We consider the problem of developing safe Artificial Intelligence methods for high-frequency time series, typical of financial data. The goal is to extend Accuracy metrics and Shapley values, making them applicable to various machine and deep learning models in time series analysis. The study compares AI models, such as autoregressive networks and recurrent neural networks (LSTM and GRU), with other machine learning models. The approach is applied to Bitcoin price data, using traditional financial prices as explanatory variables. The analysis finds that recurrent neural networks outperform traditional neural networks and other machine learning models. Although Bitcoin prices are mainly influenced by their past values, differently from traditional financial assets, recurrent neural networks are able to capture the role of these assets in predicting Bitcoin prices, and to explain their role, which is shown to vary across different prediction horizons.