Findings from our Research on Applying Deep Learning to Long-Term Investing

 

Euclidean has spent the last year working on a thorough investigation of the application of deep learning techniques to long-term investing. One outcome of this process is a peer reviewed paper that will be published this December in the proceedings of the Time Series Workshop at the 2017 Neural Information Processing Systems conference. This paper was a collaborative effort between Euclidean and Zachary Lipton, currently a scientist at Amazon AI and assistant professor at Carnegie Mellon University.

You can listen to John and Zack discuss the paper on Bloomberg's "Odd Lots" podcast.

 

A Quick Summary 

In this paper, we introduce how we use deep learning techniques to predict future company fundamental data, such as earnings, revenue, and debt, from past fundamental data.  We also show that these predictions can be used to meaningfully improve the investment performance of widely researched and commercially applied quantitative investment strategies that use valuation ratios.

The results reflect our use of long-short-term-memory recurrent-neural networks (RNN) and multi-layer perceptrons (MLP) to predict future company fundamentals from a time-series of past company fundamentals. We were motivated to do this by the intuition that future fundamentals should have a bigger impact than current fundamentals on a stock’s future price. We confirm this intuition through an empirical study demonstrating that knowledge of future fundamentals (a clairvoyant prediction) would, if possible, dramatically improve the investment performance of simulated portfolios. We then show that portfolios constructed with valuation ratios that use predicted fundamentals outperform portfolios constructed with valuations ratios based on current fundamentals. If you would like to skip all the technicalities of the paper, here are the out-of-sample results over the period 2000-2016:

 
 

As a quick bit of orientation, note these definitions:

  • S&P 500: the total return of the S&P 500 over the period.
     
  • Market Average: the equal weighted return of a portfolio of all stocks in the paper's investment universe.
     
  • Price-LSTM: a recurrent neural network using the fundamentals outlined in the paper in order to predict relative return directly (as opposed to trying to predict next period fundamentals).
     
  • QFM or quantitative factor model: a traditional factor model that ranks companies on trailing 12-month EBIT divided by current Enterprise Value.
     
  • LFM-Linear: a linear neural network predicting fundamentals at the next time step and using predicted EBIT divided by current Enterprise Value to rank companies.
     
  • LFM-MLP: a multi-layer perceptron predicting fundamentals at the next time step and using predicted EBIT divided by current Enterprise Value to rank companies.
     
  • LFM-RNN: a recurrent neural network predicting fundamentals at the next time step and using predicted EBIT divided by current Enterprise Value to rank companies.
     
  • MSE and CAR are the mean squared error of the fundamental predictions and compound annual return of the simulated portfolios, respectively.

Research Lessons

Through this research we learned some lessons about where deep learning can be used successfully and where it cannot when applied to questions relating to long-term investing.   

 

Lesson 1 – Directly Predicting STOCK RETURNS with Deep Learning

One big lesson is that using deep learning to directly predict returns does not perform any better than what can be acheived with a linear model. This may be because there is so much noise in the relationship between fundamentals and price changes. As we suspected that the relationships between fundamental data may have a larger signal to noise ratio than the relationship between fundamentals and price, we took the example of sequence-to-sequence prediction in natural language processing and considered the idea of predicting future fundamentals from past fundamentals and then using the predicted future fundamentals as inputs to a factor model.  This is where we found better results.  

 

Lesson 2 – Predicting Future Fundamentals Has Real Value

Formulating the problem in this way also led to the question: if we were able to predict fundamentals perfectly (clairvoyantly), how large is the performance opportunity?  Well, it may not surprise you to find that one could (in theory of course) achieve outstanding performance with clairvoyant access to future fundamentals.  We reference it here, however, because there has been a lot of discussion within the finance community as to whether there is any value in the use of reported earnings in investing models.  Our results suggest that the better you can predict future fundamentals, the greater your opportunity for future returns and this, of course, motivates additional research.

 

Lesson 3 – DEEP Neural Networks Outperform Linear Models

We observe, and share results explaining, that deep and recurrent neural networks can predict next period fundamentals better than what can be achieved with linear models. Furthermore, we also found that using the deep neural network’s predictions of next-period EBIT in a factor model outperforms the naive approach of using trailing twelve months EBIT. The implications of this are encouraging for further exploration of deep learning techniques in long-term investing.