After the invention of Bitcoin and a peer to peer electronic cash system based on the blockchain, the market of cryptocurrencies increases rapidly and attracts substantial interest from investors and researchers. Cryptocurrencies price volatility prediction is a challenging task owing to the high stochasticity of the markets. Econometric, machine learning and deep learning models are investigated to tackle the stochastic financial prices fluctuation and to improve the prediction accuracy. Although the introduction of exogenous factors such as macro-financial indicators and blockchain information helps the model prediction more accurately, the noise and effects from markets and political conditions are difficult to interpret and modelling. Inspired by the evidence of strong correlations among cryptocurrencies examined in previous studies, we originally propose a Weighted Memory Channels Regression (WMCR) model to predict the daily close price of cryptocurrencies. The proposed model receives time series of several heavyweight cryptocurrencies price and learns the interdependencies of them by recalibrating the weights of each sequence wisely. Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) components are exploited to establish memory and extract spatial and temporal features. Moreover, regularization methods including kernel regularizers and bias regularizers and Dropout method are exploited to improve the generalization ability of the proposed model. A battery of experiments are conducted in this paper. The results present that the WMCR model achieves the state-of-art performance and outperforms other baseline models.