Intraday stock price modeling is a challenging task due to the noisy and highly volatile nature of short-term stock markets variations. The new advances in the application of generative adversarial networks (GAN) in many areas, especially in the financial sector, allow researchers to develop tools capable of making more accurate predictions. This article proposes a framework for improving intraday stock forecasting by using synthetic examples to train a prediction model. The framework relies on conditional Wasserstein GAN with gradient penalty and on a mode-normalization procedure to generate highly realistic data which are fed, alongside the real ones, to an LSTM to predict future stock variations. The usefulness of the proposed framework is assessed on real stocks data using quantitative and qualitative criteria. Our experimental results show a significant improvement in forecasting accuracy.