This paper aims to evaluate whether metropolitan house-price downturns can be predicted out-of-sample, how predictability varies with the forecast horizon and whether machine learning classifiers improve on a naive persistence benchmark.
Using a panel of 94 US metropolitan areas drawn from the Federal Housing Finance Agency (FHFA) All-Transactions House Price Index (1978–2026; 18,083 metro-quarters), a downturn is defined as a metropolitan index that is net lower h quarters ahead. Eight classifiers (four support-vector-machine kernels, random forest, gradient boosting, a feedforward neural network and logistic regression) and a persistence rule are compared across horizons of 1, 4 and 8 quarters. Models are evaluated with a leakage-safe expanding-window panel protocol, tuned on the area under the receiver operating characteristic (ROC) curve and assigned a decision threshold that maximizes the training geometric mean of sensitivity and specificity. Inference uses a quarter-blocked bootstrap.
Metropolitan downturns are predictable out-of-sample, but a persistence benchmark tuned to the same operating criterion as the models captures most of that predictability. On the geometric mean of sensitivity and specificity, the primary metric, the tuned benchmark reaches 0.68–0.76 across horizons and the best classifier 0.75–0.82. Computed within each test block, the classifiers’ area under the ROC curve is 0.62–0.79. After correction for multiple comparisons across models and horizons, the one-quarter random forest improves on the tuned benchmark in the main analysis (Holm-adjusted p = 0.008). This finding does not hold with seasonally adjusted labels or equal training sizes. No single algorithm dominates, and parsimonious linear and logistic specifications are competitive throughout. Under an asymmetric loss that weights missed downturns more heavily than false alarms, the classifiers reduce total misclassification cost relative to persistence at all horizons.
The index is a smoothed, periodically revised, not seasonally adjusted series; the evaluation uses the standard feature-outcome alignment rather than real-time data vintages; kernel models are trained on capped subsamples for tractability. A transaction-based purchase-only index and a full local-fundamentals panel are identified as priority extensions.
The study provides a metropolitan-panel early-warning comparison anchored to an explicit naive benchmark. The benchmark quantifies the modest marginal value of machine learning: one comparison is significant after correction in the main analysis, but that finding does not hold in the robustness checks; the study also contributes a horizon-specific attribution of metropolitan vulnerability and a cost-sensitive assessment of operational value.
