Paper: arXiv 2301.05693

Authors: Fateme Shahabi Nejad, Mohammad Mehdi Ebadzadeh

Abstract

Applying machine learning methods to forecast stock prices has been one of the research topics of interest in recent years. Almost few studies have been reported based on generative adversarial networks (GANs) in this area, but their results are promising. GANs are powerful generative models successfully applied in different areas but suffer from inherent challenges such as training instability and mode collapse. Also, a primary concern is capturing correlations in stock prices. Therefore, our challenges fall into two main categories: capturing correlations and inherent problems of GANs. In this paper, we have introduced a novel framework based on DRAGAN and feature matching for stock price forecasting, which improves training stability and alleviates mode collapse. We have employed windowing to acquire temporal correlations by the generator. Also, we have exploited conditioning on discriminator inputs to capture temporal correlations and correlations between prices and features. Experimental results on data from several stocks indicate that our proposed method outperformed long short-term memory (LSTM) as a baseline method, also basic GANs and WGAN-GP as two different variants of GANs.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper adapts DRAGAN and feature matching to address training instability and temporal correlations in stock price forecasting. It provides reasonable mathematical formulation for the loss functions and compares performance empirically against LSTM, standard GAN, and WGAN-GP baselines.

Research Flowchart

  flowchart TD
    A[Research Goal: Improve Stock Price Forecasting Stability & Capture Correlations] --> B[Data Inputs: Historical Stock Prices & Technical Features]
    B --> C[Preprocessing: Windowing & Conditioning for Temporal Correlations]
    C --> D[Model Training: Proposed DRAGAN + Feature Matching Generator/Discriminator]
    D --> E[Performance Evaluation vs. Baseline LSTM, Basic GAN, & WGAN-GP]
    E --> F[Key Outcome: Superior Forecasting Accuracy, Stable Training & Reduced Mode Collapse]