ABSTRACT

Can deep learning approaches achieve similar accuracy as statistical models in predicting 
the stock market? This research introduces an approach integrating Generative Adversarial 
Networks (GANs) with Long Short-Term Memory (LSTM) generators and Convolutional 
Neural Network (CNN) discriminators to predict stock trading actions. By analyzing and 
computing on historical Open, High, Low, Close and Volume (OHLCV) data of the stock 
market, we generate a wide range of input features using Fourier transforms, auto-encoders, 
and several technical indicators. Extreme Gradient Boosting (XGBoost) is employed to 
filter and extract important features. The extracted features are fed into the LSTM 
generator, whose output is then fed to the CNN discriminator to discriminate the buy, sell, 
or hold signals. Results demonstrate the model's effectiveness in capturing temporal and 
spatial patterns, with implications for refining stock trading strategies and enhancing 
financial decision-making for maximizing profits.