STOCK MARKET PREDICTION USING GENERATIVE ADVERSARIAL NETWORK (GAN) – STUDY CASE GERMANY STOCK MARKET

Abstract

Using neural networks in economic time series data is a relatively new and unexplored field. A lot of companies and economic research mostly use logistic regression or statistical approaches when trying to predict stock market movements. Neural networks have become a frequent tool for prediction in recent years, and this approach has been confirmed to provide more reliable and better solutions in terms of predictive accuracy. Within the wider context of the current debate on the use of neural networks in stock market prediction, we suggest an innovative methodology based on a combination of neural networks. In our analysis, we use a Wasserstein Generative Adversarial Network (WGAN) on the German stock market as an example. We present how a trading strategy could be established based on the model’s predictions and how it can be compared with other models in terms of returns. Overall, the WGAN monthly prediction outperformed Random Forest by 36%, the benchmark by 32%, and LSTM by 26% in the testing period. Our results also suggest that the WGAN model has, on average, higher returns than a pure investment in the index. Furthermore, WGAN is less volatile, which is always the preferred option for investors. Using neural networks for stock index prediction and confirming that the WGAN investment strategy brings higher returns compared to commonly used models is the main contribution of this paper to the current debate.

Data:
Received: 23 Oct 2024
Revised: 30 Oct 2024
Accepted: 1 Nov 2024
Published: 15 Nov 2024