Abstract
We consider the problem where an agent aims to combine the views and insightsof different experts' models. Specifically, each expert proposes a diffusionprocess over a finite time horizon. The agent then combines the experts' modelsby minimising the weighted Kullback--Leibler divergence to each of the experts'models. We show existence and uniqueness of the barycentre model and prove anexplicit representation of the Radon--Nikodym derivative relative to theaverage drift model. We further allow the agent to include their ownconstraints, resulting in an optimal model that can be seen as a distortion ofthe experts' barycentre model to incorporate the agent's constraints. Wepropose two deep learning algorithms to approximate the optimal drift of thecombined model, allowing for efficient simulations. The first algorithm aims atlearning the optimal drift by matching the change of measure, whereas thesecond algorithm leverages the notion of elicitability to directly estimate thevalue function. The paper concludes with an extended application to combineimplied volatility smile models that were estimated on different datasets.