Over the past few years, investment managers within every market segment have considered how machine learning might, if at all, be added into the investment decision making process. Lured in by the potential for automation of decision making – and the hopes of enhanced speed, accuracy and juicier risk-adjusted returns – machine learning approaches and data science techniques have become a hotbed for research teams on both the sell and buy side of the market.
Man continues Reign over Machine for Earnings Estimates
Despite the volume of headlines and the swathes of research underway in this area, the exact details of the benefits these teams are extracting has remained reasonably elusive. Shedding light on this matter is recent research from Macquarie - which indicates that an entirely machine learning- based approach to estimating company earnings underperforms analysts’ fundamental based consensus estimates. Also, a quantamental approach (the combination of quantitative and fundamental) techniques, a relatively new field, also provided weak evidence of improving outcomes against the purely fundamental approach. In short, incorporating automatically generated estimates of earnings forecasts did not deliver improved results. 1
The team at Macquarie set out to build on the findings of a new academic paper by Ball and Ghysels (2018) which presented encouraging results - suggesting that a quantamental combination of forecasts is more successful than consensus applied in isolation. A particular motivation for pursuing this research further was the potential for adding significant value in the decision-making process. In the MiFID II climate we are now operating in, with sell-side research taking a considerable hit, automated decision making would be of enormous value to asset managers.
David Einhorn's Greenlight Capital was down 0.1% for the first quarter, underperforming the S&P 500's 6.2% return. In their letter to investors, which was reviewed by ValueWalk, the Greenlight team said a lot happened during the first quarter even though they made just a handful of changes to the portfolio and essentially broke even. Q1 Read More
While the research from Macquarie does not look great for quantamental techniques, keep in mind that the predictive model from Ball and Ghysels (2018) did beat consensus consistently in cases of high forecast dispersion and among small firms. This could present an opportunity for alpha identification – but with those parameters, alpha capture would be difficult to achieve.
Crucially the Macquarie research looks only into the ability to automate earnings estimation. 1 While Macquarie was left wanting, many top name asset managers have also been pursuing the approach towards alpha generation. BlackRock, Point72, Third Point and the Tudor Investment Corp have all been pivoting experimental research in this direction over the past three years. Big names they may be, but none have yet presented, or even hinted at, robust findings in this space.
2017 was not a good year for quantitative asset management - with low volatility and bull trend environments failing to provide the inefficient backdrop required for capturing opportunities through automated strategies. Many quant funds suffered. As a prominent example, the Clinton Group Inc., a quantitative hedge fund managing more than $2.9 billion into 2017, slumped close to 5.5 percent in 2017 after delivering consecutive year on year positive returns since its 2006 launch. 2
We might easily surmise that the market climate leading us out of 2017 and into 2018 will additionally prove far from ideal for continuing experimentation in this space. The jury remains decidedly out (and arguably down) on quantamental investment decision-making techniques.
1 Macquarie Research, Equities, Quantamentals, "Can machines replace sell-side analysts?" 26th January 2018