LoopGAN: A Novel Multi-Step Generative Architecture for Sequential Stock Forecasting

This paper introduces “LoopGAN,” an advanced generative model architecture designed for sequential stock forecasting, leveraging the strengths of Conditional GAN (CGAN) and Least Squares GAN (LSGAN) to enhance prediction accuracy in highly volatile financial time series. LoopGAN's unique recursive prediction mechanism allows each output to feed back as input, creating dynamic, extended forecasts across multiple time steps. Extensive evaluation of the Dow J ones Industrial Average (DJIA) dataset, highlights LoopGAN's superior performance, achieving a Mean Absolute Percentage Error (MAP E) of 0.005739. This result surpasses traditional models, with LoopGAN outperforming LSTM and RNN models, which registered MAPE scores of 0.006803 and 0.008023, respectively. These findings underscore LoopGAN's robustness, offering a marked improvement in prediction accuracy and confirming its reliability for complex financial forecasting tasks.