We present a Bayesian approach to ensemble learning for time series forecasting, designed to address the challenge of model adaptation in non-stationary environments. Classical ensemble methods typically assume fixed model parameters and static combination rules, which limits their adaptability under temporal drift. Our method generalizes ensemble construction by framing both submodel weight assignment and parameter adaptation as problems of probabilistic inference. Using Bayes’ theorem, we derive an online update rule that balances new observations with prior knowledge, enabling coherent uncertainty quantification. Submodel errors are transformed into ensemble weights via a parametrized function, with parameters governed by Gaussian priors and updated using a maximum a posteriori (MAP) approximation. This allows for fast, gradient-based updates without discarding accumulated experience. We evaluate the approach on a short-horizon directional forecasting task across 40 foreign exchange time series using eight neural network-based submodels. The proposed method consistently outperforms static ensembles, with small but statistically significant improvements in accuracy and increase in profits in simulated setting. Beyond empirical findings, this work contributes a tractable and theoretically grounded framework for adaptive ensemble learning with Bayesian guarantees.