of sentiment data. The challenge, suggests Lucena Research co-founder & CEO Erez Katz, is that there isn’t enough historical data with which to conduct thorough quantitative research.

“Quant research consumes historical data to predict the future. Since crowd-based sentiment is a relatively new phenomenon, model training and backtesting are still somewhat limited, so big funds use such signals with great interest but also with caution,”

Martin Froehler, CEO of algorithmic trading systems marketplace Quantiacs observes that while machine learning offers a broad set of advanced tools to detect patterns in historical data, this data is very noisy and mathematically pretty close to a so-called random walk.

The random walk theory suggests that stock price changes have the same distribution and are independent of each other, so the past movement or trend of a stock price or market cannot be used to predict its future movement.

“If not applied correctly, machine learning will find patterns where there are none and overfit the historical data,” says Froehler. “Overfitted models usually have very low, if any, predictive quality.”

The benefit for managers is that it is possible to search the data for patterns at a whole new scale, potentially detecting patterns that no human would find.

On the downside, machine learning methods typically have a lot of parameters or degrees of freedom to fit the model to historical data, which doesn’t always result in the best models on new data.

“It can also be tricky to formulate machine learning problems for financial forecasting, as standard objective functions such as the lowest error/ least deviation don’t work. They often return trivial estimations – such as ‘don’t trade’ signals – since the data is so close to a random walk.”

Flannery says the major challenges to applying machine learning to asset management are getting the right data into a format that is amendable to the downstream workflows and structuring the problem correctly.

“There are a lot of tools and methods that can produce good results, but they require data and structure,” he says. “The term machine learning shouldn’t be mistaken for something that can overcome bad research or business processes.”

“There are a lot of tools and methods that can produce good results, but they require data and structure,” he says. “The term machine learning shouldn’t be mistaken for something that can overcome bad research or business processes.”

There must also be sufficient relevant historical data to train the machine learning algorithms and extract statistically accurate predictions.

In most non-financial machine learning applications there are many contemporaneous data sets, but in financial markets there is only one history and the histories of different financial instruments are often closely related, particularly in risk on/risk off environments.

LIMITED UNDERSTANDING

Also, machine learning only really works if the current environment if similar conditions are contained within its historical training data and often fails on unexpected or unpredictable events, adds Bateson. “It has no concept – at present – of understanding, for example, the implications of a Macron win versus a Le Pen victory in the French presidential election.”

According to Henri Waelbroeck, director of research at multi-asset trade automation solutions provider Portware, the main danger lies in model risk. “In 2009 we all suffered the consequences of inadequate copula or probability distribution models, which seemed to work very well until the MBS implosion.

Machine learning techniques are also models that can be wrong in a way that is not apparent until something happens. In a world where asset pricing is increasingly driven by quant strategies, diversity of strategy is going to be essential and I worry that there might not be enough regulatory disclosure to track this.”

In this context the human element remains crucial. Edgar van Tuyll van Serooskerken, head of quantitative strategies at Pictet Wealth Management observes that any investment opportunity identified by the technology is studied by his investment experts before it is acted upon.

This is important, says Folts, because in a complex system with so many inherent uncertainties it simply is not possible to include every single variable that could impact a solution.

Folts notes that Mary Cummings of the Institute of Electrical and Electronics Engineers (IEEE) Computer Society says that allowing a human to coach a highly-automated system produces results up to 50% better than if the automation were left to its own devices.

Folts notes that Mary Cummings of the Institute of Electrical and Electronics Engineers (IEEE) Computer Society says that allowing a human to coach a highly-automated system produces results up to 50% better than if the automation were left to its own devices.

BAM Lexicon posted a +1.6% return in February versus +0.3% for CTAs (commodity trading advisors) and is up +1.2% year-to-date says Bateson, who also refers to data from Eurekahedge showing that artificial intelligence/machine learning (AI/ML) hedge funds have outperformed CTAs and trend followers consistently over the last five years with higher returns and Sharpe ratios.

“In addition, the AI/ML index is negatively correlated to the average hedge fund and has virtually zero correlation to CTA futures and trend-following strategies,” he adds.

However, in light of the observations of Folts and van Tuyll van Serooskerken highlighting the value of human analysis and Eurekahedge’s view that losses suffered following Donald Trump’s US election victory illustrate machines’ inability to always predict the future, the revelation that returns don’t always match the hype is perhaps less of a surprise.

However, in light of the observations of Folts and van Tuyll van Serooskerken highlighting the value of human analysis and Eurekahedge’s view that losses suffered following Donald Trump’s US election victory illustrate machines’ inability to always predict the future, the revelation that returns don’t always match the hype is perhaps less of a surprise.

So although the Eurekahedge AI Hedge Fund Index has largely beaten the average hedge fund, Bloomberg data indicates that it lagged the S&P 500 Index in 2012, 2013 and 2014.

It is inevitable that as the technology evolves and machines learn from their experiences, their ability to predict human behaviour and therefore factors affecting markets will improve.

Katz does not expect asset management to become completely automated any time soon, though. “Eventually, machines will dominate decisions but people will still need to apply their intuition and discipline in order to gain the upper hand,” he concludes. “We are still light years away from removing human cognitive reasoning from the investment decision process.”