The stock plate division can help investors accurately understand the overall characteristics of stocks and determine the scope of investment. Most of the stock clustering algorithms only extract common indicators of stock investment which cannot describe the stock data well. So in this paper, we extract numerical statistical features such as mean, variance and low-dimensional manifold representation. Moreover, those clustering algorithms ignore the fact that different importance features would make different contribution to the clustering, and rarely evaluate the extracted features, those features with less information or high correlation with each other may result in poor generalization ability, low accuracy of the trained model, and poor visibility of clustering effects. So we use entropy weight method to delete features with less information, delete features that are highly correlated with each other, all obtained features and the corresponding entropy weights are multiplied to make the different importance features have a different contribution to the stock sector division. Finally, in the aspect of clustering algorithm, we use the improved Fuzzy C-means clustering (FCM) algorithm which can effectively avoid falling into local optimal solution, thereby improving the clustering accuracy. The experiment shows that the proposed algorithm outperforms the traditional algorithm in both clustering results and the objective indicators.