We add the Bernanke-Gertler-Gilchrist model to a modified version of the Smets-Wouters model of the US in order to explore the causes of the banking crisis. We test the model against the data on HP-detrended data and reestimate it by indirect inference; the resulting model passes the Wald test on output, inflation and interest rates. We then extract the model's implied residuals on US unfiltered data since 1984 to replicate how the model predicts the crisis. The main banking shock tracks the unfolding 'sub-prime' shock, which appears to have been authored mainly by US government intervention. This shock worsens the banking crisis but 'traditional' shocks explain the bulk of the crisis; the non-stationarity of the productivity shock plays a key role. Crises occur when there is a 'run' of bad shocks; based on this sample they occur on average once every 40 years and when they occur around half are accompanied by financial crisis. Financial shocks on their own, even when extreme, do not cause crises - provided the government acts swiftly to counteract such a shock as happened in this sample.