District level rice yield forecast models were developed for 13 districts of the Brahmaputra valley of Assam using yield data and weekly weather data during 1990-2012 at vegetative (F1) and mid-season (F2) stages of rice crop through modified Hendrick and Scholl technique. The models were validated using independent data set of three years (2013-15). Stepwise regression technique was used for fitting the model and decided best by highest R2 and lowest percent error. The coefficient of determination (R2) ranged from 0.32 (Jorhat) to 0.88 (Kamrup) in F1 stage and 0.29 (Sonitpur) to 0.92 (Kamrup) in F2 forecast. In general, F2 forecast models were found comparatively better in forecasting rice yields than F1 models. Inclusion of BSSH along with temperature (maximum and minimum), rainfall and relative humidity (morning and afternoon) increased the accuracy of the yield forecast models, compared to the model developed without BSSH. Maximum temperature and relative humidity were the major weather parameters in determining rice yields in most of the districts located in central and upper part of the valley. On the other hand, rainfall in combination with maximum temperature and relative humidity were found relatively more important in the districts located in the lower part of the Brahmaputra valley.