A Guide to Exploiting Predictable Correlations and Volatility

During their latest episode of the VALUE: After Hours Podcast, Taylor, Carlisle, and Rasmussen discussed A Guide to Exploiting Predictable Correlations and Volatility. Here’s an excerpt from the episode:

Tobias: This might be a state secret, but how does Dan get the data for all of his bonds?

Dan: You can get a lot of data from Capital IQ. It has the 10 ten years of bonds pretty well. And then anything other than that, Bloomberg is the only really reliable. You kind of have to trade bonds, you really need Bloomberg.

Tobias: Do you want to tell us a little bit about the multi-strat fund? That’s launched or that’s just launching?

Dan: Yeah, we launched it about two years ago. We’ve changed it. We tried some things that didn’t work and then we’ve really been improving the model. But what we’ve come to is that, I think a lot of investors are very focused on expected return. You want to maximize your expected return. But expected return is also really hard to predict, as we’ve talked about. It’s really hard to know. If you take every stock and try to rank them by expected return, which is what all equity investors are trying to do, it’s really, really hard. And if you look at the R squared on factors, how well factors predict the expected return for stocks, you’re getting in the 5% to 10% R squared range. Like, it’s a real edge, but it’s a lot of noise.

Jake: It’s not much though. Yeah.

Dan: It’s a lot of noise. It’s not much. It’s really hard to predict. What we’ve learned is that, actually, what’s much easier to predict is correlations and volatility. So, if you just take a weighted average correlation matrix, you say, “Hey, let’s give it a half-life of a month or three months, and look at the correlations between stocks and bonds and oil and value and size,” and whatever, that that correlation matrix is pretty stable. It changes over time, but you can predict next month’s correlations pretty well with that weighted half-life type history of recent correlation matrices, such that the R squared on that might be– It’s hard to think of what an R squared means for a correlation matrix. But if you think of some equivalent, you’re probably getting into like 70 or 80% R squared. Like, you can really predict correlations pretty darn well by relying on [unintelligible [00:50:59].

And then volatility is also really predictable. So, last month’s volatility, you take in the VIX and you take in recent, like last month’s volatility, you can get a 40% or 50% R squared predicting next month’s equity volatility. And if you try to predict bond volatility and oil volatility and whatever, you’re going to get pretty good at that too. It’s just pretty auto correlated. It moves a lot, but it’s auto correlated. So, if you say, “Well, gee, let’s imagine, I can’t say I have no view on expected returns.” I think stocks return what they long-term averages. I think bonds return long-term average. I think oil returns long term average. Everything just has a long-term average return. So, I think the market follows a random walk. I have no view of anything.

But I think that volatility and correlations move around a lot. You run that through an optimizer, you’re going to get very different portfolios every different month, because if stocks and bonds are really highly correlated, gee, you’re going to take down your exposure to one or the other, because you don’t need both. And if stock volatility goes up a lot and you have the same expected return forecast, then your forecast of Sharpe dramatically went down. So, you’re going to say, “Well, gee, I want to take down my equity allocation, not because I have any negative view on equities, I have the same expected return view.” But for that volatility, they’re just less of a good buy right now. I’m just getting less Sharpe for the same products, I’m going to reduce my weight, and maybe I’ll take it up in something that’s less volatile than normal.

And so, what we started is basically building this giant database of every single stock categorized by factor, bonds, both sovereign and corporate with factors, and then commodities, oil, copper, gold and currencies, the major tradable currencies, and saying, “Hey, let’s run an optimizer where we look at their volatility and correlation structures.” You take some rough bench regularized to some rough 60-40 like benchmark and then say, “Gee, can I improve outcomes, because I’m really good at predicting volatility and correlations?” And the answer is, “Gee, yes, you can.” You can really, really dial up Sharpe and you’ll take bets that you might be really–

For example, right now we’re quite short the Mexican peso. We have no view on the Mexican peso. In fact, our model is told that the Mexican peso is a 0% expected return, always. We never have a view. It just happens that right now the Mexican peso is really negatively correlated with a lot of bets that we want to take. So, we like value, and it turns out that the Mexican peso– When value does well, the Mexican peso does badly or something. So, it ends up loading up on the Mexican peso to diversify our value long. And you’re like, “I never would have thought of that. That’s totally nuts to me.”

Tobias: [crosstalk] It’s what I’ve been missing.

Dan: It’s what we’ve all been missing, clearly. But when you think of why it did that, it actually makes a lot of sense. It actually works decently well. And then we’ve said, “Okay, well, gee, now what if we could actually predict, have some edge in expected return? Is there some way where we can make better expected return forecasts?” We looked at everything we could try to predict. Can we time value? Can we time size? Can we time Treasuries? Can we time high yield? And for 90% of things, we found that we couldn’t. There’s no ability. Nothing we came up with, we threw the kitchen sink at it. Nothing worked out of sample. Just all a failure. Like, we have no ability to predict whether Japan is going to do better in the US next month. We have no ability to predict the US equity intercept.

But for some things they are predictable or more predictable. So, momentum, I just described to you how equity momentum works really well under 600 basis points, not well between 600,000 as reversals over 1,000. You plug that in, you’re actually a big improvement in your ability to forecast momentum returns. And then you can apply a similar logic. One of the logic we talked about high yield spreads is size. When high yield spreads go– When they’re going widening, blowing out, size does worse. When they’re coming in, size does better. When spreads are really wide, size does better. When they’re really tight, size does worse. And then, gee, you can actually get a 10% R squared in predicting the size premium, for example.

And then oil is another example, where oil is really driven by high yield spreads. When high yield spreads blow out, oil sells off. When high yield spreads come in, oil does well. When high yield spreads are just bumping around, oil just goes randomly oscillates in a dramatically unpredictable way. But you start to layer on all of these things, and you accumulate all of them into rules, and you write software to trade them, that’s what we’re trying to build, is to try to build all these insights into, basically, software that can trade all of these ideas and understand the volatility and correlation matrix across 39 correlation pairs. That’s the essence of what we’re trying to do, which I’m pretty excited about. It’s been a huge, huge research effort, both building out the infrastructure to do it, doing all the research, and then learning how to actually trade it and how to make it work.

You can find out more about the VALUE: After Hours Podcast here – VALUE: After Hours Podcast. You can also listen to the podcast on your favorite podcast platforms here:

Apple Podcasts