Can Perturbations Help Reduce Investment Risks? Risk-Aware Stock Recommendation via Split Variational Adversarial Training

TL;DR

This paper tackles the challenge of risk-aware stock recommendation by addressing the inadequacy of profit-focused models to control downside risk. It introduces Split Variational Adversarial Training (SVAT), which splits adversarial training to be robust to perturbations of profitable stocks while being sensitive to risky ones, and couples this with a Variational Perturbation Generator (VPG) to model diverse latent risk factors. The approach yields a rough risk quantification through ranking entropy and demonstrates superior performance on six real-world datasets, including crisis periods, with notable improvements in risk-adjusted profits and volatility reduction. The work advances interpretable risk modeling in stock recommendation and provides a flexible framework that improves other baselines when integrated with SVAT components.

Abstract

In the stock market, a successful investment requires a good balance between profits and risks. Based on the learning to rank paradigm, stock recommendation has been widely studied in quantitative finance to recommend stocks with higher return ratios for investors. Despite the efforts to make profits, many existing recommendation approaches still have some limitations in risk control, which may lead to intolerable paper losses in practical stock investing. To effectively reduce risks, we draw inspiration from adversarial learning and propose a novel Split Variational Adversarial Training (SVAT) method for risk-aware stock recommendation. Essentially, SVAT encourages the stock model to be sensitive to adversarial perturbations of risky stock examples and enhances the model's risk awareness by learning from perturbations. To generate representative adversarial examples as risk indicators, we devise a variational perturbation generator to model diverse risk factors. Particularly, the variational architecture enables our method to provide a rough risk quantification for investors, showing an additional advantage of interpretability. Experiments on several real-world stock market datasets demonstrate the superiority of our SVAT method. By lowering the volatility of the stock recommendation model, SVAT effectively reduces investment risks and outperforms state-of-the-art baselines by more than 30% in terms of risk-adjusted profits. All the experimental data and source code are available at https://drive.google.com/drive/folders/14AdM7WENEvIp5x5bV3zV_i4Aev21C9g6?usp=sharing.

Figures (11)

• Figure 1: An example showing that for regression and classification methods, accurate stock prediction models LSTM (MSE $\downarrow$) and CNN (Acc. $\uparrow$) may earn less profit than inaccurate models ARIMA (MSE $\uparrow$) and MLP (Acc. $\downarrow$). The returns of the five stocks SQQQ, REGN, ACOR, ZBRA, and TREE are selected from the NASDAQ stock market at 02/23/2017.
• Figure 2: (a) Daily returns of two stock recommendation models. Both models achieve similar total profits ($35.7\%$ for Model 1 and $35.2\%$ for Model 2) but with different volatilities. (b) Illustration of the split adversarial training. $\delta$, $\delta'$ are perturbations and $\$$ denotes the stock return. Best viewed in color.
• Figure 3: (a) Workflow of the proposed SVAT framework. (b) Architecture of the variational perturbation generator. The perturbation $\bm{\delta}_i$ generated from $\mathbf{z}_i^{\text{post}}$ is engaged in adversarial training while $\bm{\delta}_i$ generated from $\mathbf{z}_i^{\text{prior}}$ is utilized to compute the ranking entropy in the testing environment.
• Figure 4: The curves of daily returns and cumulative investment return ratios of all models backtested on the three normal datasets. Best viewed in color.
• Figure 5: The curves of daily returns and cumulative investment return ratios of all models backtested on the three crisis datasets. Best viewed in color.