Preface
This is a very valuable article, so I’ve translated it for you.
Full text
On August 28, 2023, my company, Newfound Research, will be 15 years old. It feels a little ridiculous to say that. I know I’ve told this story before, but I never imagined the company would become what it is now. I started the company when I was an undergraduate and named it after Newfound Lake, a lake in New Hampshire that my family frequents. I fully expected the company to close within a year and I’d go to Wall Street to pursue a career.
But 15 years later, we’re still here. Somehow, this milestone feels bigger than any recent birthday I can remember. I’m so grateful for everything this company has given me. I’m thankful for my business partner, Tom, and I’m thankful for the people who have worked here and given part of their careers to this company. I’m thankful for the customers who have supported this company. I’m thankful for all the friends I’ve made in this industry. And I’m grateful for people like you who have given me a platform to explore ideas that I’m passionate about.
As this anniversary approaches, I’ve thought deeply about my career. One of them is that I’ve pondered all the lessons I’ve learned over the years. I thought a fun way to celebrate was to take the time to write down the ideas and lessons that have influenced my thinking.
So, without further ado, here are 15 lessons, ideas, and frameworks I’ve learned over the course of 15 years.
1. Risk cannot be eliminated, only transformed.
I am pursuing a master’s degree in computational finance at Carnegie Mellon University. This financial engineering program is an interdisciplinary collaboration between the finance, mathematics, statistics, and computer science departments.
In practice, it is a study of the theoretical and practical considerations of pricing financial derivatives.
I don’t quite remember when it dawned on me, but at some point I realized that there was a broader pattern in every task. The instruments we price always involve some degree of risk transfer. Our goal is to identify that risk, figure out how to isolate and extract it, package it into the right type of product, and then price it to sell.
Risk drives the whole equation. Pricing is about understanding the distribution of potential returns and trying to determine “fair compensation” for the various risks and assumptions we make.
For every buyer, there is a seller and vice versa. Ultimately, the seller who is unwilling to take the risk will have to compensate the buyer for taking the risk.
Ultimately, when you construct a portfolio of financial assets or even a strategy, you are expressing a view of the risk you are willing to take.
I liken the idea of a risky portfolio to a lump of Play-Doh. As you diversify your portfolio, the playdough smears across the risk space. For example, if you move from an all-equity portfolio to an equity-bond portfolio, you may reduce your exposure to economic contraction but increase your exposure to inflation risk.
The play dough doesn’t go away; it just gets spread around. By doing so, you become sensitive to more risks, but less sensitive to any single risk.
I should add that the idea of conservation of risk is by no means unique to me. For example, Chris Cole has said on several occasions, “Volatility is never created or destroyed, only transformed.” In 2008, James Saft wrote in Reuters, “Economic volatility is a bit like energy, it cannot be destroyed, it can only be transformed from one form to another.” In 2007, Swasti Kartikaningtyas, in an article on the role of central counterparties in the Indonesian market, noted, “A simple law of entropy in finance is that risk cannot be eliminated, it can only be transferred between parties.” In his 2006 book, Preventive Risk Management, Mark Jablonowski states, “Risk cannot be eliminated, it can only be divided.” In 1999, Clark and Varma, in their article on long-term strategic planning in business, said, “Risk, like matter, cannot be eliminated.”
My point is simply that this idea is by no means new or unique. But that doesn’t mean it isn’t important.
2. No pain, no premium
The concept of “no risk, no reward” simply reminds us that in the long run, we get paid for taking risks. And, eventually, the risk is likely to manifest itself and cause losses in our portfolios. After all, if there is no risk of loss, why would we expect to be rewarded above the risk-free rate?
Modern finance is largely based on the principle that the greater the risk taken, the higher the expected return. Most people seem to understand this notion instinctively when buying stocks and bonds.
But we can generally expect the same to hold true for many investment strategies. For example, value investors are likely to take a higher risk of bankruptcy when purchasing stocks, and that’s what gives them their returns.
What about for strategies that aren’t necessarily risk-based? What about strategies like momentum strategies that have more behavioral explanations?
At a higher level, we need to make sure that the strategy is hard enough to hold onto to prevent the premium from being arbitraged away. If an investment approach is seen as easy money, enough people will adopt it and the inflows will drive the excess returns away.
Therefore, it almost seems that at some point, specific strategies, especially low-frequency strategies, need to be difficult to stick to for any premium to exist. Pain is ultimately what keeps strategies from becoming overcrowded and allows the premium to exist.
3. Diverse, cheap BETA is just as valuable as equally diverse, expensive ALPHA.
I would categorize this lesson as “something that is obvious but probably needs to be mentioned”.
Our industry is obsessed with finding alpha returns. But in most cases, a portfolio doesn’t actually care if an investment is truly alpha or beta.
If you own a portfolio and can introduce a novel source of diversified beta, not only is it likely to be cheaper than any alpha you can access, but you’re likely to have more confidence in its risk premium.
For example, if you are invested only in equities, finding a way to wisely introduce bonds is likely to have a greater effect on your portfolio over the long term and lead to a higher level of confidence than trying to figure out a better way to pick stocks.
For most portfolios, beta will drive the majority of returns over the long term. Therefore, it will be more fruitful to exhaust beta sources first before looking for novel sources of alpha.
By the way, I’m pretty sure I stole the title of this lesson from someone, but I can’t find the person who originally said it. If it was you, please forgive me.
4. Diversification takes many forms
In 2007, Meb Faber explored the use of 10-month moving averages as a timing model for various asset classes in his paper “Quantitative Tactical Asset Allocation Methodology”. This will probably go down in financial history as one of the most well-timed papers ever published, given that it was closely followed by the 2008 crisis and how well the 10-month simple moving average model performed in protecting your capital from this event. This is probably the paper that spawned thousands of tactical asset allocation models.
In 2013, I wrote a blog post showing that the performance of this model is highly sensitive to the choice of rebalancing date, and Meb originally wrote the paper on a month-end rebalancing schedule. Theoretically, there is nothing stopping someone from running the same model and rebalancing on the 10th trading day of each month. In the blog post, I show how the strategy performs when applied to every possible trading day change in a month, from the first to the last trading day of the month. Even though the long-term returns were statistically indistinguishable, the short-term dispersion between these strategies was striking. And so my obsession with the luck of rebalancing timing was born.
Soon after, my good friend Adam Butler pointed out that the choice of a 10-month moving average was arbitrary. Why not 9, why not 11, why not 200 days? Why simple moving averages instead of exponential moving averages or simple time series momentum? As I have seen in rebalancing schedules, there is no statistical difference in long-term returns, but there is significant dispersion in short-term returns.
This dispersion can cause managers to exit the business. Ultimately, I see three dimensions to diversification: what to invest in, how to invest, and when to invest.
What to invest in is the traditional diversification that everyone is very familiar with. It’s diversification across securities or assets. This is what you invest in.
How to invest is the process of making investment decisions. This includes diversification across different investment styles - for example, value versus momentum - but it also includes diversification within styles. For example, how do we measure value? Or what trend model and velocity do we use? When to invest is the rebalancing schedule.
Just as traditional portfolio theory tells us that we should diversify because we won’t be compensated for taking on specific risks, I think the same applies to how and when we invest. Our goal should be to diversify ruthlessly across all non-compensated bets.
5. Philosophical Limitations of Diversification: If You Spread All the Risk, You Shouldn’t Expect Any Returns
One of the most common due diligence questions is: “When does this strategy not work?” This is an important question to make sure you understand the nature of any strategy.
But the fact that a strategy doesn’t work in certain circumstances is not in itself a criticism. It’s to be expected. If a strategy works all the time, and everyone does it, then it won’t work anymore.
Similarly, if you are building a portfolio, you need to take some risk. Whether that risk is some kind of economic risk, process risk, or path-dependent risk is irrelevant - it should be there, lurking in the background.
If you want a portfolio that has absolutely no risk in any scenario, you are essentially asking for a true arbitrage or an expensive way to replicate the risk-free rate.
In other words, if you eliminate all risk from your portfolio - and again, thinking about this is like rubbing a ball of clay really, really, really thin across a very large plane of risky scenarios - then returns should only converge to the risk-free interest rate.
If not, you have an arbitrage opportunity: simply borrow at the risk-free rate and buy your risk-free, diversified portfolio.
But arbitrage opportunities don’t come easily. Especially for low-frequency strategies and portfolios of low Sharpe ratio asset classes. There is no magic combination of assets and strategies that will eliminate all downside risk in all future states of the world.
A corollary to this argument is what I call the frustration law of active management. The basic idea is that if an investment idea is perceived to be both alpha and “easy”, investors will allocate to it and erode the associated premium. This is just basic market efficiency.
So how does a strategy become “difficult”? Well, the manager may have a significant informational or analytical advantage. Or the manager may have structural barriers to accessing deals that others don’t have the opportunity to pursue.
But for most major low-frequency edge advantages, the “difficulty” will be behavioral. The strategy must be persistent enough to avoid being destroyed by arbitrage.
This means that in order for any disciplined investment approach to achieve excess returns over the long term, it must experience periods of low returns in the short term.
But we can also say the opposite: in order for any disciplined investment approach to underperform in the long run, it must experience periods of excess returns in the short run.
The frustration for active managers is not only that their investment approach will have to underperform from time to time, but also that poor strategies will have to outperform. The latter may seem confusing, but consider that an intentionally poor strategy can simply be reversed - or shorted - to create an intentionally good strategy.
6. It’s usually the unconscious betting that breaks you.
I once read a comic strip - I think it was Sidebar, but I’ve never been able to find it - that joked about the end of the world coming after a bunch of scientists in a lab said, “Great, it worked!” and then it came.
It’s rarely what we plan to do that knocks us down. Instead, it’s more the unintended bets that sneak into our portfolios - the things we don’t realize until it’s too late.
For example, in the mid-2010s, it was widely believed that European stocks were very cheap compared to U.S. stocks. However, investors who were keen on European equities were penalized.
Simply swapping U.S. stocks for foreign stocks introduces a significant currency bet. Europe may indeed be undervalued, but given that valuation reversions typically take years to materialize, any analysis of this trade should include the cost of hedging the currency exposure, or at least an opinion on why the implicit shorting of the dollar is a worthwhile bet.
But it could be argued that the analysis itself is wrong. Lawrence Hamtiel has written on this topic many times, pointing out that both cross-country and time-series valuation ratio analyses can be grossly skewed by industry differences. For example, U.S. equity indices tend to be more heavily invested in the technology sector, while European indices are more heavily invested in consumer goods. When the sector differences are normalized, the valuation gap narrows considerably.
Those who choose Europe versus the U.S. intend to make valuation bets. Unless they were careful, they also made currency and sector differential bets.
It’s rarely the intended bet that brings you down.
7. Portfolio all the way down
I don’t remember when this idea came about, but it’s one of my favorite mental models. It’s a funny way of saying “there are 18 levels of hell below the basement”.
Every portfolio and its decisions can be broken down into being bullish on some things and bearish on others.
This may sound bland, but it’s extremely powerful. Here are some examples.
- You are evaluating a new buy-only active fund. To isolate the fund manager’s operations, you can take the fund’s position and subtract their benchmark position. The result is a net-neutral long-short portfolio that reflects the manager’s active bets - long their overweights and short their underweights. This can help you determine the type and size of their bets and whether they cover their expenses. 2.
- If you intend to sell one position for another, the trade is the same as holding an existing portfolio and overlaying it with a long/short trade: long what you want to buy and short what you want to sell. This allows you to examine the attributes of the overall trade (both the additions and subtractions). 3.
- If you want to understand the contribution of the different steps of portfolio construction to risk or return, you can progressively view the changes as a long-short portfolio. For example, for a portfolio of equal-weighted holdings of 50 stocks in the S&P 500, you could compare (1) equal-weighted S&P 500 minus S&P 500, and (2) equal-weighted 50 stocks minus equal-weighted S&P 500. Separating each construction step into longs and shorts allows you to understand the return properties of that step. In all of these examples, evaluating the portfolio through a long-short framework provides meaningful insight.
In all of these examples, evaluating the portfolio through a long-short framework can provide meaningful insight.
8. The more diversified the portfolio, the higher the hurdle rate for time selection
Timing strategies may be the most tantalizing song in finance. It sounds simple enough. Whether it’s a market risk reserve or some kind of investment strategy, we’d like to say, “When the timing is bad, don’t do that.”
After all, all stock factors were popularized in the 2010s when factor timing became popular. I’ve read many articles that say you can buy certain factors at certain stages of the economic cycle. One article used a large number of economic indicators that could be easily tracked to simultaneously define which phase of the economic cycle you were in, and then rotated different factors based on the system.
And the performance was phenomenal.
So, to test this idea, I decided to run opposing scenarios. If I kept the definition of economic regimes the same, but purchased a basket of factors at each stage of the cycle completely at random, then with just a few factors, four regimes, and three baskets of factors purchased by each regime, you could violently traverse all possible combinations and create a distribution of their return outcomes.
Not surprisingly, the results in that article are in the top percentile. You know what else? A plain, equally weighted combination of factors. That’s when you should ask yourself, “How confident am I in this approach?”
Because the evidence shows that beating plain vanilla diversification is really, really hard.
There are a few ways to get a sense of this, but one of my favorites is to look clearly into the future and ask yourself, “How accurate do I need to be to beat a well-diversified portfolio?” It’s not a difficult scenario to model, and for a reasonable level of diversification, the accuracy numbers climb quickly.
Ultimately, timing is a very low-breadth exercise. To quote AQR’s Michele Aghassi: “You’re comparing the present with the usual.” And mistakes accumulate forever.
In almost all cases, it’s easier to find things that diversify your returns than it is to improve the accuracy of your return forecasts.
As a corollary to this lesson, I would add that the more predictable something is, the harder it is for you to profit from it.
For example, suppose I have a system that allows me to predict the economic system we live in, and I have a model that predicts which assets will do well in that economic system.
If I can accurately predict the economic system and the market is relatively efficient, I probably shouldn’t be able to know which assets will do well under which system. Conversely, if I can be absolutely certain about which assets will perform well under which regimes, I may not be able to accurately predict those regimes.
If markets are even relatively efficient, then the easier it is to predict something, the less likely I am to profit from it.
9. Certain signals are only valuable in extreme situations.
I recently received a chart that plots the valuations of large-cap, mid-cap and small-cap stocks in the United States. Valuations are expressed as the average combined z-value of price to yield, price to book value and price to sales. The average z-value for large-cap stocks is about +1, while the average z-value for small- and mid-cap stocks are all about -1.
The implication of this chart is that it may be wise to rotate small and mid-cap stocks based on these relative valuations.
This immediately brings me to lesson #6 about unintentional betting.
For example, are historical measures still relevant today? Prior to 2008, the large-cap index was heavily weighted towards financials and energy stocks. Today, the index is dominated by technology and communication services. And we’ve had a zero interest rate policy era for a full decade. How do interest rates above 5% today affect the refinancing opportunities for small-cap vs. large-cap stocks? What about the industry differences between large caps and small caps? Or profit margins? Or exposure from foreign revenue sources? How does this analysis treat unprofitable companies? Is the price-to-sales ratio still a useful metric when sales are generated across the entire organization?
You may be able to adjust your analytical methods and figures to take many of these points into account. But there may be many other points that you haven’t thought of at all. This is noise.
Almost every signal has noise.
The question is, “How much noise?” We think that the more noise a signal has, the stronger we need the signal to be in order to croire it to have any effect. When we can trade an accurately measured signal within one standard deviation, we may only have confidence in a crudely measured signal at a higher significance level.
10. Under strong uncertainty, a 50-50 split may be an optimal decision.
During the factor wars of the mid-2010s, there was a heated debate among firms around the optimal method of portfolio construction - hybrid or integrated.
The hybrid approach argued that each factor should be constructed separately and placed in its own position.
The integrated approach argues that stocks should be scored on all the factors simultaneously, choosing the ones with the highest composite scores.
Both views have strong supporters. Goldman Sachs supports the hybrid approach and AQR supports the integrated approach.
I spent months agonizing over which method to use. I read papers, did empirical analyses, and even derived the expected factor efficiencies under each approach.
In the end, I couldn’t convince myself that I was completely on either side. So, what did I do? Fifty-fifty.
Half of the portfolio is managed with a hybrid approach and half with an integrated approach.
The real reason is that it’s just a decentralization of the decision-making process.
Whenever I’m faced with highly uncertain choices, I often fall back on the 50/50 approach.
When I argue about which option structure to use, there is no clear winner, I go 50/50. When I argue that I should use two completely different modeling methods, there is no clear winner, I go 50/50.
Fifty-fifty at least provides an incremental step in the decision-making process and implicitly creates dispersion to help hedge against uncertainty.
11. always ask: “What’s the deal?
In July 2019,Greek 10-year Treasury yields were nearly identical to U.S. 10-year Treasury yields.
By December, the Greek 10-year Treasury yield was 40 basis points below the U.S. 10-year Treasury yield. How can this possibly make sense? How can a country like Greece make U.S. bonds look like high-yield bonds?
When something seems ridiculous, ask yourself a simple question: What is the deal? If it is ridiculous, how do we profit from it?
In this case, we could have considered going long the US 10-year and short the Greek 10-year in a convergence trade. But we soon realized an important factor: what you actually gain is not in percentage points, but in currency. At that point, the trade suddenly became problematic. In this case, you will receive dollars and owe euros. If you try to pre-hedge the trade with a cross-currency basis swap, any difference in earnings will largely disappear.
Perhaps a more relevant financial figure is the spread between Greek and German 10-year government bonds in the second half of 2019, which trades between 150-275 basis points. No longer completely unreasonable.
When financial commentators talk about the absurdity of the market, ask yourself, “What is the trade?” By thinking about how to profit from the absurdity, it is often possible to bring to the surface the reasons for analyzing the error.
12. The trade-off between Type I and Type II errors is asymmetric
Academic finance is obsessed with Type I errors. The literature is littered with Alpha strategies that are significant at the 5% level. The literature wants to avoid reporting false positives.
However, in practice, there is an asymmetry that must be considered.
What is the cost of false positives? Unless the strategy is deliberately chosen, the performance of trading a false positive should be just the noise minus the transaction costs. (And the opportunity cost of capital.)
What is the cost of a false negative? We missed the Alpha.
Now consider how the first type of false concern can bias your choice of strategy. Are they easier to data mine? Are they more likely to be crowded? Are they less likely to incorporate new market characteristics without meaningful history?
Once we recognize this asymmetry, it may actually be wise to lower the statistical requirements of the strategies we implement.
13. Behavioral time is decades longer than statistical time
I recently stole this from Cliff Asness. This has very little to do with any actual portfolio construction ideas or useful mental models. It simply recognizes that managing money in real life is very, very, very different from managing money in a research environment.
In backtesting, it’s easy to look at a multi-year pullback period for a low Sharpe ratio strategy and say, “I can live with that”. When it’s a multi-decade simulation, a few years can seem like a small ripple - just a statistical event on the path of the curve from bottom left to top right. As your eye wanders over the return curve, you experience the years of retracement in your mind’s eye in just a few seconds.
However, in the real world, a multi-year retracement feels like a multi-decade retracement. When you say, “This performance is within the standard confidence interval of our expected Sharpe Ratio strategy, and we can’t find any evidence of a problem with our process,” that’s not much comfort to those assigned to you. The client will ask you to perform an attribution analysis. The client will ask you if you’ve considered explanation X or Y. Sales will plummet. Clients will redeem.
For anyone considering a career in money management, it’s important to time adaptive behavior.
14. Jensen’s Inequality
Jensen’s Inequality basically says: “The function applied to the mean is not necessarily equal to the mean after the function is applied”.
What does this mean and how is it useful? Consider this example. You are building a simple momentum portfolio.
You start with the S&P 500, rank it based on its momentum score, pick the top 100 stocks, and then equalize the weights.
But you remember lesson 4 and decide to use multiple momentum signals to diversify your risk.
The question here is: do you average all the momentum scores and then pick the top 100, or do you create a portfolio using each momentum signal and then average those portfolios?
Jensen’s inequality tells us that these two approaches lead to different results. This is essentially the hybrid vs. integration debate from Lesson 10. The more convex a function is, the greater the difference in results is likely to be. Imagine if we had chosen the top 20 or the top 5 instead of the top 100, the combinations could have been very different depending on the momentum signals.
Here’s another simple example. You have 10 buy/sell signals. Your function is to go long if the signals are positive and short if the signals are negative.
If you average your signals first, your position is binary: always open or closed. But if you apply your function to each signal and then average the results, you end up with a series of weights whose distribution will depend on the correlation between your signals.
You can see that Jensen’s inequality plays a huge role in portfolio construction. Why? Because nonlinearities are everywhere. Portfolio optimization? Nonlinear. Maximum or minimum positions? Nonlinear. Ranking-based cutoffs? Non-linear.
The more nonlinear the function, the bigger the gap. But it also helps us to understand how certain portfolio construction constraints can help us reduce the size of that divide.
Ultimately, Jensen is telling us that averaging things out before or after the convex step of your process can lead to completely different portfolio outcomes in order to achieve diversification.
15. Backtesting is just a lottery for a randomized process.
As the saying goes, no one has ever seen a bad backtest.
Our industry as a whole has good reason to be skeptical of backtesting. Almost every experienced quant trader can tell you a story about a time when they were young and naïve enough to run backtests, desperately searching for the holy grail strategy to overfit and overoptimize.
Less sophisticated actors may even launch products with these backtests, using the results as marketing materials to sell to potential investors.
Most investors would be perfectly justified in simply ignoring them. I might even be in favor of banning the display of these backtests.
But that doesn’t mean that backtesting is ultimately futile. We should recognize that when we run a backtest, it’s just a sample of a larger random process. After all, historical prices and data are a record of what happened, not a complete picture of what could have happened.
Our job as researchers is to use backtesting to try to understand what the underlying stochastic process looks like.
For example, what happens when you change the parameters of our process? What happens when we change our entry or exit times? Or what happens when we change our slippage and impact assumptions?
One of my favorite techniques is to change the investable universe, randomly removing a portion of the universe to see the sensitivity of the process. Similarly, randomly remove time periods from the backtest to test regime sensitivity.
Injecting this randomness into the backtesting process can tell us how much of an anomaly our single backtest actually is.
Another excellent technique is to intentionally introduce foresight bias into your process. By explicitly using the ability to see into the future, we can find theoretical upper bounds on our achievable results and develop confidence intervals for more reasonable accuracy assumptions that our results should have.
Poorly done backtesting is worse than no backtesting at all. You’re better off trying to reason your way through the process with a pen and paper. But in my opinion, well-done backtesting gives you a pretty good idea of the nature of your process, which is what we ultimately want to understand.
16. The market is usually right
Yes, did I say 15 ideas and lessons? Here’s an additional lesson that took me longer to learn than I care to admit.
The market is usually right in most cases. Repeatedly applying Lesson 11, “The Nature of the Trade,” has made me realize that most seemingly ridiculous things may not be so.
That’s not to say there aren’t exceptions. If we see a $20 bill on the ground, we can certainly pick it up, and cash and futures trading in cryptocurrencies in the year 2021 instantly comes to mind. With limited institutional capacity and a nearly insatiable demand for leverage from retail investors, the implied financing rates for perpetual contracts and futures on highly liquid tokens such as bitcoin and ethereum have reached over 20%. I doubt that’s as close to free money as I’ll ever come across.
But that’s usually the exception.
This last lesson was a mind shift for me. Instead of saying “the market is wrong” as soon as I saw something, I assumed that the market was right and that it was me who was wrong. This forced me to list a number of potential causes that I might have missed or overlooked, and to eliminate these explanations before building confidence that the market was indeed wrong.
Conclusion
If you have done this, thank you. I appreciate your generosity of time. I hope some of these ideas or lessons resonate with you, and I hope you enjoyed reading them as much as I enjoyed reflecting on the concepts and putting together this list. It will be interesting for me to review the other 15 and see how many of them have stood the test of time.
Until then, happy investing.
