Autoencoders for Conditional Risk Factors and Asset Pricing

This chapter shows how unsupervised learning can leverage deep learning for trading. More specifically, we'll discuss autoencoders that have been around for decades but have recently attracted fresh interest.

Unsupervised learning addresses practical ML challenges such as the limited availability of labeled data and the curse of dimensionality, which requires exponentially more samples for successful learning from complex, real-life data with many features. At a conceptual level, unsupervised learning resembles human learning and the development of common sense much more closely than supervised and reinforcement learning, which we'll cover in the next chapter. It is also called predictive learning because it aims to discover structure and regularities from data so that it can predict missing inputs, that is, fill in the blanks from the observed parts.

An autoencoder is a neural network (NN) trained...

Autoencoders for nonlinear feature extraction

In Chapter 17, Deep Learning for Trading, we saw how neural networks succeed at supervised learning by extracting a hierarchical feature representation useful for the given task. Convolutional neural networks (CNNs), for example, learn and synthesize increasingly complex patterns from grid-like data, for example, to identify or detect objects in an image or to classify time series.

An autoencoder, in contrast, is a neural network designed exclusively to learn a new representation that encodes the input in a way that helps solve another task. To this end, the training forces the network to reproduce the input. Since autoencoders typically use the same data as input and output, they are also considered an instance of self-supervised learning. In the process, the parameters of a hidden layer h become the code that represents the input, similar to the word2vec model covered in Chapter 16, Word Embeddings for Earnings Calls and SEC Filings...

Implementing autoencoders with TensorFlow 2

In this section, we'll illustrate how to implement several of the autoencoder models introduced in the previous section using the Keras interface of TensorFlow 2. We'll first load and prepare an image dataset that we'll use throughout this section. We will use images instead of financial time series because it makes it easier to visualize the results of the encoding process. The next section shows how to use an autoencoder with financial data as part of a more complex architecture that can serve as the basis for a trading strategy.

After preparing the data, we'll proceed to build autoencoders using deep feedforward nets, sparsity constraints, and convolutions and apply the latter to denoise images.

How to prepare the data

For illustration, we'll use the Fashion MNIST dataset, a modern drop-in replacement for the classic MNIST handwritten digit dataset popularized by Lecun et al. (1998) with LeNet. We...

A conditional autoencoder for trading

Recent research by Gu, Kelly, and Xiu (GKX, 2019) developed an asset pricing model based on the exposure of securities to risk factors. It builds on the concept of data-driven risk factors that we discussed in Chapter 13, Data-Driven Risk Factors and Asset Allocation with Unsupervised Learning, when introducing PCA as well as the risk factor models covered in Chapter 4, Financial Feature Engineering – How to Research Alpha Factors. They aim to show that the asset characteristics used by factor models to capture the systematic drivers of "anomalies" are just proxies for the time-varying exposure to risk factors that cannot be directly measured. In this context, anomalies are returns in excess of those explained by the exposure to aggregate market risk (see the discussion of the capital asset pricing model in Chapter 5, Portfolio Optimization and Performance Evaluation).

The Fama-French factor models discussed in Chapter 4 and...

Summary

In this chapter, we introduced how unsupervised learning leverages deep learning. Autoencoders learn sophisticated, nonlinear feature representations that are capable of significantly compressing complex data while losing little information. As a result, they are very useful to counter the curse of dimensionality associated with rich datasets that have many features, especially common datasets with alternative data. We also saw how to implement various types of autoencoders using TensorFlow 2.

Most importantly, we implemented recent academic research that extracts data-driven risk factors from data to predict returns. Different from our linear approach to this challenge in Chapter 13, Data-Driven Risk Factors and Asset Allocation with Unsupervised Learning, autoencoders capture nonlinear relationships. Moreover, the flexibility of deep learning allowed us to incorporate numerous key asset characteristics to model more sensitive factors that helped predict returns.

...