Cliff Asness, chief investment officer and founder of quantitative hedge fund AQR Capital Management, has reservations about the value of machine learning, particularly in cases where limited data is available.
Speaking at the Ritholtz Wealth Management and the Information Management Network’s Evidence-Based Investing Conference in New York on November 2, Asness said there was a clear split between the young and older generations of quants at AQR as to whether machine learning was the paradigm shift it has been heralded to be.
He confirmed AQR is currently exploring the applicability of machine learning and artificial intelligence to various trading and data analysis problems to find patterns in large, unstructured datasets, but stops short of embracing the technology as a cure-all.
“Machine learning worries me,” he said, adding that on data mining: “We’ve been handed a H-bomb instead of an A-bomb.”
Asness pointed to satellite imagery as an example where the amount of data available is not sufficient to draw meaningful conclusions.
However, AQR is exploring the applicability of machine learning to areas of trading where the hedge fund has no strong view, but “tons of repeatable data”, such as determining when to roll futures, Asness said.
The firm is also using decision trees to better forecast the default risk of companies – the technique allows AQR to predict the probability of a corporation defaulting or migrating to a lower-quality credit state based on ‘non-linear’ input variables such as financial statement ratios and industry-specific data.
The work is informing AQR’s factor-investing strategies in the credit markets: the firm says forward-looking default risk is the “fundamental anchor” of the value premia in corporate bonds.
Asness said ultimately the proliferation of machine learning has little impact on AQR’s bread-and-butter strategy, which is rooted in the Fama/French factor model. Machine learning is unlikely to change the fund’s preference for the value and momentum factors, for instance, he said.
“We have satellite data of what stores people go into, but it doesn’t really interact with value, momentum or volatility,” he said.
Asness was also critical of the industry’s tendency to lump together quantitative or systematic factor-based strategies with big data strategies, arguing the latter tend to have high Sharpe ratios and alphas that decay quickly, as opposed to systematic value investing that has been proven academically to endure over multiple business cycles.
“Machine learning is another type of quant strategy but it doesn’t overlap with the factors,” he said.
Correction, November 3: This article was amended to clarify Asness’s distinction between machine learning and data mining
- Asia moves: BlackRock picks new Asia head, Credit Suisse boosts regional solutions, and more
- We need a different approach to supervisory stress-testing
- Bank risk manager of the year: UBS
- People moves: Asia hires at Credit Suisse, new UBS data role, NatWest takes UBS's Duclos, and more
- Risk solutions house of the year: HSBC