The correlation-based network has become an effective tool to investigate the correlation of complex financial systems [1,2]. Different methods have been proposed to probe the complex correlation structures of financial systems including the threshold method, the minimum spanning tree(MST) [3], the planar maximumly filtered graph(PMFG) [4] and a strand of other methods [5–۱۱]. The common aim of all correlation-based networks is to seek for a sparse representation of the high dimensional correlation matrix of the complex financial system. Unlike other eigenvector-based methods (e.g., the principal component analysis) which decompose the variance of the system into a few dimensions, the correlation-based methods directly map the dense correlation matrix into sparse representation. The easy implementations and straightforward interpretations of those methods make them quite popular in complex system analysis, especially for complex financial systems. Recently, the correlation-based network has been used for portfolio selection in which some risk diversified portfolios are constructed based on a hybrid centrality measure of the MST and PMFG networks of the stock return time series [12]. It is well known that the financial system has its own temporal property which makes it extremely hard or even impossible to forecast. Thus if we want to construct our portfolio in a proper way, we have to consider the temporal attribute of the financial system. In this work, we study the correlation-based networks of stock markets by using the temporal network approach. Specifically we have analyzed the temporal evolution of three major stock markets of the world, namely, the US, the UK and China. Based on a centrality measure of temporal network, we also construct some portfolios that consistently perform the best under two portfolio optimization frameworks. Our work is the first research that incorporates the temporal network method into the study of complex financial system. The temporal evolution of the topological structures can be used to access the information of market instability. The effectiveness of the temporal centrality measure in portfolio selection depicts the importance of the temporal structure for the stock market analysis. The remainder of the paper is organized as follows: Section 2 gives the data description and the methodology we use through the paper. Section 3 presents the main results of the paper including the topology analysis of the stock markets and the applications to the portfolio optimization problems. Section 4 provides our conclusion.