CBGAN: Predicting Stock Prices Using Combinatorial Generative Adversarial Networks

This paper proposes a combinatorial generative adversarial network (CBGAN) for stock market prediction. The generator integrates LSTM and GRU to capture long-term dependencies and improve efficiency, while the discriminator combines CNN and MLP for feature extraction and classification. The model uses MSE loss for single-step prediction and Wasserstein distance for multi-step prediction. Experimental results show that CBGAN-MSE excels in single-step prediction, while CBGAN-WGP offers stability in multi-step prediction. The model performs better in stable market conditions without unexpected disturbances.