Dynamic Uncertainty-Weighted Fusion for Multi-Source Financial Data: A Bayesian Framework for Time-Varying Source Reliability

Abstract

The integration of diverse data sources—from quantitative market metrics to qualitative news sentiment—is pivotal for modern financial forecasting. However, prevailing data fusion techniques are critically hampered by their reliance on static weighting schemes, which operate under the untenable assumption of constant source reliability. This paper confronts this fundamental misalignment with the non-stationary nature of financial markets, where the predictive utility of any single source, such as news sentiment, is inherently dynamic and subject to regime shifts. We introduce a novel Bayesian framework for Dynamic Uncertainty-Weighted Fusion, designed to explicitly model and adapt to the time-varying reliability of multi-source data. The core of our methodology is a probabilistic mechanism that continuously estimates the time-dependent uncertainty (σ_i, t) of each source, inferred from its recent predictive consistency. These evolving uncertainty measures are then transformed via a softmax function to generate a set of dynamic, probability-weighted coefficients. This architecture enables the fusion model to autonomously prioritize the most credible information streams at any given time, effectively discounting sources that become noisy or unreliable during periods of market stress or transition. By formalizing a principled, uncertainty-driven approach to adaptive information integration, this framework establishes a new paradigm for building robust, self-calibrating financial decision-making systems.