Several machine-learning practitioners joined Ahcene Gareche, head of quantitative strategies at AXA IM Chorus for machineByte’s Hong Kong ThinkTank on June 19, 2019.
Participants debated the performance of factors over the last two years and discussed how best to use machine-learning trading techniques to trade options, use data to find optimal hedging strategies and build risk models.
More Data, More Problems
The quality of alternative data sets continues to be a major sticking point for practitioners interested in employing machine-learning tactics in their trading strategies. One participant noted that alternative data is most useful when applied to trend-following momentum strategies for estimating expected returns over a long time horizon or for high-frequency trading applications, but not much in the middle.
Regardless of the end user or strategy, alternative data are very hard to incorporate into models and mostly are “a waste of time” as they frequently contain simple mistakes and need extensive cleaning. Even if the data is good, the participant explained, it can still take him three hours to clean his data.
He continued that there’s a significant mismatch between the people who want to sell alternative data and pricing on the data sets, adding that he couldn’t understand what U.S. managers did with this data.
Data vendors have largely not made their way to Asia yet, but some Asian managers are buying data through U.S. firms or obtaining it in other ways.
The group discussed the pros and cons of several types of strategies, including genetic algorithms versus random forest for selecting optimal combinations of factors.
Genetic algorithms, which are inspired by Charles Darwin’s theory of natural selection, can be used for knowledge discovery, according to one participant. He is currently developing a GA algo that can look at the universe of factors and come up with the best combination of factors within a portfolio.
Genetic algorithms are superior to neural networks, he added, because GA doesn’t suffer from explainability problems while NNs do.
However, one attendee disagreed about how useful the algo was, saying that GA was extremely susceptible to overfitting, which has caused it to fall out of favor among many practitioners. Instead, he preferred random forest methods, particularly because it did a better job of fitting any tail within the data.
The first speaker noted that GA was primarily an optimizer and that he tried to control the algorithm via the inputs and methodology.
Attendees highlighted several other applications where they’re using machine learning, including portfolio construction, event-driven strategies and news filters to flag relevant news headlines.
Some asset managers with business in China are working in new ways to get around the current lack of data about the country’s businesses. One way is by identifying theories with potential and then drawing on institutional knowledge of how U.S.-based managers solve the problem. But since the micro structure of every country is different, managers tweak these strategies ported from the U.S. to fit their specific issues.
One participant noted that, because China has many state-owned enterprises, his firm doesn’t apply the same factors to each one, instead deciding to split the firms based on other criteria. The process still employs machine learning, but it’s a supervised approach.
No commitment 30 day free trial
We deliver education to quantitative professionals on the latest innovations and developments in machine learning in investment management and other industries, such as tech, healthcare, transportation and education.
Get the latest exclusive news, innovations and approaches at financial institutions, tech firms and companies across the whole spectrum of machine learning.