In this 10‑minute keynote at the IDSS conference, Dr. Jake Xia of Harvard Management Company outlines a four‑part view of data—collection, storage, processing, and prediction—and argues that finance must move beyond traditional price‑only datasets toward richer behavioral and causal information. He illustrates the challenge with concrete examples such as evaluating hedge funds and trading strategies, noting the multidimensional complexity of performance, personnel, and strategy shifts. Xia critiques the industry’s reliance on historical regressions, calling for new theoretical frameworks that can handle the growing volume of non‑price data, including social media and human behavior signals. He discusses the endowment’s unique long‑term horizon (aimed at supporting generations of Harvard), its liquidity needs, and the tension between diversification goals and the convergence of institutional investors into similar asset classes, which erodes diversification benefits. In response to audience questions, he acknowledges the scarcity of commercial behavioral‑analytics products, emphasizes the need for academic research, and differentiates Harvard’s internal, research‑driven approach from the Yale‑style outsourced model. Throughout, Xia balances humor with a serious call to action for both practitioners and scholars to develop better causal models and data‑driven investment insights.
Chapters
So thank you, John. Thank you, Amanta and Andrew, for inviting me to this IDSS conference. So as we speak, this is a data conference, and I'm the last data point on your calendar keeping you away from your family starting the weekend. So I'm actually quite used to that, you know, with the last name starting with an X. That's why I named my first child with the first name starts with A, to hopefully someone count the first name first. So being the last one speaking, the upside is a lot of the things already been said, but the downside is what's the insight I can add? So I feel some kind of a pressure to come up with some new ideas. which I'm struggling with. So by listening to all the talks, I said, okay, maybe there are a few things I can think of. Hopefully will be some food of thoughts to, you know, push us forward and for the future IDSS work. So when we talk about data, when I think about data, I really think about 4 aspects. So, I have no slides, by the way, so that hopefully will not go over 10 minutes. So, data is about collecting and the storage and the processing. The last thing is really, I think, the most important thing is about prediction. So, many said, you know, data is not useful until you turn into actionable insights.
Yeah.
But how do we turn it into actionable insights? So we have to figure out the causal relationships to uncover them, make the prediction work. So that's my aspect of how I look at the data. In my work at the Harvard Management Company, we obviously use a lot of data, and we do all these things. So one of the important questions Let me give you an example. How do we evaluate a hedge fund? How do we evaluate a trading strategy? So, we receive a lot of emails, a lot of phone calls, a lot of pitches. Some student just won a competition. What's the right word? Challenging contest. With money as an award. And how do we judge it? You know, also on the other hand, with the hedge fund had the 20 years successful record, should we put money in that hedge fund? So this is about data, but this goes beyond, way beyond the data. How do we look at the information, for example, goes beyond prices, return history, and all the context related to that data. So that's what I'm focusing on, and hopefully that provides a different angle. So I'm not going to claim we know all the answers. So we are searching. We are searching for methods to find these causal relationships, make us predict better. But I just want to comment on bit on the research direction, since this is about big data, and since John said I actually have a degree in this department, so I need to sound technical. And so, I feel in the last 40, 50 years, in the financial industry, when people, when we collect the data, we are very much focused on collecting price information, price data. So obviously there's a lot of work in the theoretical framework proposed which helped us to price options and many other things. It's all a lot of quants when they started in financial industry, including myself. I was a pricer. I was a pricing model builder. But I think now this has to change. Because the data is richer than that. We've got to focus on behaviors. We need to collect human behavior data. We need to have models which can understand better about human behavior. So that goes way beyond price data. Some of the previous talks mentioned social media, collecting data in Twitter and keyword searching text algorithms, and I think particularly I felt Sandy's talk was very intriguing, and I don't know if Sandy's still in the audience. I'd love to chat with him more on that, but I feel we are just seeing the tip of the iceberg of a lot more data coming to us. A lot of times I feel we have much more data than actually we can process. Why? Because we don't have the right theoretical framework to understand it. Simply doing historical regression, finding traces of patterns, it's not enough. So just like doing physics and engineering, I think it's a point of time we need to— I urge especially people in academia spending more time truly understand What the context, what is the related behavior around the price movements, and after all, it's the people, it's the subjects who actually move the price. So that's what I'd like to leave you with. So I'd like to save more time for questions if there's any, but if there's no question, the good news, you can start your weekend sooner.
Thank you.
Is that okay?
So you mentioned that one of the parts that you have to do at the management company is assess hedge funds. I've been surprised when I've met with people at hedge funds that with a few notable exceptions, the level of data analysis and statistical capabilities are remarkably thin, typically less than, say, the average baseball team's. Did you observe the same thing?
It really depends. So I think we've seen— I mean, we obviously have a lot of people who have very sophisticated statistical tools and models. I mean, sometimes even probably quite comparable to universities' statistics departments, but they obviously have different sources of data and they focus on different usages. So I think the challenge is when you look at the hedge fund, there are a lot of variables. The people there change, the strategies shift, and the positions also vary. And to draw a conclusion on the past performance, it's a multidimensional complex problem. And it's not a study can be put under controlled environment. It's very hard to do.
Okay. Audience questions? You can't all be that anxious to get out of here. Yes, there's a hand up in the middle. In terms of behavioral analytics— Why don't you wait for the microphone? It's on the way. In terms of really deep-dive behavioral analytics, not that I'm necessarily pushing a vendor, but we've done some work with a product from— called Digital Reasoning, which sort of does start to try to get into fraudulent behavior activity much more deeper than good, bad, neutral. Are there any other products on the market other than self-developed products for inferring behavioral analytics that you'd either recommend or have any interest in?
That's a tough question. So, I'm not aware of, and so I think there are a lot of research work being done, but not in the product form.
met.
Thank you. Question in the back there. Please correct me if I'm wrong, but I believe you guys and MIT and Yale all follow the Yale School of Investing.
How does big data inform your analysis of how that's been doing over the last, you know, since 2008? First of all, I'd like to claim we are not following the Yale model. We like to say, you know, there's a Harvard model, which differentiates us from the rest of the endowments. It's really we manage significant amount of money internally, which Yale and MIT all follow sort of kind of outsourcing external fund-of-funds model. And so, but you have a very good question there that I think you're referring to is kind of the convergence of the crowd behavior, which is very true that the endowment evolved over the years starting from purely simple safe bonds, U.S. Treasuries, expanding to stocks, international stocks, blue chip, then emerging, then private equity, real assets, farmlands, and so on. So it's becoming very complex, broad range, for the purpose of diversification, but in the end, when all the big pensions, foundations, sovereign wealth funds all move into the same direction, So that diversification benefit is diminishing, which is very much my concern too. When the market actually goes through like 2008, as we saw, everybody lost money. And so I think that's the challenge to the real money, the real asset management industry. How do we truly diversify? and finding new investment opportunities, given there's so much money chasing for return at this point.
Question kind of along the same lines. As an endowment fund, you could have a very unique position in the market in the sense that you've got a longer time horizon than traditional mutual fund companies you've heard of from today. Can you talk about both time horizon standpoint and the fact that you've invested a lot more as an endowment in alternative assets, private equity, hedge funds, other kind of aesthetic investments? To what extent can you use analytics, big data, sort of modern-day technology differently than if you were managing a straight public equity fund?
Yeah, that's a great question too. First of all, on time horizon, so we like to say we have 10+ years, and, but it's actually longer than that. So the purpose of the endowment is really to support multi-generations of high education, research, scholarships. So Harvard has a history about 380 years, so we like to say we are targeting for another 380 years. But obviously nobody can see that far. And so we do have the longer time horizon, which is supposedly give us flexibility or tolerance of volatility. But in practice, in reality, that such flexibility is very limited because we spend— every university is pretty much in the same situation. About 5% every year to support the university's operating budget. I think typically the large endowments nowadays support 35% to 40% of the annual budget. So the endowment needs to be reasonably liquid and has limited tolerance for downside risk. So with that, we have a target to beat, essentially inflation. You have to grow your assets faster than the inflation in order to support future operations. So with those combination, the combination of those goals, and so your room of operating becomes a lot smaller, but we still like to look at longer time horizon rather than short-term public markets trading. So it's still a very different approach. With regarding using data analytics, the kind of the good thing about it, because I used to work at a bank in the industry, so we have a bit more research angle or luxury to spend time, think about those problems. So we are implementing things which wherever, like from this type of conference, we can learn. So that's, I think, the difference I see. But translating that into actual results, that's where we still need to see more.
