Empirical Analysis of Neural Architectures and Side Information in Financial Time Series Forecasting

Abstract

This thesis investigates the predictive capabilities of neural networks in financial time series forecasting, focusing on predicting the weekly close price of the SPY index. We explore the integration of options-derived features alongside traditional price data, compare recurrent architectures and transformer-based models, and evaluate multiple training strategies. Our key contributions include: (1) evidence that options-derived input features improve both error metrics and directional accuracy; (2) a comparison study of four training methods (one-step-ahead, direct multi-step, simulation error, and teacher-forcing); (3) the development of a bidirectional GRU-LSTM hybrid model that outperforms standard recurrent networks in multi-step forecasting; and (4) a novel coarse tokenization approach for discretizing continuous financial data, which improves first-week prediction performance when used in transformer models that use an asymmetric attention mechanism. Overall, this thesis illustrates the importance of input design, model architecture, and training methodology in neural financial forecasting. We conclude by outlining directions for future work, including cross-asset generalization and further exploration of tokenization schemes for transformer-based models.