We develop a model of financial crises with both a financial amplification mechanism, via frictional intermediation, and a role for sentiment, via time-varying beliefs about an illiquidity state. We confront the model with data on credit spreads, equity prices, credit, and output across the financial crisis cycle. In particular, we ask the model to match data on the frothy pre-crisis behavior of asset markets and credit, the sharp transition to a crisis where asset values fall, disintermediation occurs and output falls, and the post-crisis period characterized by a slow recovery in output. We find that a pure amplification mechanism quantitatively matches the crisis and aftermath period but fails to match the pre-crisis evidence. Mixing sentiment and amplification allows the model to additionally match the pre-crisis evidence. We consider two versions of sentiment, a Bayesian belief updating process and one that overweighs recent observations. We find that both models match the crisis patterns qualitatively, generating froth pre-crisis, non-linear behavior in the crisis, and slow recovery. The non-Bayesian model improves quantitatively on the Bayesian model in matching the extent of the pre-crisis froth.