Andrew Lo, a professor of finance at the Massachusetts Institute of Technology, dialled in from Boston to talk about regulation, markets and the future of machine learning.
Lo pioneered the adaptive markets hypothesis, which describes markets as a complex evolutionary ecosystem, populated by different stakeholders that adapt to changes on the basis of certain behavioural traits and biases.
Regulatory reforms tend to tighten and loosen in cycles. When markets are doing well, regulators usually end up taking their foot off the pedal, allowing risks to build up again over time.
Lo spoke about how regulation can be enforced in an adaptive manner so as to keep the systemic risk of the banking system stable.
“[You] really require a deeper understanding of human behaviour, because ultimately that is the source of this kind of pendulum swinging from lots of regulation to very little regulation. Once we understand that it is really our own perceptions of risk and how we deal with it that really drive these kinds of swings, we can better manage those swings to try to attenuate them and develop a more stable financial system,” Lo said.
As a key first step, Lo highlighted the importance of developing ways to measure systemic risk in the financial system. “We don’t yet have an aggregate measure of system-wide risk,” he said, “and until we start measuring that kind of systemic exposure it is very difficult for regulators and policymakers to respond to them.”
Cryptocurrencies is another market where an adaptive view could help understand the price dynamics better, argued Lo.
“I think this is an area where having this kind of an evolutionary perspective provides some really interesting guidance, which is difficult to obtain in any other way,” he said.
He added that regulated exchanges alone will not help reduce the risks in this market, and that a broader involvement of central banks and regulators would be the way for cryptocurrencies to be accepted into the mainstream.
In the past, we have used machine learning to verify various predictions. But at some point machine learning can be used to actually formulate those hypothesesAndrew Lo
Lo also spoke about the usefulness of machine learning in making sense of data and applying adaptive algorithms to develop trading strategies. One key criticism from detractors is that with machine learning there is always a risk of data mining and overfitting. Lo argued one way to fix this is through rigorous testing.
“The only way we can reliably deal with this issue to some degree is to go into this exercise with some kind of perspective and world view about what it is we are hoping to find, and after the fact testing rigorously whether or not in fact that thesis is justified by the data we have collected. If it’s not, we have to basically reject it and collect new data to formulate a new hypothesis,” he said.
Lo added that the next major development in machine learning research is to use the technique to come up with the hypothesis itself.
“This is probably the next frontier for machine learning. In the past, we have used machine learning to verify various predictions. But at some point machine learning can be used to actually formulate those hypotheses,” said Lo.
Lo added his future projects involve combining behavioural science with computer science to encode human behaviour algorithmically. Many algorithms, such as portfolio optimisation and mean variance optimisation, show how people ought to behave, but how people do behave hasn’t been encoded algorithmically.
“Once we have those algorithms, we can then begin designing more sophisticated portfolio management tools and risk management tools to be able to help individuals deal with their own behaviours in a more systematic way,” Lo said.
This, Lo added, would require bringing together computer science, AI, machine learning, data analysis, neuroscience and psychology.
“We are now in a position where all these different fields are coming together in ways that are really exciting. I see lots of potential breakthroughs over the course of the next few years,” concluded Lo.
1:00 Adaptive regulation
4:20 How systemic risk can be measured
7:30 How adaptive markets hypothesis can be applied to cryptocurrencies
11:35 Quant funds investing in cryptocurrencies
14:46 Impact of the unwind of quantitative easing on the banking system
16:36 What happened during the Vix move of February 6
18:58 How machine learning can help in implementing adaptive algorithms
29:02 How can adaptive markets hypothesis be applied in trading strategies
31:50 Tools to avoid data mining
35:09 Can you learn from a machine learning algorithm?
38:47 Are machine learning algorithms black boxes?
45:50 Future projects – combining computer science with behavioural science
To hear the full interview, listen in the player above, or download. Future podcasts in our Quantcast series will be uploaded to Risk.net. You can also visit the main page here to access all tracks or go to the iTunes store to listen and subscribe.
- People moves: SocGen adds in prime services, Deutsche fills new rates hole, HSBC makes model move, and more
- Quant Finance Master’s Guide 2019
- Princeton tops inaugural Risk.net quant master’s ranking
- Credit risk quants are hitting the tech gap
- Does credit risk need an expected shortfall-style revamp?