Separating stock market signals from noise | Alberto Rossi, Georgetown University

David Sol: Alberto, welcome. Great speaking with you. You're a Professor of Finance at Georgetown University. You're also the Director of AI Analytics and the Future of Work Initiative, and soon to be a Senior Resident at the Brookings Institute, so congratulations with that. You've published an awful lot on machine learning, Robo Advisory, and FinTech. What is actually Robo Advisory and what does it actually mean for the financial services market? Alberto Rossi: So we really define Robo Advisory as any sort of advice that is generated by an algorithm and then that is tailored to clients' characteristics. And these kind of Robo Advisory services became popular, I mean, quite frankly, there were kind of initial forms of Robo Advisory already in the 90s and early 2000s, but they became more and more popular after the financial crisis around 2011, 2012, and so on and so forth. And there is a great promise for these tools. We know that a lot of the population is cut out of financial services or human financial advice when it comes to investing their own portfolio because human financial advisors have what we call an opportunity cost of time. They will want to cater to investors that have a lot of assets under management. And so through automation, you can imagine that you can not only improve the type of advice that you give to individual investors, but you can also expand dramatically the number of people that can. David Sol: So reaching more people at a lower cost. Alberto Rossi: Exactly, which is what we generally define, kind of, financial inclusion aspect of Robo Advisory. That's what excites me the most about Robo Advisory services. David Sol: And what is it that your research looks at in the Robo Advisories particularly? Alberto Rossi: So most of the work we do is trying to understand, first of all, what are the effects of Robo Advisory services on investors' portfolio locations, so how do their portfolio changes, how we can reduce some of the mistakes that people make. For example, one of the typical mistakes that American investors, or retail investors, tend to do is to be overly exposed to US equities, so something that we call home bias puzzle. And Robo Advisors are particularly good at exposing investors through ETFs through international markets. David Sol: So more globally and more diversified. Alberto Rossi: More diversified portfolio. So their reduction in risk is one of the crucial aspects. And then other aspects that I study when it comes to Robo Advisors is trying to understand for example, what is the optimal way of providing Robo Advice, right? One thing that we know is that you can have very good forms or very good algorithms advising individuals, but if they don't trust them, of course, it's going to be all done for nothing, right? So in many cases, you have that asset managers have started to develop, kind of, hybrid forms of advice where you have that some of the tasks are delegated to the algorithms, and some of the tasks instead are still retained by a human financial advisor. So effectively you have that these algorithms are there to empower the human financial advisor in the sense that it allows them to, first of all, become more effective, kind of cater to more clients, and so reduce the cost of providing it. David Sol: So you believe that there's an ideal combination of both human engagement as well as Robo Advisory to instill trust. Is that the best way of going about it? Alberto Rossi: Well, right now I would say, yes, a lot of the forms of Robo Advisory services that have been deployed to kind of use this hybrid structure, and this hybrid structure has been particularly effective. We have some research showing that the quality or the amount of interactions between the human and the client really affects how much trust the individual clients have in the algorithms, allows asset managers to retain more customers during market downturns, which is usually typically the time where a lot of the clients tend to kind of quit Robo Advisory services. But of course, this is a very dynamically kind of evolving space. You have that large language models are coming into play. You have that the humanisation of some of these chatbots, is potentially, can create or help instill trust. It remains to be seen how much we're going to be able to introduce these new technologies when it comes to Robo Advisory services. David Sol: Can we switch, Alberto, to another piece of research? Very interesting what you've done in machine learning, and it makes me think about my time as a portfolio manager and an analyst doing fundamental analysis, trying to understand companies, looking at the cash flows, price earnings ratios, to somehow come up with, or a way of differentiating, good companies from maybe not-so-good companies. You've looked at signals and looked at more than eighteen thousand different signals. So tell me a bit about that research. Alberto Rossi: The way we created them was, just to, by effectively thinking about how a fundamental investor would have done, which is basically going, parsing through all the financial reports and construct, and all the accounting information related to all those companies, and try to come up with all the possible, kind of, accounting ratios. David Sol: So, price earnings ratio, momentum, book-to-market, but then eighteen thousand of them. For which you definitely need AI. Alberto Rossi: It starts becoming relatively heavy from a computational perspective. Some of the models we train took almost thirty days on a thousand core. So it's like having a thousand of your MacBook Pros next to each other, crunching the numbers for almost a month. David Sol: And you looked at financial data over several decades- Alberto Rossi: Yes David Sol: Trained those signals. eighteen thousand signals. What was it telling you? What were the insights from it? Alberto Rossi: The point of our work was the following, is that what we've seen is that over the last couple of years, a lot of researchers, a lot of asset managers have started to use these kind of signals that were popularised over the past couple of decades, like book-to-market, dividend-price ratio, or momentum. These were all signals that people have discovered to be useful in the past. And so if you start using those signals and you feed them into a machine learning model, you're kind of cheating, right? Because you basically are exploiting information that you would now have known were you an investor manager. David Sol: Selection bias. Alberto Rossi: Exactly. You have selection bias and a look-ahead bias. So these prior studies show phenomenal performance. You have, like, annual returns of sixty percent, effectively, like, printing money almost. But then what we said is like, look, if you really want to take this seriously, you have to go back and kind of pretend you didn't have any information as to what signals could be useful or what signals could not be useful. And so you have to create the universe of them and then you have to feed them all into a machine learning model. And what we do find is that once you kind of do that, the performance is not as striking. You basically have that these machine learning tools are effective, but they will give you an outperformance compared to the market of maybe, I don't know, five to ten percent per year, which once you start introducing also implementation costs and all the aspects related to market impact for different trades, it shows you that some of the promises of these tools may not be as outstanding as people would argue. And I think also another aspect that I think is very important is that there is a little bit of this narrative out there that you can feed a machine learning or AI model anything, right? If you think about Open AI, that you would say, okay, look, we basically got the whole internet, kind of read it. David Sol: Whole library. All the library. Every single piece of information out there we fed it into these foundation models and look, we have large language models that are pretty effective. To a certain extent, you can feed a lot of the information. These tools are designed to discard information that is not relevant. So some of the signals there may not be relevant, but this is not as powerful as people would expect. David Sol: So still rubbish in, rubbish out. You need to look at what, how do you feed your models, what is the input data, and what's your objective? Alberto Rossi: Exactly. So you have to be kind of careful and very thoughtful about what we call feature engineering, right? So trying to feed your model the, kind of, the right amount of inputs in the right proper form. Because if you don't do this, you're going to have that there will be, they will be able to express some signal from all the noise that there is in the data, but they're not going to be as effective as you would want them to be. David Sol: So your research is in essence, also a cautionary tail for people using AI and misunderstanding these models. Alberto Rossi: The research we have is basically informing all those asset managers, like pension funds, or endowment funds, that are sometimes kind of investing in hedge funds and so on and so forth, that may kind of advertise phenomenal returns, and I'm not saying that you shouldn't go for those. What I'm saying is that you should be very cautious as to what are the kind of the signals that these hedge funds are using or some of these other investment managers because you should not be expecting these incredible returns going forward. And even though in the backtests that they present to you, they will show very, very large outperformance compared to the market. David Sol: So AI is of course a very hot topic, also in the financial services industry. Now, if I was a Chief Investment Officer of an asset management firm, if I would have to build a strategy around AI, what would you advise me? Alberto Rossi: I think that the crucial aspect is to be cautious in making sure that you, kind of, are careful about backtesting, making sure that you try to eliminate all the look-ahead biases that implicitly we all do whenever we go and use historical data to train our models and then be as complete as possible, like, try to make sure that you are always preparing for the worst-case scenario, not necessarily pick the one model that has performed extremely well because maybe of a couple of trades over a certain period of time. David Sol: Well, Alberto, thank you for sharing your insights from your research. It's been a pleasure speaking with you and I’m looking forward to reading some future research on AI going forward. Thank you. Alberto Rossi: Thank you so much, David. Thank you.