Andrew Lo, finance professor at the MIT Sloan School of Management and quant investing pioneer, believes large language models (LLMs) can bridge the gap between fundamental and quantitative investing in a way that was unfathomable five or ten years ago.
While there are still a lot of “hallucinations” or misleading results in AI outputs, Lo said the finance industry is now at an “inflection point” in creating more effective ‘quantamental’ investment strategies which would bring together the best of both worlds.
Lo, who founded quant alternatives asset manager AlphaSimplex in 1999, said the biggest difference between quant and fundamental strategies, or the obstacle standing between their integration, is fundamental strategy’s lack of scalability.
“You don’t need a lot of quants to be able to manage billions, even hundreds of billions of dollars. It’s incredibly scalable,” Lo said during a presentation at the Machine Learning in Quantitative Finance Conference at Oxford University.
“Fundamental analysis has its limits, because it just takes time and effort to bring all of these disparate pieces of information, much of which is not numerical,” he said. Taking his own healthcare investment venture as an example, Lo said the average fundamental analyst can only manage to follow 20 companies and come up with around five trades a year, and the only way to scale up is by hiring more fundamental analysts – until now.
“It turns out that generative AI and specifically large language models, can now process both text and numerical information together,” Lo said.
“It’s not just a bag of words that machine learning algorithms, when first applied to do natural language processing, tried to accomplish. It is much, much more powerful, and that is really a sea change which will allow fundamental analysis to be scaled.”
While Lo acknowledged that no one has discovered how to specifically implement this scale-up yet, he is optimistic that a solution will come out in the near future.
“Can quant and fundamental analysts live happily side by side and be able to be more productive than either by themselves? That’s what I mean by ‘quantamental’ analysis,” Lo said.
But to make further progress on this integration, it’s essential that investors figure out how to conduct fair backtests of strategies involving LLMs.
“I can run a backtest now of ChatGPT using data from 10 years ago. But the problem is ChatGPT has all the information over the last 10 years,” he explained. “I need to make ChatGPT dumber by somehow excising the last 10 years of knowledge, because otherwise it can just use the data for prices over the last 10 years, and yes, ChatGPT can perfectly predict future prices, given future prices.”
“So here’s an idea [for the big AI firms]… create a financial product that allows you to license large language models by vintage – monthly large language models that are trained on data only as of that month and prior.
“You can’t necessarily replay fundamental analysis over time, but you can certainly now run the quantitative part applied to fundamental data in a perfectly legitimate fashion.”
Pattern spotter
LLMs also have the potential to upend one of the most debated corners of investing, technical analysis, and lead to huge implications in fields like currency trading where pattern spotting is still prevalent, Lo said.
Technical analysis refers to the method of examining past trading patterns of a security, commodity, currency or index to predict future price movements. Analysts decode the geometric shapes in graphs and charts that they surf through every day – ‘head and shoulders’, ‘double top’ and ‘double bottom’ and so forth – to identify signals and trading opportunities.
Technical analysis’s reliability is often questioned and according to Lo, the very practice of it is looked down upon elsewhere in the industry with some referring to it as “voodoo finance”. It stands in contrast to fundamental analysts’ close examination of company balance sheets and macroeconomic factors and hinges its forecast on hints from historical market data, which are also highly open to interpretation.
“Someone once said… fundamental analysis is to technical analysis, like astronomy is to astrology,” Lo said.
But at the heart of technical analysis is visual cues and pattern recognition, and Lo said AI has proven to be so much better at identifying them than humans.
“They can actually identify patterns that humans cannot even begin to comprehend. We use double bottoms, head and shoulders, triangles, because we’re used to those shapes. How many shapes can large language models identify?” he said.
“Large language models will become extremely good at trading foreign currencies among other asset classes.
“Because in foreign currencies, if you’re trading, there’s really nothing else. Not a whole lot of changes between 10:03 and 10:04pm or am in terms of the trade deficit. So what are you going to do? You need to trade based upon something. This is an area where large language models can completely revolutionise the field.”
The ultimate task
In three years, Lo is optimistic that AI will be able to perform the most important mission in fund management: satisfying the fiduciary duty as defined by the law.
Lo is currently conducting research around LLMs’ role as financial advisers and said it can already satisfy two out of three crucial requirements to provide sound financial advice: having expertise in the domain, and being customisable. The final criterion is developing trust and ethics, which Lo believes will be achievable in the foreseeable future.
The boundary between human and AI is not as defined as we think, Lo said, and he pointed to prior academic efforts that aims to understand intelligence as a whole regardless of their forms, including Norbert Wiener’s exploration of cybernetics and John von Neumann’s posthumously published study, The Computer and the Brain, that discusses how the brain can be viewed as a computing machine.
“My personal hero, Marvin Minsky… said ‘I don’t want to build a computer I can be proud of, I want to build a computer that can be proud of me’.
“I believe that today we are on the verge of building computers that can be proud of us. That is both exhilarating and absolutely terrifying.”
