Neural network application to perception of local risk

Abstract

If the markets are efficient, no other portfolio should consistently outperform is theperformance of according to Markowitz model optimized portfolios. The startingpoint of this research is the observation that some traders consistently make superiorprofits. Knowing that insiders and manipulators are rare cases, we draw our attentionto what dealers name as a 'feeling'. The expression that they 'feel' is true becausethey do not use explicit deterministic rules for calculation of expectations aboutrisk&return. Human brain implicitly learns concepts and rules. Our aim in this paperis to isolate one such 'feel' or method to show their existence and usefulness inportfolio management techniques. As a target for our research, we selected theconcept of dynamic local risk perception. One of the rriathematical structures thatlearn through experience withom specifying explicit rules, is the neural network(NN) structure. From various possible NN types we used back-propagation NN tomeasure and document this 'feeling', because they are the most widely usednetworks and the theory of them are the most complete one. We worked on<;ukurova and Eregli stocks and attained reasonable correlations for predicting riskyand normal days. This 'feeling' is one of the most simplistic concepts that tradersuse, which can't be explicitly defined but can be mimicked by NN. We believe thatNN's can be used to explain many such rules that can't be explicitly defined but areof great advantage to portfolio managers.