Fishery managers often rely on forecasts of future population abundance to set allowable harvest quotas or exploitation rates. While there has been substantial research devoted to identifying environmental factors that can predict recruitment for individual populations, such correlations often degrade over time, thereby limiting their utility for management. Conversely, examining multiple populations at once to detect shared, spatially structured patterns can offer insights into their recruitment dynamics that are advantageous for forecasting. Here, we develop a population dynamics model for natural origin coho salmon (Oncorhynchus kisutch) stocks in Washington State that leverages spatial and temporal autocorrelation in marine survival to improve one-year-ahead forecasts of adult returns. Executed in a Bayesian hierarchical integrated modelling framework, our spatiotemporal approach incorporates multiple data types and shares information among stocks to estimate key biological parameters that are informative for forecasting. Retrospective evaluation of one-year-ahead forecast skill indicated that the spatiotemporal integrated population model (ST-IPM) outperformed existing forecasts of Washington State coho salmon returns by 25–38 % on average. Moreover, the ST-IPM estimates parameters that were previously non-identifiable for many stocks, and propagates uncertainty from multiple contributing data sources into model forecasts. Our results add to a growing body of work demonstrating the utility of spatiotemporal and integrated approaches for modelling population dynamics, and the framework developed here has broad applications to the assessment and management of coho salmon in Washington State and elsewhere throughout their range.