Machine learning prices property well but treats every transaction as independent, discarding the market cycles and persistence that housing markets are built on. This study tests whether a hybrid framework fixes that: four models — multiple linear regression, random forest, support vector machine and gradient boosting — fitted to twenty years of one-family dwelling transactions in Brooklyn, New York City, then paired with an ARIMA model applied to their residuals.

Hybridisation improved every model. Prediction error fell by roughly 35% and average R² rose from 0.37 to 0.63, with Diebold–Mariano tests confirming the gains were statistically significant. The residuals carry real temporal information that the machine learning models threw away — and the same logic applies wherever cross-sectional risk factors and shifting market conditions both drive the answer, premium rating included.