Vivek Viswanathan - Quant Equity in China (S4E10)

Hey everyone, Cory here, thanks for tuning into another episode of Flirting with Models. If you're enjoying the show, I'd greatly appreciate it if you'd take a moment to rate, review, and most importantly, share with a friend. A word of mouth is how this podcast grows. And if you'd like to learn more about Newfound's platform of Returnstacked Mutual Funds, ETFs and model portfolios, head over to Returnstacks.com. Now on with the show. All right, 3, 2, 1, let's go. Hello and welcome everyone. I'm Cory Hofstein and this is Flirting with Models, the podcast that pulls back the curtain to discover the human factor behind the quantitative strategy. Cory Hofstein is the co-founder and chief investment officer of Newfound Research. Due to industry regulations, he will not discuss any of Newfound Research's funds on this podcast. All opinions expressed by podcast participants are solely their own opinion and do not reflect the opinion of Newfound Research. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of Newfound Research may maintain positions and securities discussed in this podcast. For more information, visit thinknewfound.com. This season is sponsored by Simplify ETFs. Simplify seeks to help you modernize your portfolio with its innovative set of options based strategies. Full disclosure, prior to Simplify sponsoring the season, we had incorporated some of Simplify's ETFs into our ETF model mandates here at Newfound. If you're interested in reading a brief case study about why and how, visit Simplify.us/flirting with Models. And stick around after the episode for an ongoing conversation about markets and convexity with the convexity maven himself, Simplify's own Harley Bassman. Vivek Visvetatin is the head of research at Raylian Global, a quantitative asset manager focused on generating alpha from investing in China and other inefficient emerging markets. Our conversation circles around three primary topics. The first is the features that make China a particularly attractive market for quantitative investing and some of the challenges that accompany it. The second is Vish's transition from a factor-based perspective to an unconstrained characteristic driven one. Finally, the critical role that machine learning plays in managing a characteristic driven portfolio. And at the end of the conversation, we're left with a full picture of what it takes to be a successful quantitative investor in China. I hope you enjoy my conversation with Vivek Visvetatin. Vish, welcome to the show, really excited to have you here. Thank you for joining me. Thank you. Yeah, I got you right. You're about to take a sip of water there. That's the what I like to do to my guests, catch them all. Well, let's begin for the guests who maybe haven't heard of you before with a little bit of an intro in some background. Absolutely. So, I went to the University of Chicago for undergraduate, mentioned Abban, MFE from UCLA, and a finance PhD from UCI over the years. I worked very briefly at Morningstar in 2005, quickly fired for telling them they should change or give rid of their equity mutual fund rating system because funds at a T5 star status still under performing the next funds. But my morning wasn't quite as kind as that when I said it, so that may have contributed to the fire. After that, I worked at Research Fluids and worked on equity factor research and cross-sectional research and other asset classes. And in 2016, Jason Sue and I, along with a few other spun off, the Asia arm of Research Fluids to start growing global advisors. At Brilliant, our focus has been on China, primarily Chinese equities, but also quite a bit on Chinese commodity futures and bots. And we've recently launched a China A shares ETF in the U.S. R.A. Y.C. At Brilliant, we implemented more sophisticated quantitative models using alternative data, expected returns built from machine learning models and optimization. So first question for you would be, why so much focus on China specifically? What makes China particularly attractive to attack from a quantitative investing perspective? Yeah, that's an excellent question. First, I want to draw some lines. When I say that China is a great market for quantitative investing, I'm specifically talking about China A shares, which are listed onshore and mainland China in Shanghai or Shenzhen. There are mainland Chinese stocks listed in Hong Kong, which are A shares, red chips and P chips. And there are China ADRs listed in the U.S. We generally find that alpha comes from investor mistakes. So Chinese stocks listed in Hong Kong or the U.S. are going to be reasonably efficient. It's a stocks listed in Shanghai and Shenzhen that we believe have a lot of alpha. And just to emphasize this point, from a return correlation perspective, A shares are as different from H shares and ADRs as the U.S. is from emerging markets. That's how different it is. So do keep in mind we're talking about onshore China, that's China A shares. So China A shares is a huge market, not as large as the U.S. obviously, but it's punching in a similar way class. The U.S. has a $47 trillion total market capitalization while China A shares has a $13 trillion total market capital. In 2020, the U.S. had a monthly average trading volume of $6.5 trillion while China A shares had $2.4 trillion in monthly average trading. The reason why you want to find alpha in a huge market is that you want it to be liquid enough to trade and deep enough that the alpha won't disappear after you deploy a modicum of capital toward the strategies. You might be able to find a lot of alpha in a small market, but you can't actually invest significantly in capital towards China A shares, A shares, it luckily is huge. So the second thing you want to look for is whether other institutional investors are out before. Of course, if you look at large cap U.S. equities, 90% of active managers underperform the market over any given tenured period. In China, the average equity mutual fund outperforms the market after cost despite having a 1.5 percent management fee on average. The average international investment in China A shares outperforms. In other words, stocks with high international holdings outperform those with low international holdings. So in the U.S., you need to be in the 90th percentile among fund managers just to perform as well as a benchmark. In China, being in the 90th percentile means delivering solid outperforms. But that's the active management in general. What about quant investing in particular? First, I'll talk about the syncontics of factors, but you can beat the factor framework. But we'll talk about that later. So if you invested in a standard basket of long short factors, think McLean and Potif 2016 factors, integrate into a long-only portfolio. In the U.S., you'd earn a.36 information ratio from 2010 to 2020. In the EM, you would earn a.91 information ratio. In China, you would earn a 1.06 information ratio. Just so this is clear, these are factors that were discovered in the United States. And we're blindly implementing them in China and earning a 1.06 information ratio. That's pretty incredible. If instead of looking at standard factors, you look at China-specific factors using alternative data-specific to China. Those earned an IR of 1.43 since 2010. And that's yet another source of alpha in China-Asia's. And it makes up about a third of our access returns. Now, for what it's worth, all of those information ratios are pre-transaction costs. You probably need to shave about.2 off of those to get to post-transaction cost numbers. So just to quickly sum up that long-winded answer, we think China-Asia's is a good place to apply quant investing because it's very liquid, so it can absorb a lot of capital flows. Because other institutional investors tend to outperform. So we have a sensor's genuine alpha to be found. And because standard quant factors do well and China-specific quant factors do even better. Can you expand a little bit on that notion of China-specific quant factors, like maybe some examples of what those signals are? Absolutely. So, the first one I want to discuss, which we're publishing a paper on in June, is the AH Premium. I mentioned that 8 shares are listed in mainland China and Shanghai and Xinjiang, and 8 shares are listed in Hong Kong. Some companies actually dual list in both mainland China and Hong Kong. That means that we have two identical shares, which the same voting rights, same dividend rights list on two different exchanges. Since the two shares can't be converted into each other, this isn't a pure arbitrage play or anything. But you would expect the prices to be relatively similar, since they're the same share on different exchanges. But the thing is, the prices aren't even close. On an equally weighted basis, dual listed 8 shares are on average 75% more expensive than their 8 share counterparts. On a cat-weighted basis, they're on average 33% more expensive. And there are some stocks like the Chinese fashion company, La Chapelle, it sounds French, but it is indeed Chinese, which is 4.5 times more expensive than its 8 share, while other more stayed, maybe boring stocks like Ping On, have an 8 share price that is actually identical to its 8 share price. Now, it turns out that because Hong Kong tends to have more sophisticated investors, a high 8-age premium does not vote well for the 8 share, because that means Hong Kong investors think that stock is overpriced. And a low-age premium means that 8 share will perform well because Hong Kong investors think the stock is fairly price or underpriced. Now, if you overweight low-age premium stocks and underweight high-age premium stocks, you can earn a.89 information ratio from that alone, so that's pretty good for one signal. Next, let's talk about pled shares. So in many countries, including the United States and China, you can use shares you hold in a publicly listed company's collateral for a loan. If you can't pay back the loan, the bank sells the shares. This is what you do if you want to keep your shares, but you might want money for a house or a boat or whatever. In China, directors and managers might pledge $50 million, $100 million or much more for a loan. Now, obviously, they are not buying a boat with that money, so what do they do? They're lending the money back to the firm. So in China, if you're not a state-owned enterprise, it's surprisingly hard to get a loan, and issuing bonds or equity requires a fairly lengthy regulatory process. But if shares are pledged as collateral, it's much easier to get a loan from a bank. So managers will pledge their own shares as collateral, get a loan, and lend that money back to the firm, effectively doubling down on the firm. That is a vote of confidence. They would only do that practically if they believed in the firm. Now it also suggests that they are not profitable enough or have enough cash to get money from traditional sources. So this signal only positively predicts if you control for profitability in other variables. Now if a firm has large total pledge shares after a long period of time, that suggests a firm can't pay back the loan, and that's a negative sign. So there are two signals to be had here. One is new pledge shares, which is a positive signal after controlling for other variables, and total pledge shares, which is a negative signal whether or not you control for other variables. So just to go over one more, I mentioned foreign investors outperforming. The foreign holdings in various stocks are disclosed every day. Just looking at Northbound Connect holdings, which is a particular type of foreign holdings, but every month you overweight high Northbound Connect holding stocks and underweight low Northbound Connect holding stocks. Then from July 2016, when the data started to now, you would have earned an information ratio of 1.92. That sounds incredible and obviously you shouldn't assume that you would earn exactly that information ratio in the future. But it suggests you can get a tailwind from falling to behavior of other smart money investors. By the way, one last thing. I am not the driver of China signals. That is the head of China Research Donald Ho, who works out of Hangzhou China. I don't want to take credit here. We have an entire team in China driving this. In any case, those signals give some color about what's so nice about China Asures. There is so much data. There are so many mandatory releases for firms, many of those releases contain information that you can use to predict return. Much of the data is either on wind, which is a data provider in China similar to Bloomberg, or in a well-organized format online that can be easily scraped. The signals I discuss just begin to scratch a surface. There are so many more signals that one can build from the sheer volume of data that comes from these Chinese firms. What I find somewhat contradictory here is you're talking about disclosures and very frequent disclosures. It sounds like that is leading to an increased amount of information in the marketplace. I think from an academic perspective, we would expect that to lead to a more efficient market. I think it actually doesn't change the efficiency. In fact, you can find more inefficiencies with all this data. How do you reconcile that fact? About 80% of the volume in China Asures is driven by retail trading. These folks aren't reading these releases or crunching this data. They certainly aren't building natural language processing algorithms to process this data faster than any human could read it. China has the perfect combination of things to generate alpha. A ton of data and a large segment of investors who are not fully or optimally utilizing that data. Of course, that will change over time. As that happens, we need to find more data and build models that are captured at data. If all goes well, eventually, the market will become efficient and 90% of active managers will. Underperform. Just like in the US. Then we'll pack our bags and tell everyone to buy a cat way to index one. But we're probably a few decades away from that. From an outsider's perspective, it seems like the regulatory regime in China can be quite fluid. What sort of challenges or even potential opportunities does this present for you as an investor? The regulations are constantly changing. The tightest stock market is fairly young. The Shanghai Stock Exchange was founded in 1990 and Shenzhen right after that. So regulations are still triangulating on the preferred regulatory regime. So significant regulatory changes still occur with some regulatory. Back in 2015, there was a stock market bubble on crash localized to China asia's. The Chinese regulators blamed malicious short sellers, which obviously is not why the market crash. But whatever. Let's get into that. And they banned most investors' ability to lend shares, which effectively banned short sell. Now, in June of 2020, they allowed securities lending, again, for most investors. And if you look at the graph of short selling, it looks like it's at the very beginning of the escrow. So it's in its exponential growth phase. Now, the cross-sexual dispersion of short selling can presumably be turned into a signal now. As you know, historically, on average, shares that are sold short and we're heavily have lower expected returns than those with lower short interest. But we haven't done that research yet. So that's a new potential signal that came from regulatory change. Another recent regulatory change was the removal of the IPO PE restriction. So let me explain what that is. Back in 2014, the government created a cap on the price-to-earnings ratio of IPO firms at 23. This wasn't even a formal rule, by the way. It was just an informal guidance that was followed perfectly. Now, this also wasn't a secret. You can read about it in the news. It's just another interesting feature of Chinese markets. Some of the rules are not technically codified, but neither are they a secret. The point behind this rule that capped IPO PE at 23 was to make sure that retail investors who love signing up for these IPOs can get a great deal and make a ton of money. Now, as you can probably guess, because these IPOs are discounted, they are heavily oversubscribed. So you have to enter a lottery to get in, but it's fantastic if you do get them. But before this rule was initiated, IP was run-deprized by 10 or 15%. After this rule was implemented, IP was run-deprized by 200 to 400%. I got those numbers from Ding, Sinclair, and use a recent working paper called redistribute the wealth Chinese IPO market reforms. I want to make sure I credit them. Anyway, last year, they removed the 23 PE limit for a small subset of the market, the Gem and Star Markets, which are small cap tech heavy boards. And they'll probably eventually remove it for the market as a whole. So as long as this rule is in place, you'll probably want to subscribe to IPOs. In the Gem and Star Markets, it's not nearly so compelling anymore. For what it's worth, if you are a foreign investor using Hong Kong Connect to invest in China Asia's, which is the usual way to invest, you can't even subscribe to IPOs. So this rule change doesn't matter much to us. But if you are a QE investor or have a fund in China, you can subscribe to IPOs, and you probably should try. These are just two of the recent changes. Some of the other ones are daily price limits in the Gem and Star Markets, where increased from 10 to 20% recently. Regulators are also getting more aggressive about de-listing firms that engage in accounting malfeasance. But the key here overall is you really have to keep your ear to the ground when it comes to Chinese regulatory changes. One way to do so is to have a person on the ground, which is genuinely very helpful. But when it comes to regulations at least, you can generally just read the news and you should be okay. So you spent a good part of your career, I think it was nine years at research affiliates, which was a firm that was a real innovator in the smart beta space. I know since joining Rayleigh, and you've sort of shed off that factor-oriented view of the world. And I'm curious as to what led to that change. This one's going to be a long and technical one, so forgive the soliloquy I'm about to give here. So let me clear out some potential semantic misunderstanding here. When I say I don't have a factor view of the world, I mean anomalies as linear factors. I believe that the equity market factor is an extremely useful tool. I believe the industries and countries drive the covariance matrix. So not having a factor view of the world is specifically not viewing anomalies as risk factors or linear univariate mappings to return. And I hope to make that claim that when you think about anomalies, you need not think about risk factors and about long short portfolios built on linear characteristics sorts. Those are useful for publishing papers and for a heuristic understanding, but they are not useful for portfolio construction or for mapping how characteristics relate to expected returns. There are few high-level ideas I want to touch on. The first is where the idea of factors come from. Factors are an idea that stems from arbitrage pricing theory or APT, that the expected return of an asset is a linear combination of priced risks. And you might argue that even if you don't believe in APT with respect to all these anomaly factors, you can still believe that factors are useful to them. I'll address that later, but for now, I want to talk about why the characteristics that we see that predict returns should not be considered priced risk factors. There are a few things we would expect to see if we believe things like value, gross profitability, cash flow and cool post earnings announcement drift, new sentiment momentum, whether those were priced risk. So first, one that few people talk about is that we would expect to see products that sold high valuation, low profitability, negative earning surprise, and every other negative factor told you can imagine, because those would be great hedges for whatever latent risk investors are trying to avoid. If these anomalies are risk, it stands to reason that investors would want to hedge these risks and would happily pay their counterparties to hedge these risks. But those products largely don't exist. There's no hedge or latent risk factor product that earns a negative expected return in exchange for risk hedging. No one is selling these supposed insurance products to investors. Moreover, no one has figured out what risks this multitude of factors is meant to price. What risk is profitability a proxy for? What risk is conservative accounting a proxy for? What about earnings growth? What about momentum? No one really seems to know. Now to certainly the case of value in size embodies some aspect of risk. Imagine you have two firms with exactly the same expected cash flows, but one is riskier than the other. Then the riskier one will have a higher discount rate and will be smaller cap and deeper value, so if there are latent risk, they will probably show up at least in part in value in size. But it's really difficult to make the case for the multitude of other six. And of course, if factors were price-risk, we would expect to see co-verances driving returns not characteristics, but they don't. That was found by Daniel and Titman in 1996 with respect to value, but it's also been shown about most other anomalies as well. In fact, my thesis shows this with regard to the 97 McLean and Pontifactors, at least the 86 in compute globally. It's characteristics that drive return. It's true that there have been attempts to recover this idea that if you squint hard enough co-verances or driving returns, so in 2019, Gukheli and Shosh used a variant of an auto-encoder and account for the predictability of returns in that encoder and say you can compress the information in characteristics and outperform other linear asset pricing models. And of course you can, because as we'll discuss the relationship between characteristics and returns is non-linear. So if you put a non-linear model against a few linear ones, you're probably going to win out. But that doesn't mean you found price risks. It just means that auto-encoder is a good at compressing information and non-linear models are performed linear ones. The real test would be putting it up against a neural network that didn't have a bottleneck layer in it. In other words, does the act of compressing your factor space improve or hurt your predictability? And from all of our analysis, compressing the factor space hurts. You really want that 100 plus signals helping your prediction. And yet another piece of evidence against the risk factor approach to understanding anomalies is how we construct co-verance matrices. When we build co-verance matrices, you might account for market industries, countries, of a component of return, and admittedly, size and value. But you wouldn't account for post earnings announcements, drift or operating profitability of factors. They are not meaningful parts of the co-verance structure. Now, why does this matter? Who cares whether anyone thinks factors are price risks? It matters because if you believed that factors were price risks, you would sit on one side of the trade forever, expecting to get paid for taking that risk. You would expect that someone would happily pay you a premium to take on this factorist. But it's a behavioral anomaly. Then you're keenly aware that the market might become efficient to that signal. But before I talk about why factors are not even particularly useful as tools, I want to talk about the number of factors. So there's this idea of a factor zoo brought up by John Cochran, that there are too many factors and there's a mystery here. And indeed, he's right, there isn't mystery. But the way much of the literature has tried to explain it is not correct. Way back when I read Harvey Lio and choose and the cross-section of expect return and later Hoshway and James replicated anomalies. And if you read those papers, you might think that most cross-sectional anomalies are data state. And that markets are more efficient than the anomaly you literature suggests. Or that most anomalies can be collapsed into other anomalies, so that there are only a handful of anomalies out there. And by the way, I briefly believed all of those things, right? So very briefly though, like maybe a few months, some pretty quick empirical tests can refute them. So as it turns out, you really want those 100 plus signals in your expected return model. The first bit of evidence for this is Jacobs and Muller's 2019 paper anomalies across the globe. They found that if you create cap-weighted long-short portfolios out of the 241 anomalies that have been found in the literature and equally weight them, that is equally weight the factors, then you will earn significant excess returns in 38 out of 39 markets. And that's significant at a.01 level, the odd one out happens to be turkey. Now let's think about this for a second. These anomalies were discovered in the US. If they were the result of data snooping, then they would not produce excess returns in other markets, but they deliver excess returns on average in 38 out of 39 markets. Now you might argue that the returns might be driven by only a handful of these 241 anomalies, maybe 5 out of the 241 let's say. That's where my thesis comes in. I didn't look at 241 anomalies, but I did look at the 86 out of 97 McLean and Ponte of anomalies that can be calculated globally from 1995 to 2018. And 44 out of 86 of those earned a significant Pharma French 3-Pactor Alpha, and that's at a.05 level. Two of those 86 are size and value themselves, so they can't produce a Pharma French 3-Pactor Alpha. 44 out of 84 anomalies deliver a significant Pharma French 3-Pactor Alpha globally. That is over half of the anomalies found in the literature. That's too many significant anomalies for the result to be driven purely by luck. Now you might ask, why does the Pharma French 3-Pactor Alpha matter? Why not just look at whether the returns are significant in their own right without looking at the Pharma French 3-Pactor Alpha? The answer is that almost every signal is a quality signal. My quality signals as a whole tend to be negatively correlated with size and value. AQR showed this with respect to size and their size matters. The ideal holds for a variety of quality factors, even ones that come from alternative data sources. So, that is why the PhF3-Pactor Alpha is so important and why you will see so many anomalies with insignificant return but significant PhF3-Pactor Alpha's. Now I want to briefly talk about Hoshuane-Jang's paper replicating anomalies because there's a particular issue with it that some of these other papers that try to collapse anomalies have as well. They first look at factors that earn significant returns in their own right, and then they try to explain only those factors. The insignificant factors are never tested against the PhF3-Pactor Model. But this is the issue with that approach. If you regress insignificant factors on a set of other factors, those insignificant factors might have a positively significant alpha. And that is now a new anomaly. You don't get to ignore those guys. If you're as surprising an anomaly generates more significant alpha than there were significant factor returns in the first place, that is not a successful model because there is now more significant factors than before. A linear combination of factors is still a factor. So, I hate on two things already. What is that? Factors are behavioral anomalies. The second is that there are many anomalies. Now we have these 100 or more potential anomalies and we don't believe there is factors. So, what are they? They are information about long horizon, expect a return of the firm. And this is information that can interact and be non-linear and it can be incorporated into price in one period or another. And this is exactly how we would view the world if we never learned about factors. We would look at a company and say, it has this gross profitability, this operating profitability, these are tools. I read this negative piece of news and this positive thing from their annual report and they're being sued by a competitor for anti-competitive practices. When you take all this information and you get a value for the firm and you compare that to the price and you say, this stock is overvalued or undervalued. You want your model to effectively do the same thing. Practically that means you need machine learning model. At the very least you need linear ridge so you can prevent yourself from overfitting to those 100 plus signals. But if you want interactions and non-linearities, you need gradient boosting and random force. If not, beat forward no networks. You don't have to predict the value of the firm, though. You practically are predicting return. But you want to include all of those signals and throw on price and valuation as additional features. Now, once you have your model, how do you improve your product? Instead of trying to build new cross-sectional factors, you're trying to provide information to your model about the value of this firm so it can learn more about expected return. So you're still engineering features. You're just not building factors. In summary, these anomalies aren't risk factors. They're characteristics that predict return. And they can do so through interactions and non-linearities. There are many, many characteristics and you want to search for this data everywhere you can and use the latest tools to try to encode information about stock, expected returns. So I think for a lot of quants who got their education in the last 15 years, it can be very difficult to mentally unchain themselves from the traditional factor framework. And often it sort of requires seeing a non-factor framework in practice to sort of break that mental box. So I'm curious whether you could go into maybe what that non-factor framework does look like in practice versus the more traditional linear models. Absolutely. First, you need an expected return model. And assuming you're predicting cross-sectional equity returns, that model should utilize some subset of things that fall under machinery. If you aren't trading in individual stocks, then what you do is dependent on the amount of data that you have. In cross-sectional equities, you generally have a lot so your models can be far more sophisticated. If you're looking at commodity futures on a monthly horizon, I might use linear rich, but the cross-section is probably not big enough for anything more complicated. If you're looking at asset allocation on the other hand, it's even fine to use static weights there because it's hard to build conditional expected returns in that space that are better than unconditional expected returns. When you're predicting cross-sectional equity returns with 100 signals, you do not want to use OLS because you will overfit and get no excess returns in your auto sample backtests. I can't emphasize enough how different your experience will be using OLS versus using a model with some form of regularization built in that prevents your parameters from overfitting. But if you're predicting on an annual horizon, you can rely on linear rich and maybe random force. If you're predicting on a quarterly or monthly horizon, you want linear rich, random force and gradient boosting, you might be able to use neural networks on a monthly returns, but we don't. If you're predicting on a final horizon like daily or intraday, then you definitely want to use neural networks, probably only neural networks. Now the next and most important step is you predict returns in each time step. Let's say every month, using models that could only be fit in prior months. If you're predicting a return in July 2010, you can only use models that fit data up until June 2010. In other words, you can only use models that were fit to prior period data to predict the next periods return. That is how real life works and you want your model to behave similarly. These are what are called pseudo out of sample or quasi out of sample backtests. We tend to just call them out of sample backtests because the word backtests already implies that they can't literally be out of sample. Now let's go back to our prediction returns in 2010 example. Let's say your data set starts in 1995. In 2010, you have 15 years of data to fit your returns. In 2020, you have 25 years of data to fit your returns. So your backtest assumes you have less information than you do now. That does mean your backtests will somewhat understate your ability to earn returns, but that's probably more than offset by the fact that markets get more efficient over time. Now, once you have your expected returns, you need a COVID-19 matrix. There are many different ways to do this while you probably want to count for structural sources of covariance like industry, country, size and assume the remaining variance is residual. But there are other ways. For example, using some number of principal components and then assuming the remaining variance is residual works fine too, you also need a count for various decay horizons and you need to shrink the loadings on sources of covariance. I'm going to skip over that because it's probably its own conversation and it's just not good podcast conversation. You need a whiteboard or something. Now, you have your machine learning expect returns and you have this covariance matrix. You want to do mean tracking or optimization with respect to your benchmark. If you're roll to a return investor, if you're trying to max my sharp ratio, you do mean variance optimization, you also want to have constraints on industry weights, country weights and individual stock weights. It's difficult to communicate your model that you're uncertain about the covariance matrix. And constraints are clumsy but effective way to say, look, I did my best on the covariance, let's take it with a grain of salt, let's make sure not to get too crazy here with the offset and bets. I mentioned earlier that an equally weighted basket of China factors has something like a 1.5 information ratio from 2010 until now. Using all these methods on global and China specific signals and a walk forward out of sample weight, an optimized portfolio and China large cap will produce an IR of about 2.8 after transaction costs since 2010. A walk forward optimized portfolio in China small produces IR of about 3.5 after transaction costs. It will do that partially by building good expected returns but also by drastically reducing tracking error. At that information ratio, your tracking error will be between 4 and 6% so your back-tested expected return will be 12 to 24% excess returns, I'm sorry. Given such a low tracking error, you might get classified as enhanced indexing despite your expected return. That's something to be wary of. To deal with this, you can relax your constraints and reduce your risk of version but your information ratio will fall. You cannot maintain very high information ratios while taking high tracking error because of the zero lower bound on weights. Now if you have a long short portfolio, you can kind of go to town here. Now there's some things worth realizing about expected return loss. If the model doesn't think you can perform well in a market, it's not going to bullshit you. If you run the same walk-forward optimized model in US large cap, you would earn an excess return of 2% and an information ratio of 0.6 after transaction costs since 2010. In Hong Kong, you would get an information ratio of zero after transaction costs. You wouldn't earn any excess returns on the market. If you run an EM large, you get about 1.6 after transaction costs. If you run in China large, it's about 2.8. If you try to do walk-forward prediction in an efficient market, you're going to have mediocre results as you wouldn't realize. If 90% of the active managers in the US underperform over any 10-year period, it would be strange if we found a model that did incredibly well. Now another useful insight is that if you build a solid covariance matrix, your optimized model may show no significant factor tools using traditional factor attribution. And we actually had a discussion about this at our firm just last week. We fed one of our newly launched products into some factor attribution software and low and behold, we had virtually no factor tools relative to our benchmark. The one factor tool we did have momentum did not perform all over the past three months or so. And yet our active return over this benchmark that we had no significant factor tools against was a positive 5%. And this is because factor tilts don't really matter. They don't tell you they expect a return of the portfolio. They don't even do a good job classifying the risk of your portfolio, a profitable retail stock in China A shares. It's not going to be any more correlated with the profitable tech stock in China A shares than an unprofitable tech stock in China A shares. In the end, even though you're using value signals and profitability signals and many other signals, those very well may not show up as linear factor tilts since your models nonlinear and accounts for covariance risk. And naturally clients might have the question, if you can use factor attribution, how do you know if you're taking the same risk across two portfolios and across two managers, you would take the correlation of the active return of the two portfolios. Now, if you're worried that this is too backward looking, you can build an expected covariance matrix of the underlying securities and how they'd expect a correlation of the two portfolios. That's hard, but at least it's accurate. If you're using factor loadings to compare portfolios, you might as well use astrology for all the good that's going to do you. So one of the fundamental struggles that almost every quant deals with is the non-stationarity of data. It seems to me like that problem potentially compounds when you start talking about interaction and nonlinear effects. But before we even get there, when we talk about a market like China where we mentioned there are big regulatory regime shifts, how do you think about model construction where you just may not have the depth of data necessary for analyzing a signal because of a meaningful regime shift occurs? So if the regime changes rapidly and unexpectedly, then you'll just be cut off guard anyway. Nothing you can really do about that. But let me give an example in China that may be helpful in understanding what one might do to a regime shift that you seek. So in China, the government is trying to reform SOE's and make them more efficient. We have SOE classifications and SOE stands for Stato and Enterprise, by the way. We have SOE classification as a parameter in our expected return model. But if we believe the reforms would be successful, we should remove that variable. Now if we don't believe the reforms will be successful, we should keep that variable. And for what it's worth, we haven't decided what to do yet. But the key here is we need to act before the regime has shifted or soon after. Otherwise, the model will learn about it without our help. Now one thing you can do to ensure that the model can handle changes in the relationship between features and expected return is to have a decay on the data, such that it learns more from recent data and less from data from the distant past. That's a good idea in general. But it will handle slow changes in the market environment, not fast changes. But for what it's worth, changes in the relationship between features and expected return tend to be slow moving. With regards to adjusting models by hand in general, I would strongly caution against it. If you have a strategy earning an IR of 3 or even an IR of 2, what forward-looking thing can you add to improve the model? It's much more likely you'll hurt the model as improvement. And I'm embarrassed to say I know this from personal experience. We used to adjust our model weights when they did things that we deemed unintuitive. It turns out that's a pretty bad idea. The model is learning from mapping features to expected returns using 25 years of data and tens of thousands of stocks and optimally trading off between expected return and tracking error given a particular risk of version. Our intuitions are probably not going to help the model balance there. Now, if your model is missing something, it's generally a good idea to add it directly to the model. For example, let's say you find that your biggest negative contributes to portfolio performance come from failing to react to news. And the model currently has no way of reading the news. Instead of you reading the news and manually adjusting portfolio weights, you can use an NLP model, maybe Google's burn or something, can code the news and use that to predict return. You wouldn't use the full encoding of a document. That vector would be too large. And instead you might extract sentiment or some other component of information. That way the model can learn optimally without human intervention. When you look at the large basket of characteristics that you've compiled over time, how many of those signals or characteristics do you think would fall under traditional factor classifications? For example, gross profitability would be a profitability or quality factor as an example. The brief answer is something like 50% fall under typical factor classification. Those signals fall under familiar rubrics like value, size, accounting, conservatism, investment conservatism, profitability, earnings and revenue growth. But I want to use your question to talk about signal classification systems in general. Because I think it can inform signal research. And forgive me if I'm hijacking the question a little bit. And critically, I want to point out I'm talking to Quants here. This is not how you want to present your signals to your clients. Just how you want to think about the philosophy of signal tax on it. So let me talk about how my research and portfolio management team think about signal classification because it's in some ways different but in some ways extremely familiar. First, I want to do something similar to what a QR's quality minus junk paper did and take it in a slightly different direction. As you might remember, they look at the Gordon growth model divide by book value and say everything on the right side of that equation can be considered quality. I'm going to do something similar but with the residual income valuation model. The residual income valuation model says that the value of a stock is equal to the book value plus the discounted sum of residual income. It's very similar to the dividend discount model and in fact, it's derived from it. But it uses residual earnings instead of dividends and it's added to a book. So let's divide through by book value and what you would have is the price to book value of a firm is equal to one plus the discounted future return on equity of the firm. So this is pretty damn similar to what the AQR paper did. But again, I want to take it in a slightly different direction. This equation shows that there are three signals that predict long horizon return. Market capitalization, which is on the left side of the equation before we divide through by book value, price to book value, and current and future profitability. So you have three signal categories, size, value, and current and future profitability. Size and value are more or less one signal each. Obviously there are 20 different ways to define value but the ones that perform better do so because they're leaning on current and future profitability. The current and future profitability category encompasses almost every signal in existence besides size and that. So it encompasses account and conservatism. So firms with low cash flow of rules and have higher subsequent earnings. Firms with low net operating assets tend to have higher subsequent earnings. Current and future profitability includes investment conservatism. So firms that invest less tend to be more profitable in the future, presumably because they're not moving into their lower ROI projects. Current and future profitability includes profitability signals obviously. So return on equity, return on net operating assets, return on assets, profit margin. It includes productivity signals like asset turnover and change in asset turnover, which also predict future earnings. It includes default risk measures like Altman Z score or Campbell, Hilcher, and Ziloggy's default probability, both of which negatively predict earnings. Even low volatility, positively predicts us, presumably because firms that have growing earnings are less likely to exhibit market volatility as they are further away from default. Kind of a Martin model, the stock is a call option on the firm there. Even momentum predicts earnings, though momentum of course partially reverses after the first one. And indeed other market signals also predict earnings as well, which is somewhat surprising. And all those signals I mentioned before in China, low age premium, those stocks have higher subsequent earnings, high foreign holding stocks have higher subsequent earnings. Firms with new pled shares can't have higher subsequent earnings, while firms with long-term pledge shares, high long-term pledge shares can have lower subsequent earnings, exactly the same direction as a return prediction. And I'm not the first person to realize it. McLean Ponte and Angover wrote a paper in 2018 called anomalies and news, and they found that stock return anomalies were six times higher on earnings in the last few days. And this was consistent with biased expectations regarding earnings. We might call this class of signals at predict earnings, quality signals. But I think it's important that folks don't get confused by the knowing glitch. Under this framework, there is no such thing as the quality factor or the quality signal. Quality represents dozens if not hundreds of signals, often weekly or completely uncorrelated with each other. There are independent bits of information about future earnings. So your expected return model should be something like 80% or 90% quality signals. And that's only if you have different specifications of value. If you have only one specification of value, you might be 99% quality signals. But again, they're just independent bits of information about current and future earnings. Because people could be many different things when they say quality. I prefer the verbose, current and future earnings or current and future profitability. Because it makes it abundantly clear what's going on. When someone proposes a signal in our research group, we ask one question, does that signal predict future earnings and how? That predictability of earnings is also why all quality signals rely on people under reacting to information. If people were overreacting information, that would flow right into value. If investors under react to a piece of information, we need that information in the model. If investors overreact to a piece of information, that signal will be captured in value and that information itself is not strictly necessary in the model. Now, there's an entire other class of signals, which includes everything from high frequency trading signals to short-term reversal and co-skuness. Those are signals that predict return, but then reverse. In general, those signals are extremely high turnover. If you're predicting on a daily horizon or higher frequency, those signals make up the vast majority of your predictors. But for a monthly or less frequent model, those signals likely serve little purpose. Unfortunately, I don't have a good term for those, so I just call them high decay. And I believe that is a complete taxonomy of signals for expected Socrates. Value, size, current and future profitability, and high decay signals. But critically, this is a taxonomy for quants. It aids in organizing our research. It's not the best taxonomy for dark of the class because it requires that entire lengthy discussion I just gave. By the way, if someone is convinced they have a signal that doesn't fall into that rubric, please reach out to me. I would love to discuss it. So machine learning remains surprisingly still a very hot button topic among quants. There are those who don't think there's a lot of value. There's some who just view it as a branch of statistics, and there's others that have wholesale rebuilt their firms from the bottom up to incorporate a machine learning mindset. I'm curious as to how you gain the confidence to make it a wholesale part of your process and why? Absolutely. The shortest answer as to why we use it is because predicts returns better than not using machine learning. So I actually described the process that we went through to get to machine learning because we were trying to solve a problem. And it wasn't clear at the time that machine learning was even the ass. So we had been using the long only factor approach in all of our strategies. That's where you just wait upwards or downwards based on characteristic scores. And there's a question of what factor strengths to use. There's a question of over what dimensions a particular signal predicts returns is within industry, within country, across industries, there's a question of correlations between factors and how you adjust other factor strengths once you introduce a new factor. There's a question of overfitting because instead of living in a world where you have to predict in a walk forward out of sample manner, you have to select your factors and strengths in set. So you know your back tastes always looks far better than your live performance. And I knew we had to find a better way. Now for what it's worth, I do not come from a machine learning background. All I knew was that I needed to map these signals to expect a return out of sample. And so we tried various methods. We tried ordinary least squares, that overfit, and remember we were predicting out of samples, so we got pretty bad results. We tried Lasso, partial least squares, and principle component regression, and they perform poorly. Now, let's think about why those perform. Because they all try to collapse the feature space. As a general rule, trying to collapse the feature space under performance methods that use a the full range of features. Linear Ridge, on the other hand, will share loadings across cold linear signals. If you have cold linear signals, you don't want to collapse them into one with PCR and PLS effectively do, or to push one out, which Lasso effectively does. Linear Ridge shares the loading between cold linear features. And that helps when you're predicting noisy variables. And of course, stock returns are noisy. So we tried gradient boosting random force and neural networks as well. All of those work fine. All networks are computation-expensive and require careful initialization. So we ended up just going with linear ridge, gradient boosting, and random force. We found that taking together, mostly regardless of the region, which you predict, they predict returns well in our walk forward out of sample models. The models earn much lower information ratios in Hong Kong and US large cap over the past decade. But those aren't our primary targets. We want to predict what on emerging markets and China in particular. And there, it worked exceptionally well. Once we had an optimization, it became a no-brainer from an empirical standpoint. The information ratio was nearly doubled from the factor approach, and the machine learning approach was walk forward out of sample, while the factor approach had unavoidable in sample bias. But we didn't immediately switch, which was probably the biggest mistake we've made of this as an investment organization. As we compared our machine learning paper portfolio results side-by-side with our factor-based live portfolio results, our machine learning strategies beat the living hell out of our old factor strategies. Over seven months of seeing these ML strategies, traums, or old strategies, we migrated the model over. And I still occasionally look at what our returns would have looked like if we had stuck with those old strategies. And let's just say, I'm glad we switched. Now let me be clear, we still have some smart beta products. And I should say there is still an advantage to smart beta, which is easy transparency. You can be transparent with the machine learning model, but it just takes longer. If you want to give someone a full specification of your strategy without discussing grading and boosting trees, then the factor approach is a clean way to do so. We have to be clear about the trade-offs here. When you choose the factor approach over the optimized machine learning approach, you're basically giving up high risk adjusted returns for ease of explanation, which strangely is sometimes a mutually beneficial trade. If you say, I use a mixed integer programming optimizer with a linear ensemble of random forests, grading, boosting, and linear rich for expected returns, and a covariance matrix built from shrunk principal components, some people just say, look, that might be brilliant or might be idiotic, but I'm in no place to judge. So just give me something I can understand. For those folks, you want smart beta. Now I want to briefly talk about why machine learning is a hot button issue for plants. The first is they might work in a space where machine learning isn't appropriate. If you're doing asset allocation and you revounce monthly, there might be a way to use machine learning to help your model, but I honestly can't think of it. That's obviously completely reasonable. Those folks shouldn't use machine learning. The second is that some folks think it doesn't really add a lot of value, despite working an investment universe with a lot of data. They might have tested and said, this actually doesn't help very much. I would encourage those folks to check which models they're using, check their signals, and check their hyper-branders. I've already discussed which models to use on which horizons, but I haven't talked much about the risk of using two few signals. If you have a multi-factor strategy with four signals, and you put only those four signals into your expected return model, you probably will not see big gains from using machine learning. There's insufficient data for the model to work with. Regarding hyper-branders, some hyper-branders like learning rate really matter. If you set your learning rate too high, your model is going to overfit and your out-of-sample back test will generate dismal results. If I'm being perfectly honest, I cannot know everyone's analysis. It could be the case that someone has figured out a proprietary non-machine learning implementation that beats the optimal machine learning implementation to not admit that possibility is to be completely diluted. Obviously, I haven't seen that implementation or that's what I'd be talking about right now. The second issue that folks might have is that they have some factor dynamism that they find value. For example, you might find that factors have momentum or that high value dispersion means that value will perform better. Expected return models actually already handle this. Represents that have momentum have momentum because their underlying characteristics are persistent, not because of anything about factors in particular. If the underlying characteristics are not persistent, you won't see the same positive autocorrelation in factor return. High dispersion in characteristics predicts higher factor returns because the stock returns are related to the underlying characteristic, not the percentile source. It matters that the highest ROE stock has 150% ROE in one period and 100% ROE in another. Those are different numbers and that difference has an effect on expected returns. A feature should drive expected returns more if that feature is more dispersed. That is as true in OLS as it is in linear-riged, grating, boosting, random, forest or neural networks. It's because we're stuck in factor world that we find short-term factor momentum or characteristic dispersion at all relevant or noteworthy. I like to compare factors to a geocentric view of the universe. The geocentric model found all sorts of interesting behaviors for planets. With the geocentric model of the universe, most planets move east to west in the sky, but occasionally they move west to east in these retrograde epicycles. That's just an illusion because planets aren't revolving around the Earth. They're revolving around the Sun in elliptical orbits without the need to reverse motion. In a similar way, all these potential factor timing models only arise because factors are poor description of the relationship between characteristics and expected returns. The last type of factor timing that I've seen is regime switching models. I personally have not seen this work in context of timing. They can build amazing backtests, but seem to turn into a random number generated as exposed. This may be due to my limited exposure, but most regime switching models I see are 90% confident they're in one regime or the other, and then you infer conditional expected returns based on those regime states. They're way too overconfident and thus they can result in bad decision-making. Now the third issue is that it's harder to explain under performance in a machine learning context. If you're implementing a value portfolio and value under performance, you can always say, "Well, value under performance," so the portfolio under performance. After all, the client was the one who bought the value portfolio. That's on your fault. Even in a multi-factor context, you can say, "Well, on average, these factors under performance, so the portfolio under performance." Factoring investing sort of gives you now. It's easier to explain under performance. It's even better when you under perform because junk stocks or glamour stocks are aggressive accounting stocks outperform. Then you can say expensive junk outperform in this period and that sounds like you're not wrong, but the market's wrong. If you use a method that's supposed to account for all this information, as much information as it can, and earn excess returns based on that information, then when you under perform, you can't really hide behind attribution, you can't really hide behind anything, you just have to say, "Look, I messed up." I like having that specific type of accountability, but not everyone does for obvious reasons. I'm hesitant to even ask this question because I'm pretty sure I'm giving you a volleyball that you're just gonna spike down in my face, but the typical arguments against machine learning are, it's a black box. It's gonna lead to overfit models when you combine 140 characteristics. It's not appropriate for non-stationary data, which is what we would assume financial data is. I'm curious how you address these arguments. So, linear rich, grating, boosting, and random forests aren't black box. They're just not OLS. linear rich literally gives coefficients that you can look at. The grating, boosting, and random forest trees can be looked at and interrogated using impulse response functions, but we do have our own attribution system, and there's nothing really proprietary about it, so I'm happy to explain it. The idea is to build expected returns incrementally with traditional signal groups. Let's say value, then value plus profitability, then value plus profitability, plus accounting conservatism, then all of those signals plus low default risk, and so on. You then build optimized portfolios with these incremental expected returns, and you can attribute weights to each of the signal groups. Once you have the weight decomposition, you can produce backtests of returns from this and do a return decomposition, and this is really helpful in an intuitive way to attribute weights, and it does so in the same non-linear way that the model does, so captures everything. But no, you're not going to get very far attributing these portfolios with linear factors. But you really don't want all drawn analogy. Let's say you saw a picture of a draft and said, "That's a draft." And I said, "How do you know?" Well, you'd probably say, "It has a long neck and these brown patches against light coat for these horn-like things on top of its head." If I asked, "Yeah, but what lines make it a draft?" Point out the specific lines on the picture that make the singer draft. That doesn't actually make any sense. You can point out the aggregate features on the picture, but you can't point out specific lines on the picture that make the singer draft. In a similar way, this decomposition can point out the weight effects of groups of features, like value or profitability, holistically, but they're not going to give you linear relationships. Now regarding overfitting, I never really understood that in context of machine learning, because the biggest thing that machine learning emphasizes is tuning hyperparameters to cross-validation sets and predicting out of sample and attest. If you only allow yourself to predict on a walk-forward out of sample basis, what are you overfitting to? Are you overfitting hyperparameters? Well, I'll happily admit, I overfit some hyperparameters, and I'm treating model choice here as a hyperparameter. I only know that Lasso, PCR and PLS do not work well now. I did not know that in 1995 when our data started. I only know that trying to collapse the signal space from 150 signals down a fewer signals is not a good idea exposed. So if I had to start in 1995, I would have happily been mixing in good methods with bad methods and only figuring out that they were bad a decade later. But that's nothing compared to the overfitting that occurs when you don't use out of sample test. The normal way to test signals is in-sample, so you know a given signal performs well. If you test a signal in-sample and get a t-stat of 6, then how do you add that to your factor model without overfitting the backtests? We have added signals to our strategy that meet our economic bar for inclusion, and sometimes our backtests to return declines. Do we now remove that signal? No, because that's our new backtests with five BIPs lower return than before, too bad for the backtests. In the live portfolio, you want the model to have that information contained in that signal if indeed it did meet the economic bar for inclusion, which is the ability to predict future earnings. So in short, machine learning models are less prone to overfitting because they're evaluated on test sets. That is not to say that machine learning doesn't suffer from overfitting. You might hear the machine learning community complaining about model architectures from methods being overfit to a particularly commonly used data sets like image set. But that's not the level of sophistication that these critics of machine learning have. Those folks are testing signals in-sample, maybe doing some t-stat correction, and then complaining that machine learning overfits. That's utter nonsense. Now, I think you also brought up non-stationary data. First, it's optimal to have model decay. You want to overweight more res observations, but not by a lot. You want something like a 15-year half-life on the weight of your observations. But incredibly, even if you have an infinite half-life, which means no decay whatsoever, you do totally fine. And that was one of the bigger surprises we had. We expected the optimal model decay to be high and a super important parameter. But empirically, it's not. So it turns out the relationship between features and expected returns is fairly slow move. If you want to make your decay high, you're going to lose valuable information from those distant observations. I'm curious when you look at the portfolio over time, you mentioned already that you actually don't load on traditional factors. But do you find that the machine learning method is simply creating certain structural permanent non-linear characteristic tilts or those characteristics that it's leaning into are those dynamically changing over time given different market conditions? If you look at the non-linear loadings, according to our attribution, they are relatively stable over time, but they will change based on things like the underlying characteristic dispersion. Our linear factor loadings definitely change over time. So if you did a traditional factor attribution on a portfolio, you would see that the loadings don't explain the portfolio very well. And in so far as a model does have linear factor loadings, they are dynamic over time. This might be seen as some sort of cool feature of machine learning models with the model dynamically timing factors. But it says more about the limitations of linear factor attribution than it does about machine learning models. So despite having moved on from a more traditional factor-based approach to investing, you actually hold still some pretty strong views, specifically there's this ongoing debate about sort of multifactor portfolio construction, whether you should take an integrated approach or a mixed approach. We were chatting before we started recording, and you mentioned that you actually don't think either approach is correct. You prefer something called a stacking approach, which I had never heard of, and I was hoping you could take a moment to explain it to the listeners. Absolutely. By the way, I'm not entirely out of the smart beta game. For quad portfolios, I think we can and should jettison the factor approach and use machine learning models and characteristics to generate expected returns and optimization to build portfolios. But if you're making index products, you need to use things that are straight forward like factor source, because you have to explain it fully to an index provider and to clients. Now for what it's worth, when someone asks for a smart beta product, we can now use much more sophisticated methods. We can build expected returns of the factors. We can optimize the factor weights, we can use alternative data. Anyway, that disclaimer out of the way, let's get to mixing versus integrating versus stacking. So even though I'm sure everyone is familiar with this debate by now, let me quickly go over the advantages and disadvantages of mixing and integration. The mixing approach is averaging single factor smart beta portfolios, while the integrated approach is averaging factor scores, and then creating a single factor from that average of scores. And the advantage of the mixing approach versus the integrated approach is you get dynamic active weights based on how confident you are about various stocks. So if your factors completely disagree with each other, you will get very small afterwards, which you want because your signals disagree with each other. And by the way, bottom up, expected return models behave like this too. If you're expected return model, things that the market is efficient and there are no abnormal returns to be earned, it will not give you any active weights. That is a good thing. Another advantage of mixing is you can mix signals that are completely different rebalance frequencies. In theory, you could mix a high frequency trading signal that trades every second with a value signal trading every year because they're traded entirely separately. Now, of course, you want to trade your value signal every month because of rebalance timing luck, which you have written extensively on. But I'm just speaking hypothetically here, integrating can't mix strategies of different rebalance frequencies generally. Now, the disadvantage of mixing is that you're always losing active weights with each additional factor. That is, your active weights mechanically reduce. Since if two signals agree on a stock's active weight, its active weight will stay the same. But if they disagree, the active weight will decline. Moreover, if two signals want to more than zero out a stock, while another signal wants to give a small underweight to the stock. The stock will get a small positive weight in the portfolio, even though two of the signals would ideally want to negate the stock's weight. All the other signals doesn't even like the stock, but doesn't hate it. That doesn't seem bright either. So we came up with this method of stacking in 2016 before we even heard of this debate, and I'm sure many, many others have independently discovered it. It's not like mixing and integrating approaches never entered my consciousness, but we quickly decided stacking is superior. So what is stacking? You create long short portfolios from various signals you want to use in your strategy. And you multiply these long short portfolios by a string, let's say 12.5%, which would mean the portfolio is 25% long and 25% short for a 25% active weight for each factor. Then you sum up all the active weights across all the factors. Note that since stock weights within a factor sum to zero, aggregating factors will have a total weight of 0%. You add these summed weights to your benchmark weights, and since factor weights inherently summed to 0, your portfolio will have weights that summed to 1. However, some of those resulting weights might be negative. You zero those out, and you renormalize the positive weights. This method is what we call stack. Now this method has all the advantages of the other methods. If your factors disagree, your active weights decrease. If they agree, your factor weights increase. This dynamic agreement and disagreement contributes to your return and information ratio. You can test this empirically by forcing your acta weights to be constant. You will find that the naturally dynamic acta weights earn a higher information ratio than a static acta. Moreover, you can mix different rebounds frequencies together by starting with low rebounds frequency signals and stacking the higher rebounds frequency signals on top of the price drifted weights of the low rebounds before. Low rebounds frequency, portfolio. This might sound difficult, but it takes just an hour or two to code up. Also, with stacking, you don't lose active weights. You can either gain or lose active weights based on agreement or disagreement. If two factors agree, you'll add active weight. If they disagree strongly, you'll lose active weight. Finally, if a stock has a 2% weight in the portfolio, and two factors want to give it a negative 3% active weight each, and one wants to give it a 0% active weight, it no completely zero out the stock. The third factor that is indifferent about the stock won't give the stock back its positive weight. For what it's worth, if you allow your portfolio to short stocks or your active weights are extremely small, then the mixing and the stacking approaches differ only by a multiplier on activities. If you're a standard long-only portfolio, these approaches have very real differences. If you're a quant, I wouldn't worry about any of that. Focus on the machine learning stuff and expect to return as an optimization. If you're a smart beta manager, hopefully that was helpful. Well, very last question for you here. We're starting to see accelerated vaccine rollout. It seems like hopefully fingers crossed. The COVID pandemic is behind us and we'll all be reopening again on the road and meeting at conferences. Curious, what are you most looking forward to? This is maybe unrelated to going back outside, but I've just started playing around on Kaggle. Kaggle is a website that allows you to compete in machine learning competitions. As somebody who has come into machine learning, they didn't study it at university, it's really fascinating to learn and compete on these different things that are completely different from predicting stock returns. You might identify the glomerulus in a kidney or you might say whether a catheter was put in correctly or not, and I've just started engaging with this is incredibly difficult and interesting and fascinating, and I love trying something new and just getting my ass kicked on it and slowly working my way up, so I'm definitely looking forward to engaging more with that. Well, wonderful. Well, thank you so much for joining me today. I really appreciate it. Wonderful. Thank you. If you're enjoying the season, please consider heading over to your favorite podcast platform and leaving us a rating or review and sharing us with friends or on social media. It helps new people find us and helps us grow. Finally, if you'd like to learn more about newfound research, our investment mandates, mutual funds or associated ETFs, please visit thinknewfound.com. And now, welcome back to my ongoing conversation with Harley Bassman. In my experience, a lot more people focus on hedging the left tail, but very few discuss managing the right. Why do you find upside convexity so appealing? It is the case that markets tend to rise slowly, so the escalator up and they fall quickly elevator down. And if you do look at 1% moves or 2% moves, there's many more down 2% than up 2% moves. And that makes sense because if you think about it, everybody's long. Everyone owns something. Who's short? Well, corporations are short because they sold the stock. They don't hedge. Corporations are short. They issue the bonds. They don't hedge. Homeowners issue the bonds, the mortgage securities to us buyers. They don't go and sell their bathroom so they don't hedge. So everyone's long, the market is long, financial instruments. And therefore, it's a challenge for them to adjust their risk profile because what they have to do to reduce their exposure, they've got to sell it to someone else and have that person get a little longer risk. That's tough to do. And thus, you see skews in the market where the puts balls higher than the call ball. Aside from the general demand to fly the man, if he'll buy puts and selling calls, which drives skew, it's just the mere nature of risk. People are not risk neutral. There's a good dollar hurts more than making a dollar. And so therefore, it's very challenging to go and move risk around. And therefore, you tend to see these out of the money calls, which in theory nobody wants because they're already long in the market. Why do you want to pay to get longer? That option tends to trade very inexpensive, sometimes crazy cheap. And when you go and take that asset, that instrument, that risk, that path dependent convexity and add that to your portfolio, you could then go and get rid of some other linear risk and end up with a superior profile. Most times, you will see out of the money calls in the equity market trading well below realized volatility. And so a product like our SPY up is almost brilliant in the way that you're buying a core index. And then you're buying another money call. Those two together give you a very enhanced upward package. And so what you're theory you can do is instead of buying 50 of an ordinary index, you can buy 48 of hours. And then you have a downtrade, you're only long 48. In an up trade, you get long more of it, maybe long 52. And it doesn't cost that much to do that because that option is so inexpensive. And this is the case in most, not all, but most asset classes where you have that kind of show up. But you have this skew creating a very cheap way to buy optionality.