What Matters Is What Is Yet To Come

{ Euclidean Q3 2019 Letter }

If you use Google’s email service (Gmail), you know it wasn't long ago that Gmail started creating pre-written replies for you. Your friend emails, “Let's grab lunch tomorrow?” and Gmail presents you with three canned responses: “Sure!” “Can't tomorrow,” and “Another day this week?” But more recently, Gmail has started completing your entire sentences. You type, “What is your favorite...” and suddenly, to the right of your message and in greyed-out font, the following text appears: “restaurant in midtown?”

The technology behind this feature is the product of very exciting innovations in the application of deep neural networks. These “language models” are characterized by a strikingly simple idea: given a sequence of words (e.g., a sentence fragment), we can train a neural network to predict the most probable next word in the sentence. Then the new sentence fragment (with the additional predicted word appended to the end) is fed back into the neural network to predict the next word. And this process is repeated until the sentence is complete or continues indefinitely to compose a paragraph, page, or larger corpus of text. Besides the Gmail team, Elon Musk’s Open AI has been feverishly building cutting-edge language models. You can give them a try here.

It was this simple idea, using deep neural networks to predict the next step in a sequence from the sequence itself, that inspired Euclidean two years ago to begin investigating a new approach to systematic value investing.

Backward-Looking Quantitative Value Investing

If you look at the history of public companies and their investment outcomes, you confront an unmistakable message about how various styles of investing have worked over time. Specifically, you notice that investing in companies offered at low prices in relation to recent operating results bubbles its way to the top in terms of persistent, long-term performance. By adhering to this type of contrarian or “value-oriented” strategy, you would have done remarkably well across long periods. You would have also realized returns that would have dwarfed those achieved by investing in more exciting or fast-growing companies priced expensively in relation to their recent financial results [1].

Yet, even if growth investors’ long-term performance leaves much to be desired, they are right about one important thing. That is, a company’s future results – and not its past performance – determine the company’s future value. And so, it is logical to try to figure out how companies might grow and evolve, with the hope of identifying opportunities for paying a reasonable price today for a bounty of future earnings.

The rub, though, is that predicting a company’s future results has been really hard to do, much like predicting what the next words in your sentence might be. We believe this is a big reason why systematic value strategies have outperformed growth strategies over the long-term [2]. Quantitative value may have done well because it is essentially a bet against others’ ability to predict the future.

Think about it this way: When companies are overly inexpensive or expensive in relation to their fundamentals, they are priced in such a way for a reason. That reason is that investors have established a consensus view on how the world is going to look going forward. That view embeds assumptions that prop up the prices of expensive companies and hold down the prices of inexpensive ones. But investors aren’t very good at accurately predicting the future. Thus, eventually, something surprising happens that challenges the consensus view and proves some of its embedded assumptions wrong. This erodes some of the lift – and alleviates some of the pressure – that previously caused certain companies to be priced, respectively, at premiums or discounts to their fundamentals.

So, the “backward-looking” quality of traditional quantitative value strategies has been a feature, and not a bug. By relying on what is known (past results) versus what is unknown (future results), these strategies have essentially taken the other side of the “prediction bet”; that has been a fruitful position for a very long period of time.

Trying for a Glimpse into the Future

Although these merits of quantitative value investing are clear, there is one caveat we can call out for certain. If a quantitative-value investor knew that a company’s future earnings would be 50% less than its trailing twelve-month earnings, that investor would choose the future earnings over the historical earnings when valuing the company.

When a quantitative value model uses trailing twelve-month earnings in a valuation model, it is doing so because those earnings are thought to be a good proxy for future earnings. Many quant-value investors will use secondary fundamental factors (such as return on invested capital, debt to EBITDA, etc.) to rule out or discount the attractiveness of a company because they believe those factors help identify when a company’s past earnings may not be a good proxy for its future earnings.

In this sense, quantitative values models are essentially forecasting future fundamentals — just not explicitly.

Future-Looking Quantitative Value Investing

Observing the strategy behind the deep learning–based language models we described earlier, where a sequence of words is used to predict the next word, we saw a path for advancing quantitative value models in a forward-looking direction. Instead of relying on backward-looking factors to rank stocks, we aspired to forecast a company’s future financial results explicitly via a deep neural network and use those forecasts to identify attractively mispriced investment opportunities.

We have previously written about our research to this end, most recently describing the positive impact we have seen quantifying the concept of margin-of-safety using forward-looking uncertainty cones. In that letter, we shared the output from an exercise to estimate the theoretical upper limit of investment performance if you could accurately predict earnings one year into the future. Indeed, and as you might think, it seems that the better you can forecast earnings, the greater your opportunity for future returns.

We were then asked by some of you whether predicting a company’s future cash flows might be more fruitful than predicting earnings. We thought this was a worthwhile question to answer. After all, as our good friend Tren Griffin says, “Cash flow rather than reported earnings is what determines value for an investor. You can’t spend someone’s opinion that a business generated an accounting profit. Earnings are only a clue about future cash flows.”

So, we ran the same experiment with free cash flow and share a view into it below [3].

Here are the three things we would like to point out from the above clairvoyant plot:

It seems that you could do very well if you could accurately predict either earnings or cash flow one year in the future. There is no difference at 12-months, although over a longer prediction horizon, predicting earnings appears to have more return potential than predicting cash flows.
In simulation, our models today capture only a small part of the potential implied here. This is why we are continually finding ways to improve our forecasts. These clairvoyant studies suggest that, as our forecasts improve, there appears to be room to increase our future expected returns.
While the opportunities associated with making 12-month forecasts of earnings and cash flows have similar potential, it may be that one is easier to predict than the other. Put another way, we might find that we can move closer to the theoretical limit on the dimension of cash flow than we can with earnings. Perhaps this is a subject we will discuss in a future letter.
***

At Euclidean, we have developed a set of deep learning technologies that forecast future earnings better than the consensus forecast and improve on, in simulation, a quantitative factor model that uses backward-looking financials. Additionally, we have seen that if we quantify the uncertainty in these forecasts, we improve the performance even more. However, our hypothetical perfect forecaster studies (clairvoyant studies) suggest that there is still much room for improvement and, therefore, in the coming year, we will invest heavily in advancing our forecasting models by examining new datasets, new deep learning architectures, and new forecasting and machine learning techniques.

Best Regards,

John & Mike

The opinions expressed herein are those of Euclidean Technologies Management, LLC (“Euclidean”) and are subject to change without notice. This material is not financial advice or an offer to purchase or sell any product. Euclidean reserves the right to modify its current investment strategies and techniques based on changing market dynamics or client needs.

Euclidean Technologies Management, LLC, is an independent investment adviser registered under the Investment Advisers Act of 1940, as amended. Registration does not imply a certain level of skill or training. More information about Euclidean, including our investment strategies, fees, and objectives can be found in our ADV Part 2, which is available upon request.

back to letters

[1] & [2] A Closer Look at Value Premium: Literature Review and Synthesis & The Current Market Environment

[3] We ran simulations with a clairvoyant model that has the ability to access future financial reports. To be clear, this was nothing more than an exercise to estimate the theoretical upper limit of investment performance if one could perfectly predict earnings one year into the future. These simulations were run for the period of January 1, 2000, to December 2014. The ending date is in 2014 because a simulation that has 48 months of clairvoyance needs to be able to “see” four years into the future. In Figure 1, we show that an EBIT (earnings before interest and taxes) / EV (enterprise value) and FCF (free cashflow) / EV factor model that uses a perfect 12-month forecast (a clairvoyant forecast) for EBIT and FCF would, if possible, achieve a 40.2% and 40.3% compound annualized return, respectively. That can be compared to a 14.7% and 11.3% annualized return over the same period for a traditional EBIT/EV and FCF/EV factor model, respectively, using trailing 12-month earnings instead of future earnings and free cashflow. As you can see, the further into the future one can predict financial performance, the higher the theoretical returns appear to be.

What Matters Is What Is Yet To Come

{ Euclidean Q3 2019 Letter }

Backward-Looking Quantitative Value Investing

Trying for a Glimpse into the Future

Future-Looking Quantitative Value Investing

Euclidean Fund I, LP Performance

Look Through Financials, as of Quarter End

Ten Largest Holdings, as of Quarter End

net performance, fund lifetime