Retrospective testing of streamflow forecast methods provides a basis for estimating the frequency and magnitude of forecast errors, and the risks associated with alternative methods of using (or not using) forecasts for water management. We report on a performance assessment of long-lead streamflow forecasts based on a resampling method that relies on categorical classification of winter climate for the Columbia River basin. Retrospective forecasts of runoff at The Dalles, OR are produced on September 30 preceding each water year, and the forecast period runs from October 1 to September 30 (one water year). The forecasting method uses a macroscale hydrology model driven by gridded observations up to the time of forecast, and forcings (precipitation and temperature) resampled from historical observations during the forecast period according to classification of the forecasted ENSO/PDO state. For purposes of retrospectively determining the winter climate state during the forecast period, a perfect forecast of categorical winter ENSO state (warm, neutral, cool) and a heuristic estimate of the PDO (warm, cool) based primarily on persistence is assumed to be available on September 30 of each year. Initial soil moisture is selected based upon model simulations up to the forecast date. Meteorological driving data are resampled from observed records for 1924-2000 according to the ENSO and PDO forecast, with the additional constraint that years in a three-year moving window centered on the forecast year are removed from the data set prior to resampling. Thus ensemble forecasts with about 12 members are produced for each water year from 1924-2001, and are subsequently adjusted to remove hydrologic model bias. The forecast probability distribution is then estimated from the ensemble of forecasts for each year. Reliability and skill of the forecasts are evaluated over the 1924-2001 period for which coincident climate and natural streamflow records are available by calculating the variance of the forecast ensemble divided by the variance of the entire observed streamflow data (a simple measure of skill relative to climatology), and the frequency with which the observations are outside the range of the ensemble predictions for each month (a measure of reliability). The frequency of observations falling above (below) the highest (lowest) ensemble member in the forecast is also estimated.