| Description: |
The building of prediction tools is considered to be one of the basic subjects in science([2]). Neural Networks (NN), with their distributed and parallel processing power, can be used as a tool to forecast stock exchange(SE) events, if these are seen as Time-Series(TS). In this paper we will present a system for SE events prediction, based on an energy function that we deduce from the Lyapunov (L- also called infinite) norm, and that we apply for our SE problem. We focus here on the mathematical deductions of the energy function and on the error minimization procedures. We present some comparative results of our method, the classical backpropagation method(BP), and the random walk generator. The L based energy function starts with the goal of minimization of the maximal error, also called maximal error sphere. We use this method, but while L is based on vector derivation, we use weight changing, therefore working with an extra degree of freedom. The final movement equations present a dead- zone behaviour, which is to be interpreted with a reaction (weight change) only for significantly large errors. The equations for the previous, hidden layers are deduced through BP of the new error function. The used NN is a two-layer feedforward net, the minimal design for the complexity of the problem. The data are some real-world data from [1], as well as some user-designed series for testing. The inputs of the system are the SE data and other economical influence factors. The data are pre-processed by a scaling procedure. We used the correlation coefficient 1 and the Theill(T) coefficient |