Financial Thought Experiment: A GAN-based Approach to Vast Robust Portfolio Selection

Chi Seng Pun, Lei Wang, Hoi Ying WongIJCAI 2020 · DOI

TL;DR

This paper introduces GANr, a novel Generative Adversarial Network (GAN) architecture designed for robust portfolio selection. By integrating a regressor into the GAN framework, GANr enables a "thought experiment" approach to finance, generating diverse market scenarios to build portfolios that are resilient to bleak financial futures, outperforming traditional methods in empirical studies.

What Problem Does This Paper Address?

Modern financial trading often involves investors mentally simulating various future market scenarios to construct portfolios that perform well under uncertainty. This process is akin to a "thought experiment." However, traditional portfolio selection methods, while sophisticated, often struggle with the vast number of assets (vast portfolio selection) and may not adequately capture the "ambiguity" of future market conditions. Existing robust portfolio selection methods are often model-based and cannot handle an arbitrarily large number of risky assets. The authors aim to bridge this gap by using GANs, a powerful tool for generating realistic synthetic data, to computationally replicate this financial thought experiment and develop a robust portfolio selection strategy for a large universe of stocks.

Methods

The core innovation is the GAN with Regressor (GANr) architecture. It builds upon the Conditional Least Squares GAN (CLSGAN) by adding a third neural network: a regressor (R).

  • Generator (G) & Discriminator (D): These components function similarly to CLSGAN, with the generator creating synthetic market data (fake scenarios) and the discriminator trying to distinguish between real and fake data. This ensures the generated scenarios are plausible.
  • Regressor (R): This network is trained to perform a specific regression task relevant to portfolio selection (e.g., minimizing portfolio risk). It takes both real and generated data as input.
  • Integration: The key novelty is how these components interact. The regressor's loss is incorporated into the generator's objective function, scaled by a parameter µ. This means the generator is not only trying to fool the discriminator but also trying to generate scenarios that make the regressor's task (and thus the portfolio's performance) as difficult as possible. This adversarial training between the generator and regressor is what drives the robustness.
  • Portfolio Selection Application: For vast portfolio selection, the problem is reframed as a regression task: predicting a target variable (e.g., stock return) based on other stock returns, subject to gross-exposure constraints. GANr is then applied to this reframed problem.

Datasets

The paper uses empirical datasets of S&P 500 index components for two distinct periods:

  • Recession period: December 1, 2006, to June 30, 2009.
  • Boom period: December 31, 2015, to December 29, 2017.

For training, the daily returns from the recent past one year are used. The authors note that this is a small sample size, which makes robust methods like GANr particularly valuable.

Key Insights and Ideas

  • GANs as Financial Thought Experiments: The paper frames GANs not just as data generators but as tools to computationally simulate diverse future market scenarios, enabling a "thought experiment" for portfolio construction.
  • Adversarial Robustness: The core idea is that by having the generator create scenarios that are challenging for the regressor (which aims to optimize portfolio performance), the resulting portfolio becomes inherently more robust to adverse market conditions.
  • Controlling Ambiguity Aversion: The parameter µ in GANr allows investors to control their "ambiguity aversion." A higher µ leads to more severe generated scenarios, reflecting a greater aversion to uncertainty.
  • Data-Driven Ambiguity: Unlike traditional robust control methods that rely on predefined models for ambiguity, GANr creates a "data-driven ambiguity set" by learning from historical data.

Key Results

  • Outperformance in Recessions: GANr portfolios (especially with µ = 0.2) demonstrated significantly better performance than LASSO and CLSGAN portfolios during the recessionary period ('06-'09), showing greater resilience to bleak market conditions.
  • Robustness Across Quantiles: QQ plots indicate that GANr portfolios consistently outperform LASSO at lower quantiles of daily returns, meaning they suffer less severe losses during downturns.
  • Comparable Performance in Booms: During the boom period ('16-'17), GANr portfolios performed comparably to CLSGAN and LASSO, with GANr showing lower annualized risk.
  • Improved Risk Metrics: GANr portfolios generally achieved lower annualized risk (volatility) compared to LASSO, and in the boom period, GANr (µ=0.2) achieved the lowest risk.

Limitations and Future Work

  • Mode Collapse: The authors acknowledge that GANs, including GANr, can still be susceptible to mode collapse, where the generator produces limited variety of outputs.
  • Parameter Tuning: The performance of GANr is influenced by the µ parameter, which requires careful selection based on market conditions (e.g., higher µ during recessions).
  • Future Work: The authors suggest exploring GANr with other GAN architectures (e.g., Wasserstein GAN), generalizing the regressor's loss function to other utility functions, and investigating the mathematical properties of GANr, such as its relation to Pearson χ² divergence.

Summary updated

BibTeX
@inproceedings{pun2020financial,
  author = {Chi Seng Pun and Lei Wang and Hoi Ying Wong},
  title = {{Financial Thought Experiment: A GAN-based Approach to Vast Robust Portfolio Selection}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/637},
}
Figure