This paper shows that individual-level heterogeneity in survival expectations derived from subjective survey information improves the out-of-sample predictions of a dynamic model of retirement and saving. We consider three approaches to modelling survival: life tables, average subjective expectations and individual-specific estimates based on reported survival probabilities. The models are estimated using Dutch data from the 1990s, a period during which workers could retire from age 59 at no penalty to pension benefits. Actuarial adjustments were introduced in the early 2000s and we use data from the period 2006-2016 to evaluate the accuracy of the counterfactual predictions. While the three models yield different preference estimates, their within-sample fit is similar. Out-of-sample forecasts do differ markedly. The models with homogeneous expectations anticipate a 4 or 5-year increase in the average retirement age in the new regime, compared with an observed increase of 2.6 years. The model with heterogeneous expectations predicts a more realistic increase of 2.7 years. Expectations matter when it comes to counterfactual predictions, even if different combinations of preferences and expectations appear equivalent within a given institutional setting.