Abstract
Financial markets are adaptive systems in which asset prices emerge and change with respect to the continuous interactions of investors, institutions, and algorithmic traders, reflecting both historical trends and real-time information flows. Traditional time-series forecasting methods, such as ARIMA or autoregressive benchmarks (ARA), rely primarily on past price patterns and are limited in capturing these exogenous signals. In this study, we introduce a cross-domain, multi-modal framework that integrates numerical market data with textual sentiment embeddings using recurrent neural networks (RNNs) and transformer-based architectures. Evaluated over 7 years, the model outperformed the autoregressive benchmark by 44.31% in cumulative returns, reduced risk-associated volatility by 21.34%, and improved the Sharpe Ratio from 0.18 to 1.38. These results demonstrate that incorporating cross-domain signals enhances both directional accuracy and risk-adjusted performance. Overall, this work highlights the potential of multi-modal computational approaches to systematically leverage behavioral, informational, and market signals, providing a robust methodology for financial forecasting and practical investment decision-making.