Explainable Artificial Intelligence methods for financial time series
Paolo Giudici, Alessandro Piergallini, Maria Cristina Recchioni, Emanuela Raffinetti
Abstract
We consider the problem of developing explainable Artificial Intelligence methods to interpret the results of Artificial Intelligence models for time series data, taking time dependency into account. To this end, we extend the Shapley–Lorenz method, normalised by construction, to Artificial Intelligence for time series, such as neural networks and recurrent neural networks. We illustrate the application of our proposal to a time series of Bitcoin prices, which acts as the response variable, along with time series of classical financial prices, which act as explanatory variables. Three main findings emerge from the analysis. First, recurrent neural networks lead to a better performance, in terms of accuracy and robustness, with respect to classic neural networks. Second, the best performing models indicate that Bitcoin prices are affected mostly by their lagged values, and that their explainability, in terms of classical financial assets, is limited. Third, although limited, the contribution of classical assets to Bitcoin price prediction is well captured by recurrent neural networks.