Hierarchical forecasting of residential construction activity: a three-stage architecture combining seasonal-trend decomposition, deep residual learning, and minimum trace reconciliation

While advanced forecasting frameworks are well established for operational and micro-level applications, a critical gap remains for frameworks capable of modelling national industrial activities, such as residential construction, that necessitate strategic management due to their profound macroeconomic and environmental footprints. To address this gap, a hierarchical forecasting framework is developed that considers time series across three aspects of residential construction activity—gross floor area, number of consents, and capital value—disaggregated by building typology (detached houses, townhouses, and apartments). Given the absence of forecasting literature at this scale and resolution, an optimal model selection workflow was executed, resulting in a novel three-stage hybrid architecture. First, a statistical model isolates the trend and seasonal components. Then, a Long Short-Term Memory network performs deep residual learning to capture non-linear relationships left unmodelled. Finally, empirical minimum trace reconciliation is applied to guarantee hierarchical coherence across the typology-level forecasts and top-level aggregates. The framework was validated using 35 years of building consent data from New Zealand. The proposed architecture consistently outperformed all isolated base models, with an average sMAPE of 8.1% for the top-level aggregates, while achieving reductions in MAE across features ranging from 6% to 28.5% against the two-stage hybrid model, and 23.3% to 39.7% against the statistical baseline. The validated model is then deployed to project New Zealand’s residential construction activity, showing an anticipated volume of 123,000 new dwellings, 17 million m 2 of GFA, and $54 billion NZD generated in capital value, during the 3-year forecasting horizon.