This paper proposes a pattern-based stock trading system which utilize the results of supervised learning by neural networks for highly volatile stock price patterns. Three types of the highly volatile price patterns which include a recent trading day on which the price has rosen to the upper limit of the market limitation, are defined, and the semantics of those patterns are briefly analyzed with typical chart examples. Each pattern is learned by its own dedicated neural network, and trading signals are produced utilizing the prediction results of those neural networks. Through the experimental results on KOSPI and KOSDAQ, it is shown that the proposed pattern-based trading system achieves better trading performance than an existing all-in-one trading system based on a single neural network and a trading system based on moving average patterns.