Financial forecasting and rules extraction from trained networks

This paper describes a forecasting approach using constrained networks. Two complementary approaches are proposed. The main property of the first approach is to lead to an efficient predictive algorithm based on backpropagation. Some units are constrained to hold the logical information of the network whereas the unconstrained unit keep the numerical information. Therefore the task of each unit is defined during the training. The second approach is focused on rules extraction. Using constrained networks, we are able to extract information from trained networks. This property is essential as it is possible to analysis, explain, extract and therefore control what happens inside trained networks. Simulation results for these approaches are reported