Towards an artificial technical analysis of financial markets

Technical analysis is generally considered a qualitative approach to financial trading: it relies on operators´ specific skill to capture significant patterns over a series of quotes, thus anticipating market movements so as to beat the market. The technique is based on a mixture of human sentiments and political/economical variables conditioning, rather than solid scientific foundations, in contrast to the efficient market hypothesis (EMH), even in Simon weak form. However, the idea to get time series relevant features and use them to forecast future behaviour of markets is remarkable. I use a particular class of neural algorithms, topology representing networks (TRN), to introduce an artificial technical analysis, based upon capabilities of neural nets to catch out essential features of financial time series, simply by considering them as data and nothing else. A competitive unsupervised algorithm is used. Both a single net and sets of nets are taken into account; all movements are assumed to come exclusively as a result of neural pattern recognition, skipping any other possible consideration of technical, fundamental or political indicators. This choice has been supported by previous results, which have proved the ability of such nets in recognising low deterministic chaos in financial time series. Data from Dow Jones Industrial Average since 1915 are employed: in order to test the robustness of this approach, I consider results over various epochs, each of them differing in that training and control set are randomly sorted in turn