|
Abstract
|
Most of the existing approaches fail to generate adequate training data for rare market events, adequately model temporal dependencies, or integrate multimodal data in a robust and explainable manner, thus limiting their efficacy in predictive and decisionmaking tasks. This work addresses the above challenges by using an integrated pipeline based on state-of-the-art approaches for synthetic data generation, feature engineering, deep learning, reinforcement learning, and explainability. The generation of synthetic time-series data that retains temporal dependencies is facilitated by the use of TimeGAN, thus enabling the robust training of models for rare and unseen market scenarios. The TFT dynamically captures both short- and long-term dependencies and encodes multimodal covariates in the improvement of feature representations. Multimodal BERT, or MM-BERT, fuses textual, numerical, and sentiment data, bringing out complex inter-modal relationships essential for downstream tasks. PPO with adaptive reward scaling is used for reinforcement learning with stabilizing trading policy training and optimizing risk-adjusted returns. Finally, SHapley Additive exPlanations (SHAP) enhances model transparency by attributing feature importance to trading decisions. This achieves the improvements across key metrics such as >85% similarity of synthetic data with real-world patterns, ~20% RMSE reduction in price prediction, and a +25% increase in Sharpe Ratio. This would be transformative in the space of financial modeling because of robust predictions, actionable trading signals, and explainable decision-making to overcome significant gaps in current methodologies.
|