Abstract
Deep learning is a powerful tool whose applications in quantitative financeare growing every day. Yet, artificial neural networks behave as black boxesand this hinders validation and accountability processes. Being able tointerpret the inner functioning and the input-output relationship of thesenetworks has become key for the acceptance of such tools. In this paper wefocus on the calibration process of a stochastic volatility model, a subjectrecently tackled by deep learning algorithms. We analyze the Heston model inparticular, as this model's properties are well known, resulting in an idealbenchmark case. We investigate the capability of local strategies and globalstrategies coming from cooperative game theory to explain the trained neuralnetworks, and we find that global strategies such as Shapley values can beeffectively used in practice. Our analysis also highlights that Shapley valuesmay help choose the network architecture, as we find that fully-connectedneural networks perform better than convolutional neural networks in predictingand interpreting the Heston model prices to parameters relationship.