Multimodal information fusion for financial forecasting via cross-attention and calibrated uncertainty

Forecasting financial markets requires synthesizing heterogeneous information sources such as historical prices, company indicators, and unstructured news, whose interactions are nonlinear and regime dependent. We investigate a cross-attention transformer framework that fuses these modalities for probabilistic financial forecasting with calibrated uncertainty. We propose a framework that anchors fusion on Chronos-T5, a transformer pretrained on large-scale time series and used here as a frozen encoder for market dynamics. Parameter-efficient projection layers map company-level indicators and daily news embeddings into a shared representation space, while bidirectional cross-attention learns how to align and weight the different sources. We evaluate the approach on daily EUR/USD forecasting, with additional experiments across currency pairs and market regimes. The multimodal model consistently outperforms autoregressive and deep learning baselines in point prediction, as measured by mean squared error and Diebold–Mariano tests, and delivers sharper probabilistic forecasts according to the continuous ranked probability score (CRPS), weighted interval score (WIS), and empirical coverage. Raw predictive distributions tend to be overconfident, but a post-hoc split conformal recalibration restores nominal coverage and improves interval quality without retraining the backbone. From a soft computing perspective, the system combines approximate Bayesian inference via Monte Carlo dropout with distribution-free calibration, within a structured cross-modal fusion architecture that improves the reliability and interpretability of multimodal financial forecasts.