A class of incremental learning procedures known as the Modified Temporal Difference (MTD) method is introduced in this paper for fixed-step prediction problems which uses the functional features of Multilayer Perceptron. The method is applied for weekly prediction of the Sea Surface Temperature (SST) from oceanographic data. Temporal Difference (TD) methods suggest how each output of a temporal sequence must be changed, whereas a back-propagation algorithm decides which part(s) of a network to change in order to influence its output and reduce the overall error. In other words, TD methods and back-propagation address temporal credit and structural credit assignment issues, respectively. While the two methods address different sides of the same issues they are quite compatible and easily combined. A new scheme is formed by combining the advantage of back-propagation and TD methods catering to fixed-step problems and is named as the MTD method. The back-propagation algorithm is modified to propagate the temporal error. For prediction problems, the exponential recency has not been found suitable due to its large negative slope. In this paper a weighing scheme is introduced in which alterations are made to past predictions according to a newly proposed recency factor. The stochastic method, back-propagation algorithm, TD and MTD methods are applied to predict the SST values in the Arabian Sea, the Bay of Bengal and Central Indian Ocean and a comparative study is made. From the study it is observed that the proposed alternative recency factor in the MTD method leads to better prediction than the exponential recency.