Generalized additive models (GAMs) were used to investigate the relationships between annual recruitment of natural coho salmon (Oncorhynchus kisutch) from Oregon coastal rivers and indices of the physical ocean environment. Nine indices were examined, ranging from large-scale ocean indicators, e.g., Pacific Decadal Oscillation (PDO), to indicators of the local ecosystem (e.g., coastal water temperature near Charleston, OR). Generalized additive models with two and three predictor variables were evaluated using a set of performance metrics aimed at quantifying model skill in short-term (approximately 1 yr) forecasting. High explanatory power and promising forecast skill resulted when the spring/summer PDO averaged over the 4 yr prior to the return year was used to explain a low-frequency (multi-year) pattern in recruitment and one or two additional variables accounted for year-to-year deviations from the low-frequency pattern. More variance was explained when averaging the predictions from a set of models (i.e., taking the ensemble mean) than by any single model. Making multiple forecasts from a set of models also provided a range of possible outcomes that reflected, to some degree, the uncertainty in our understanding of how salmon productivity is driven by physical ocean conditions.