Asymmetric prediction problems widely exist in the financial industry: buying an asset when its value decreases creates a real loss, while taking no action when its value increases results in a loss of opportunity. Therefore, a well-designed trading algorithm could potentially benefit from such asymmetry. In this paper, we make an attempt to construct such an algorithm, built upon the Neyman-Pearson classification paradigm, under which type II error is minimized while type I error is bounded under some arbitrary level alpha. One existing algorithm is the NP umbrella algorithm, which assumes independent data. However, it could be suboptimal in trading scenarios because of the assumption of independence. Hence, we introduce a modified version of the NP umbrella algorithm by using moving block bootstrap to address the problem. We then apply this new algorithm to real trading scenarios in the Chinese stock market, an example of the above asymmetry from its restriction on shorts. Analysis of stock data from the Chinese market shows the advantage of our proposed algorithm over the classical classification paradigm and the traditional regression approach.