LSTM has been proven to help understand multivariate time series since it can implement automatic feature learning and handle input sequences harnessed in an end-to-end model. Inspired by its success, this study proposed a new approach, named Multiple LSTMs, to settle the challenge of integrating multivariate LSTM and various characteristics of multiple time series to predict household leverage of China. The new method consists of multiple LSTM layers, where each LSTM is used to learn representations of multivariate time series automatically of various regions. To illustrate the performance, the new method is used to forecast household leverage in both the national level and regional levels based on multivariate time series which have 6 independent variables. The results show that the new approach provides more accurate predictions of household leverage than alternative methods. Overall, making use of multiple multivariate time series can gain a better result of the prediction of household leverage.