Application of sensitivity analysis to neural network determination of financial variable relationships
A new algorithm for stratified neural network training is described which uses the ``adaptive learning rates technique. Let the sign of the mean square error derivative with respect to the improved node weight at the epoch \(m\) be \(d_m\). The learning rate at epoch \(m\) depends on the previous \(d_k\) (\(k<m\)) by exponential averaging: \(D_{m+1}=\alpha D_m +(1-\alpha)d_m\) and the weight \(e\) is changed as \(e_m=e_{m-1}+\beta\) if \(d_mD_m>0\) and \(e_m=e_{m-1}\gamma\) otherwise. (Here \(\alpha\), \(\beta\) and \(\gamma\) are constants). This algorithm is used to train a network to predict the bond rating of major companies of the USA by some financial variables. Since these variables are highly correlated the principal components are used as input data for the network.
This page was built for publication: Application of sensitivity analysis to neural network determination of financial variable relationships
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4231250)