Anthropogenic climate change induced snowpack loss is affecting streamflow predictability, as it becomes less dependent on the initial snowpack conditions and more dependent on meteorological forecasts. We assess future changes to seasonal streamflow predictability over two large river basins, Liard and Athabasca in western Canada, by approximating streamflow response from the variable infiltration capacity (VIC) hydrologic model with the Bayesian regularized neutral network (BRNN) machine learning emulator. We employ the BRNN emulator in a test-bed ensemble streamflow prediction system by treating VIC simulated snow water equivalent (SWE) as a known predictor, and precipitation and temperature from GCMs as ensemble forecasts, thereby isolating the effect of SWE on streamflow predictability. We assess warm-season mean and maximum flow predictability over 2041-2070 and 2071-2100 future periods against 1981-2010 historical period. The results indicate contrasting patterns of change, with the predictive skills for mean flow generally declining for the two basins, and marginally increasing or decreasing for the headwater subbasins. The predictive skill for maximum flow declines for the relatively warmer Athabasca basin, and improves for the colder Liard basin and headwater subbasins. While the decreasing skill for the Athabasca is attributable to substantial loss in SWE, the improvement for the Liard and headwaters can be attributed to an earlier maximum flow timing that reduces the forecast horizon and offsets the effect of SWE loss. Overall, while the future change in SWE does affect the streamflow prediction skill, the loss of SWE alone is not a sufficient condition for the reduction in streamflow predictability.