Robust Statistical Inference in Financial Time Series Using Bayesian-Regularized Sequential Deep Learning

Long Short-Term Memory (LSTM) networks constitute one of the most widely employed architectures in modern deep learning, particularly for sequential data modeling. Their application to time series forecasting poses unique challenges, as the underlying data-generating processes often exhibit long-term trends, seasonal and cyclical components, and stochastic noise, all of which complicate accurate prediction. The predictive performance of LSTM models is highly sensitive to the selection of hyperparameters, such as learning rate, hidden state dimensionality, sequence length, and regularization strength, each exerting non-linear effects on model convergence and generalization. Despite their prominence, rigorous methodological guidelines for LSTM hyperparameter optimization remain underdeveloped in the literature. This paper seeks to bridge this gap by systematically analyzing the interplay between hyperparameter configurations and forecasting accuracy, thereby contributing to the formalization of best practices in the statistical learning of temporal dynamics.