Various hybrid approaches combining the above methods have been proposed in the literature. In the context of our research we have used the following two.
DT-ANN, consists, in its pure version, of first building a tree which identifies relevant attributes among candidate ones, then of translating it into a four layer MLP, further adapted via back-propagation to improve classification performances. It may or may not exploit the continuous information provided by security indices in the latter step. This allows one in practice to improve the reliability of decision trees while maintaining their advantages of simplicity and computational efficiency [23]. A variant of this method consists of using a fully connected three layer perceptron, with an a priori fixed number of hidden units.
DT-NN, consists of using in the distance computation of the
method only the attributes selected by a decision tree built
for the same problem. In our simulations, the attribute values are
pre-whitened
and the Euclidean distance is used, while the
appropriate number 
of neighbors is determined by trial and error.
With respect to a non-hybrid nearest neighbor approach, this method is
significantly faster and often more reliable
[24].