The abundance and distribution of most stream organisms are strongly influenced by water temperature. For bioassessments, the effects of ‘natural’ stream temperatures on the distribution of biota are usually predicted with surrogate variables (elevation, latitude), but a direct estimate of reference condition temperatures would be preferable. We used daily stream temperature data recorded at 455 reference-quality streams in the western USA to develop stream temperature predictive models. A simple empirical model based on latitude, longitude, elevation, and drainage area accounted for 78% of the variation in mean annual stream temperature (RMSE = 1.3 °C). Incorporation of easily obtained measures of long-term climate, catchment size and shape, and watershed soils significantly improved model performance (R² = 0.87, RMSE = 0.96 °C). However, 90% of the residuals still fell within a 3.3 °C band of values, which may be too large to detect many ecologically significant thermal alterations.  We are now incorporating reach-level estimates of canopy shading, solar radiation, and discharge to better characterize the physical processes known to influence local variation in stream temperatures.  These models should allow us to more precisely predict local thermal responses to climatic, landscape, and hydrologic alterations and subsequently the likely biotic responses to these alterations.