Neural network modeling has been successfully applied to making water level predictions for the coast of Texas. It has been hypothesized that the good performance of ANNs compared to other techniques such as harmonic analysis or linear regression is the result of the ability of ANN models to capture the non linear relationship between weather forcings and future water level changes. At present the ANN models are trained over at least one year of hourly data. This results in averaging out unusual events during the training phase. Selecting years for the training that have a higher proportion of extreme events such as strong frontal passages, tropical storms and hurricanes typically leads to best overall model performance. This study compares and discusses different methodologies to improve model performance during extreme events while preserving overall average performance. The implemented strategies include boosting the training cases leading to the largest errors (typically extreme events), modifying the cost function to increase the penalty due to large errors, and training the models on artificial data sets which include a larger proportion of extreme events than actual yearly data sets.