The impact of different observation suites on the forecast skill of the NASA Seasonal-to-Interannual Prediction Project (NSIPP) coupled model is investigated by a series of forecast experiments initialized with the NSIPP ocean data assimilation system. An optimal interpolation scheme is used to assimilate in situ temperature data from 1993 to 2002, including TAO array data. The impact of TAO array data on forecast skill is investigated by data withholding, i.e., only part of the TAO data are used in certain experiments. The skill in forecasting Nino-3 and upper ocean heat content is assessed. Different assimilation strategies are employed to compare the effectiveness of the assimilation. We will consider issues such as salinity adjustments and geostrophic current correction so as to achieve balanced initialization.