<p>Accurately forecasting multivariate financial time series remains a significant challenge due to volatility, nonlinear dependencies, and the difficulty of capturing both temporal coherence and predictive uncertainty. This study introduces LUX-GAN, a generative adversarial framework designed for multi-step forecasting of stock prices by integrating stochastic latent modeling, entropy-aware attention, and latent recalibration. Unlike conventional models that emphasize point predictions, the proposed architecture produces both accurate forecasts and calibrated uncertainty estimates, enabling risk-sensitive decision-making. The methodology employs a Latent Uncertainty Expander encoder to generate stochastic representations, an Entropy-Sensitive Delay Gate to allocate horizon-specific attention, and a recalibration mechanism that corrects latent shifts during training. Experiments on Indonesian stock data demonstrate that LUX-GAN outperforms a VAE-GAN baseline, achieving higher R² scores, lower RMSE, and well-calibrated predictive intervals. These findings highlight the framework’s novelty in unifying generative modeling with uncertainty quantification, offering a practical pathway toward more reliable and interpretable financial forecasting systems.</p>