We developed and evaluated empirical models to predict biological condition of wadeable streams across a large portion of the eastern United States, with the ultimate goal of forecasting for unsampled basins.  We used a RIVPACS-type predictive model to estimate taxonomic completeness (O/E) and categorize (i.e., altered vs. unaltered) the biological condition of 921 streams throughout the eastern United States.  We compared the accuracy with which two multiple logistic regression models and a random forest model predicted the category of biological condition for 100 independent subsets of validation data.  Predictor variables were limited to widely available geospatial data such as land cover, topography, climate, soils, and societal infrastructure.  The random forest model had the best overall performance, but misclassified biologically altered sites more often than unaltered sites (46% vs. 11 %, respectively).  We found, however, that classification accuracies for both condition categories could be simultaneously minimized at 20% by adjusting classification thresholds.  The most important predictors of biological condition were ecologically interpretable and corroborated existing literature.  We demonstrate a potential application of the model by forecasting biological condition in 552 unsampled basins across the Plains ecoregion of southeastern Wisconsin (USA).