Deep Learning for Time Series Forecasting: An LSTM (Long Short Term Memory) Architecture Applied to Equity Portfolios

Machine learning models and artificial intelligence techniques have increasingly been applied to financial forecasting problems involving large volumes of historical market data. This paper investigates whether Long Short-Term Memory (LSTM) neural networks can learn useful patterns from historical equity price sequences to predict short-term directional movements. Stock prices are influenced by multiple interacting variables and often exhibit highly nonlinear and noisy dynamics, making traditional linear models such as CAPM or Fama-French factor models insufficient for capturing short-term temporal structure.

This study constructs a dataset consisting of rolling windows of historical stock prices and formulates the forecasting problem as a supervised classification task. Each input sample contains 21 consecutive trading days of closing prices, while the target label represents the average price movement over the following week. Two labeling schemes are evaluated: a coarse three-class directional system and a more granular five-class system. A multi-layer LSTM architecture is trained on these samples, and the effect of hyperparameter tuning is analyzed through controlled experiments.

Results show that LSTM models consistently outperform random classification baselines. Hyperparameter optimization improves predictive accuracy in both labeling configurations, while coarser classification schemes achieve higher overall performance than fine-grained classifications. Although predictive accuracy remains moderate due to the inherent noise of financial markets, the results demonstrate that sequence-based neural networks can capture statistically meaningful patterns in historical price data. These findings suggest that properly tuned LSTM architectures provide a promising foundation for short-term financial forecasting and future portfolio construction systems.