This paper studies the predictability of long-term unemployment (LTU) and analyzes its main determinants using rich administrative data in Sweden. Compared to using standard socio-demographic variables, the predictive power more than doubles when leveraging the rich data environment. The largest gains come from adding job seekers' employment history prior to becoming unemployed. Applying our prediction algorithm over the unemployment spell, we show that dynamic selection into LTU explains at least half of the observed decline in job finding. While the within-individual declines are small on average, we find substantial heterogeneity in the individual-level declines and thus reject the commonly used proportional hazard assumption. Applying our prediction algorithm over the business cycle, we find that the cyclicality in average LTU risk is not driven by composition but rather by within-individual cyclicality and that individual rankings are relatively persistent across years. Finally, we evaluate the implications of our findings for the value of targeting unemployment policies and how these change over the unemployment spell and the business cycle.