<p>Stock investment is one of the most important aspects of financial markets, and stock selection is a challenging field when making investment decisions. In the Chinese stock markets, there is a limit up / down system, but it is not clear that whether the limit up / down data is useful. In this study, we propose two deep autoencoder (DA) architectures to learn efficient representations of technical indexes, introduce limit up features, and combine the two DAs, a genetic algorithm (GA) and support vector regression (SVR) to develop an effective short-term stock selection model with technical indexes. SVR is utilized to forecast the returns of stocks, and all stocks are ranked by the forecast returns. The top-ranked stocks are selected to construct a portfolio. The parameters of SVR are optimized by the GA, while the DAs are used to learn the efficient codes of all the technical features. The proposed model is evaluated with stocks from the Shanghai Main-Board Market (SHMM), the Shenzhen Main-Board Market (SZMM), and the Second-Board Market (also called the Growth Enterprise Market, GEM) separately. The empirical results show that the limit up features make a positive contribution to short-term trading. The GEM is the worst market for technical analysis among the three markets. We also compared the DAs with other commonly used feature selection methods, and the comparisons show the DAs are the most effective and potential methods to learn the representations of technical indexes for the short-term stock selection. Further, the top five stocks portfolio obtained by our model can achieve a high return in all three markets, which is friendly to individual investors.</p>