Introduction
The purpose of this study is to propose a recurrent neural network model for stock price pattern recognition,
and to develop a new method for evaluating the network. In stock trading with technical analysis[l],
price patterns in Japanese-style stock charts[2], such as triangles, indicate an important clue to the+trend of
future changes in stock price. An expert analyzes the charts to detect these patterns on the basis of his past
experience. In practice, a few experts have continued to watch the charts, and they are engaged in stock
trading for limited names of corporations. However, it takes a long time to become a proficient expert, and
the expert’s capability life span is short. Therefore, as the number of traded names increased, computer-aid
to the chart analysis has been strongly expected.
For recognizing specified patterns from a time sequence of stock prices, it is indispensable to develop a
normalization method for eliminating the bias due to differences in time spans and names, and to investigate
an algorithm for detecting the patterns. There is no successful rule-based approach to the stock price pattern
recognition, because such recognition is based on the expert’s subjectivities. For example, the triangle pattern
has non-linear time-elasticity and definite oscillations. Therefore, it is difficult to recognize the patterns by
means of existing statistical models and AI techniques.
In this work, a recurrent neural network model was applied to triangle recognition. Consequently, test
triangles were appropriately recognized. Furthermore, a new method for examining recurrent networks was
established by searching for temporal context transition in the model. It became clear that the model was
effective in partial elimination of mismatching patterns.