Machine Learning, Artificial Intelligence, and Robo-Advisers: The Future of Finance?
The topic of machine learning–enabled artificial intelligence (AI) is gaining increasing visibility in the world of investment management. Of particular interest is the application of AI to the development of smarter robo-advisers that some hope, while others fear, will yield “intelligent” and cost-effective investment management advice. This topic was raised by investment professionals during recent CFA Institute traveling conference that went to Central and Eastern Europe and the Middle East. AI has also been the subject of a recent European Commission (EC) consultation document, to which CFA Institute submitted a response.
What Is Machine Learning?
For the uninitiated reader, some quick defining of terms could be helpful. Machine learning is a technique that researchers, and now firms, have begun using to design more intelligent computer systems. The advantage of this approach is that instead of attempting to hard-code a computer with all the possible scenarios and actions a machine may encounter when trying to perform a certain task, machine learning allows the computer to “learn” the necessary relationships and actions involved in completing a task intelligently. This “training” involves using a large training data set that the computer algorithm can repeatedly go through (but typically with guidance and supervision) to learn through trial and error how to connect the input data (e.g., credit history, employment history, assets, purchasing history) with the desired output (e.g., the correct identification of a suitable risky portfolio). Once this is learned, the algorithm can be used on real-world data with surprisingly good (in some cases) results.
The most popular approach to machine learning is through the use of neural networks, which are computer algorithms designed to mimic the way we believe the brain processes information. The networks are created by connecting millions, if not billions, of artificial neurons (essentially input/output switches) in all possible combinations. Through the process of machine learning (i.e., supervised trial and error), certain connections are strengthened, others weakened, and some removed until the correct network of switches and connections is left that can accept an input and identify the correct output.
The input data that can be fed into neural networks has only the restriction that it must somehow be converted into 1s and 0s that are palatable to a computer. In the world of fintech, common use cases for machine-learning algorithms are credit-scoring models for fintech credit companies (e.g., marketplace lending or peer-to-peer lenders) and robo-advisors. In the former, hundreds of input parameters about the individual (including, in some cases, Facebook and Twitter usage) are fed into a neural network that attempts to find patterns that are unobservable to humans but are nevertheless correlated with credit-worthiness. It should be noted that whether these types of neural networks really work has yet to be tested by a full credit cycle. In the case of robo-advisors, similar input parameters about the individual investor are considered by a neural network designed to assess suitability, risk tolerance, and appetite. Alpha-generating algorithms are a separate angle in which machine learning is being developed in the field of investment management.
We’ve Seen These Issues Before: Algo Trading and the Black Box Problem
One issue relating to neural network–based machine learning–enabled AI applications in investment management is one familiar to readers of this blog and its posts on market structure, high-frequency trading, and algorithmic trading. The black-box issue, in which the workings of an algorithm are not understood by its user or other stakeholders and lead to potentially unintended actions or consequences, is a well-known headache for regulators trying to ensure market stability. Although some attempts have been made to check the source code of algorithmic traders, the most effective protection against algorithmic errors are circuit breakers on markets that limit the amount of damage a failing algorithm can cause.
We highlighted in our response to the EC’s consultation document on fintech an example from the world of algorithmic trading on the use of circuit breakers as a de facto solution to the black-box challenge. In the case of neural networks, the black-box issue is even more complex. Although an algorithmic trader is, at least, coded by a human, a neural network evolves by itself in response to its training regime. This consideration is relevant for regulators seeking to, in some way, certify or monitor neural networks because only a limited understanding of the behavior of the neural network can be gained by simply observing its operation or codebase.
To see the extent of this issue, consider one of the more intuitive types of neural networks — convolutional neural networks used in object recognition. These take raw pixel inputs from an image and then “search” for lines and edges, then for corners and curves, and then for more complicated structures and object parts before finally identifying the object correctly (e.g., a dog). But, this example is a relatively rare case of intuitive neural network behavior because of the visual intuition that can be observed when analyzing the different layers of this particular type of neural network. It is far less likely that any intuition can be observed in a neural network that combines age, postcode, salary, Facebook status history, and number of Twitter followers into a credit score. For this reason, it is unclear what regulators would be looking for if attempting to approve or monitor these kinds of algorithms.
Even if it were likely that a regulator would have enough expertise and resources to evaluate, monitor, and police AI algorithms market-wide, it is likely to be a fundamentally impossible task because of what I just described. The EC consultation asked for opinions on whether a technologically agnostic approach to financial services regulation is desirable. This approach is something CFA Institute has previously advocated for. The development and adoption of new technology in financial services appears to be accelerating, and there is little chance regulators will be able to “pick winners” or craft tech-specific regulations in a timely manner. We think the only effective solution is to craft regulations that apply regardless of the medium of the service delivery and, like circuit breakers on equity markets, allow some degree of failure without contagion.
If you liked this post, consider subscribing to Market Integrity Insights.
Photo Credit: ©Getty Images/Just_Super