Sensitivity analysis of asset returns to various economic variables provides investors with a useful tool to build portfolios and manage their risk. However, there are strong reasons to believe that stock exposures evolve through time and that factor models involving them are only pertinent if they use reliable estimates of future sensitivities. Both Kalman filtering and neural networks may be used to provide such estimates. While the Kalman filter is good at modelling the time structure of sensitivities, neural networks are capable of relating them to exogeneous variables in a non linear way. Furthermore, because the two approaches perform complementary tasks of sensitivity forecasting, they may be combined to achieve better performances. These procedures are evaluated in a controlled simulation experiment and in a real stock exposure analysis. Stock sensitivities to interest and exchange rates are forecasted for 90 French shares and portfolios are built accordingly.

At temperatures encountered in thermal oxidation, silicon dioxide flows viscoelastically, A reduced-dimension, generalized boundary element method for modeling such a problem has been developed. A modified Kelvin's solution is used as the kernel; constant-velocity loading is chosen to operate with a wide range of stress relaxation times. Simulation results are reported for simple cases of focal oxidation.>