Abstract: Artificial Neural Network (ANN) technique is applied to construct the non-linear mapping between output data from regional Eta model, used by CPTEC-INPE, and surface precipitation observed data over eastern Sao Paulo (Brazil) region. The goal is to generate site-specific quantitative forecasts for daily precipitation. The test was performed on five locations over São Paulo - Brazil for summer and winter periods (1997 - 2002). The paper shows the method of network structure construction, input/output data preparing and the sequence of the training. In this case a Feed - Forward neural network and learning algorithm Back-propagation were used. Model variables are the input data and precipitation data, from rain gauge, are output data. With the trained networks, the developed system gives the rainfall forecast for the next time-period (typically for next 6 hours). Additionally, a multiple linear regression was also employed, in order to compare with the ANN prediction. Statistical analysis results show the premise precipitation forecasts for the Sao Paulo region. The performance using ANN shows favourably with those generated using linear regression forecasts and numerical model precipitation prediction. This new approach indicates a potential for more accurate precipitation forecasting.