This paper applies a complex system approach to predicting downturns in housing markets. We test the Log-Periodic-Power-Law (LPPL) model on 20 years of housing price data over nine regional subsamples within the U.S. We propose a new restriction to the model to remedy estimation issues due to the low frequency of house price data. We find that the restricted LPPL model well describes the times before the downturns. The out of sample predictions of the timing of the downturns lie within a narrow time span. With the restricted model we achieve a high success rate in issuing predications before downturns and not issuing predictions before no downturns.