High fluctuation and self-similarity are typical characteristics of financial time series. Furthermore, affected by market environment, such as regular announcement of important economic data, time/date-sensitive fluctuations commonly exist in financial time series. However, the existing learning models were usually lack the consideration of essential characteristics of financial data, where both the fusion learning of multiple temporal features and the necessary attention to time-sensitive fluctuations were ignored. Inspired by this, to represent the temporal characteristics of self-similarity and reveal intrinsic feature details, in this article, time series and its features are converted into visibility graphs using the technique of Gramian Angular Fields, based on which convolutional long short-term memory (ConvLSTM) is applied to implement multifeature fusion learning. Moreover, to capture the time/date-sensitive fluctuation existing in financial time series, a subspace decomposition composed of the fuzzy control mechanism is first introduced into the ConvLSTM model, which considerably improves the prediction performance. On the basis of the proposed learning model, a concise intelligent trading strategy is designed. By using real foreign exchange data, various experiments are implemented to show the effectiveness of the proposed model.