The success, if not survival, of service businesses depends on their ability to satisfy their customers. Yet, businesses often recognize slumping customer satisfaction too late and ultimately fail. To prevent this, marketers require early warning tools. In this paper, we build upon online ratings as a direct measure of customer satisfaction and, based on this, predict business failures. Specifically, we develop a variable-duration hidden Markov model; it models the rating sequence of a service business in order to predict the likelihood of failure. Using 64,887 ratings from 921 restaurants, we find that our model detects business failures with a balanced accuracy of 78.02%, and this prediction is even possible several months in advance. In comparison, simple metrics from practice have limited ability in predicting business failures; for instance, the mean rating yields a balanced accuracy of only around 50%. Furthermore, our model recovers a latent state (“at risk”) with an elevated failure rate. Avoiding the at-risk state is associated with a reduction in the failure rate of more than 41.41%. Our research thus entails direct managerial implications: we assist marketers in monitoring customer satisfaction and, for this purpose, offer a data-driven tool that provides early warnings of impending business failures.