Three new hybrid neural models which are based upon the basic neural model put forth by McCulloch and Pitts (Haykin, 1999) and the compensatory neural models by Sinha et al. (2000), (2001) are proposed in this paper. The basic neural and the compensatory neural models are modified to take into account any linear dependence of the outputs on the inputs. This makes the hybrid models suitable for the solution of some complex problems such as chaotic nonlinear time-series and more simple problems such as linear time-series. These models are verified using a simulation example. It is shown that the hybrid neural models are superior to the basic neural model and the compensatory neural models for time-series forecasting problems.