[Paper Review] Deep Learning in Characteristics-Sorted Factor Models

Guanhao Feng, He, Jingyu+2 more|arXiv (Cornell University)|May 3, 2018

Financial Markets and Investment Strategies37 references50 citations

TL;DR

The paper develops a structural deep learning framework that generates latent factors from firm characteristics to price cross-sectional asset returns, minimizing realized pricing errors within an augmented deep factor model.

ABSTRACT

This paper presents an augmented deep factor model that generates latent factors for cross-sectional asset pricing. The conventional security sorting on firm characteristics for constructing long-short factor portfolio weights is nonlinear modeling, while factors are treated as inputs in linear models. We provide a structural deep learning framework to generalize the complete mechanism for fitting cross-sectional returns by firm characteristics through generating risk factors -- hidden layers. Our model has an economic-guided objective function that minimizes aggregated realized pricing errors. Empirical results on high-dimensional characteristics demonstrate robust asset pricing performance and strong investment improvements by identifying important raw characteristic sources.

Motivation & Objective

• Motivate and formalize the characteristics-sorted factor methodology within a deep learning framework.
• Develop a structural neural network that creates deep characteristics, sorts securities, and generates latent factors.
• Price cross-sections of stock returns by minimizing aggregate realized pricing errors.
• Align deep learning architecture with economic no-arbitrage principles in asset pricing.

Proposed method

• Introduce a feed-forward neural network that reduces input characteristics to deep characteristics.
• Implement a nonlinear sorting mechanism to create long-short portfolio weights via the deep characteristics.
• Construct deep factors from sorted long-short portfolios and realized returns.
• Model dynamic factor loadings as neural networks conditional on lagged characteristics.
• Optimize the objective to minimize the quadratic sum of realized pricing errors, with regularization.
• Specialize the framework to embed the Fama-French five-factor model as a deep learning variant.

Experimental results

Research questions

• RQ1 Can a structural deep learning model generate latent factors from firm characteristics that minimize pricing errors in cross-sectional asset pricing?
• RQ2 How do deep characteristics and nonlinear sorting affect the explanatory power of factor models compared with benchmark models like CAPM and FF5?
• RQ3 What is the incremental pricing performance when adding deep factors to traditional benchmarks?
• RQ4 How robust are the results to missing data and unbalanced panels?

Key findings

• The augmented deep factor model achieves robust improvement in pricing both for individual stocks and test portfolios.
• The MVE portfolio of deep factors with DL-CAPM or DL-FF5 yields high in-sample Sharpe ratios (4.57 and 6.49) and out-of-sample Sharpe ratios (2.49 and 3.00).
• The deep factors provide perfect hedging against benchmark models in the reported scenarios.
• Consistent linear and nonlinear exposures are found for raw characteristics such as earnings surprises and book-to-market ratios.
• The framework maintains comparable statistical and economic model fitness while enabling dimension reduction directly on characteristics.