Automate the identification of technical patterns: a K-nearest-neighbour model approach

To explore profitable patterns in historical stock prices, ordinarily a promising pattern is selected by rule of thumb, and then its performance is verified by comparing the conditional return of the pattern with the unconditional benchmark return. We adopt an alternative philosophy: without any pre-selected pattern, the proposed method explores the entire graphic space to automate the identification of technical patterns. Derived from the K-nearest-neighbour (KNN) forecast, our method calculates the graphic similarity of patterns by the distance of the price vectors, and then classifies patterns according to the graphic similarity, and finally identifies patterns which contain predictive powers of the future market movement. KNN provides an excellent tool for probing the entire graphic space formed by price patterns to obtain an overall perspective of the effectiveness of technical patterns. Not only the well-known patterns but also the unnoticed and potentially informative patterns can be probed. To evaluate the performance, our method is compared with classic KNN forecast and technical trading rules. Results indicate that the stock market is relatively efficient and technical analysis is still effective to explore excess returns.