Machine learning is increasingly used to identify inefficiencies in small- and midcap markets. But as more investors adopt the same technology, it could make generating alpha more difficult.
Since the introduction of Mifid II in 2018, analyst coverage of small- and midcaps has declined, increasing inefficiencies in that segment. These less efficient parts of the market are precisely where investors see opportunities for technology-driven strategies.
Many smallcap investors now use AI or machine learning (ML) to identify opportunities and support portfolio managers’ investment decisions. Firms including Invesco, DPAM, and Robeco are using the technology to analyze larger volumes of data and more quickly identify promising smallcaps within existing quant strategies.
“We can do more with the same team,” said fund manager Michel Bourgon of DPAM in an interview with Investment Officer.
At Robeco, machine learning is used to dynamically determine which characteristics matter for individual stocks. Based on this, the model selects smallcaps that can be included in the recently launched NextGen Global Small Cap ETF. “The ML model sees non-linear dynamics that are easy to miss with the human eye,” said Mike Chen, head of the Next Generation Quant team at Robeco.
Nick King, head of ETFs at Robeco, points to the broader role of these tools in markets. “By exploiting inefficiencies, these models make the market more efficient.”
Bourgon has also observed this effect. “Investors can make better-informed decisions.” At the same time, he believes it may still take years before inefficiencies in small- and midcaps noticeably decline.
Generating returns in an efficient market
A more efficient market can make it harder for active investors to generate alpha. As technology becomes more widely available, competition typically increases. “Because new players enter the space,” said Chen.
Bourgon does not expect this to accelerate quickly. “There is currently relatively little competition in small- and midcaps. Many investors continue to focus on largecaps. For inefficiencies to truly disappear, significantly more funds would need to focus on smallcaps. I do not see that happening anytime soon.”
According to Viorel Roscovan, research director at Invesco, it is still too early to say to what extent AI can structurally reduce market inefficiencies, as changing market regimes and implementation challenges can dominate the picture. He points to the market conditions in which many models have been tested.
Impact of regime and data quality
In the years following the financial crisis, volatility in financial markets remained relatively low for a long time and liquidity was high. Due to accommodative monetary policy, risk premiums were compressed, leading investors to take on more risk to achieve returns. “It was precisely during that period of low volatility that machine learning became popular in asset management,” said Roscovan. “In such a regime, models can appear more stable than they really are, and that becomes clear when volatility and macro uncertainty return.”
Since the coronavirus crisis, volatility has been higher and macroeconomic uncertainty greater. As a result, he said, it is difficult to distinguish between truly structural changes in inefficiencies and changing market conditions and risk premiums.
There are also practical limitations. Roscovan notes that machine learning models often exhibit more momentum-like behavior. Momentum strategies typically have high turnover, which can be problematic in illiquid markets such as smallcaps. “As a result, backtests often look much better than what real-world trading shows.”
The technology can also identify patterns that do not actually exist. “There is still a lot of noise in the data,” said Roscovan. “Extensive evaluation and human oversight are required to distinguish real alpha signals from false results.”
Humans remain indispensable
Human judgment still appears crucial when using AI in investment processes. “We outsource the identification of return drivers to the machine, but not the responsibility,” said Chen. King emphasized that people remain essential to understand and monitor the technology.
For asset managers, understanding the technology was also a longtime challenge. Robeco tested and refined the model over several years using its own capital before building the NextGen Global Small Cap ETF, to ensure it could fully understand and explain the machine’s decision-making. Chen: “During that time, the technology moved from a black box to a glass box.”
Roscovan believes that machine learning must be handled with care. “The question is not whether we should use ML, but how we implement it so that robust, usable, and reliable models are created.”
If AI ultimately leads to a more efficient smallcap market, making it harder to generate alpha, the experts are not particularly concerned. “As an active investor, you can never rest on your laurels,” said Chen. “We must constantly reinvent ourselves to stay ahead.” King added, “We also want to be able to find the inefficiencies of tomorrow.”