Subsequence dynamic time warping for charting: bullish and bearish class predictions for NYSE stocks

Advanced pattern recognition algorithms have been historically designed in order to mitigate the problem of subjectivity that characterises technical analysis (also known as ‘charting’). However, although such methods allow to approach technical analysis scientifically, they mainly focus on automating the identification of specific technical patterns. In this paper, we approach the assessment of charting from a more generic point of view, by proposing an algorithmic approach using mainly the dynamic time warping (DTW) algorithm and two of its modifications; subsequence DTW and derivative DTW. Our method captures common characteristics of the entire family of technical patterns and is free of technical descriptions and/or guidelines for the identification of specific technical patterns. The algorithm assigns bullish and bearish classes to a set of query patterns by looking the price behaviour that follows the realisation of similar, in terms of price and volume, historical subsequences to these queries. A large number of stocks listed on NYSE from 2006 to 2015 is considered to statistically evaluate the ability of the algorithm to predict classes and resulting maximum potential profits within a test period that spans from 2010 to 2015. We find statistically significant bearish class predictions that generate on average significant maximum potential profits. However, bullish performance measures are not significant.