Technical analysis : neural network based pattern recognition of technical trading indicators, statistical evaluation of their predictive value and a historical overview of the field

Abstract

We revisit the kernel regression based pattern recognition algorithm designed by Lo, Mamaysky, and Wang (2000) to extract nonlinear patterns from the noisy price data, and develop an analogous neural network based one. We argue that, given the natural flexibility of neural network models and the extent of parallel processing that they allow, our algorithm is a step forward in the automation of technical analysis. More importantly, following the approach proposed by Lo, Mamaysky, and Wang, we apply our neural network based model to examine empirically the ability of the patterns under consideration to add value to the investment process. We discover overwhelming support for the validity of these indicators, just like Lo, Mamaysky, and Wang do. Moreover, this basic conclusion appears to remain valid across different levels of smoothing and insensitive to the nuances of pattern definitions present in the technical analysis literature. This confirms that Lo, Mamaysky, and Wang's results are not an artifact of their kernel regression model, and suggests that the kinds of nonlinearities that technical indicators are designed to capture constitute some underlying properties of the financial time series itself. Finally, we complement our empirical analysis with a historical one, focusing on the origins of trading and speculation in general, and technical analysis in particular.