This paper explores how jobseekers’ expectation errors about their return to work can be used to inform public policy. While these errors are relatively simple to measure, they involve behaviour and primitive beliefs about the labor market, which makes it difficult to use them in practice to design policies. Using surveys, we measure jobseekers’ expectations about the remaining duration of their unemployment, their self-reported search effort, and their perceptions of labor market fundamentals. We then exploit the rich administrative data available from the Public Employment Service in France to study the determinants of biased beliefs about unemployment duration. We show that jobseekers’ expectations about their return to work have a predictive content and are on average too optimistic in the short run. Taking advantage of the large set of observables at our disposal, we characterize their heterogeneity by classifying respondents into groups based on their predicted biased beliefs using machine learning methods. Importantly, we can distinguish three groups with optimistic, rational, and pessimistic expectations about their return to work respectively. To understand the causes of these different biases, we show that they are correlated with different types of biased representations of the wage distribution and the arrival rate of job offers, as well as with the choice of parameters and search behavior. Finally, we study the impact of an interview with the social worker on the expectations and the declared search effort. We show that there is heterogeneity in the impact with respect to the predicted biased beliefs. This helps to inform the design of future interventions.