Housing price prediction in Ireland: a machine learning framework integrating geospatial encodings and macroeconomic indicators

Abstract

This study aims to develop and evaluate a machine learning framework for predicting residential property prices in Ireland using the complete Irish Property Price Register, a population-complete data set of 736,002 arm’s-length transactions spanning January 2010 to March 2026. It formally quantifies the marginal contribution of geospatial, structural and macroeconomic data modalities. A multi-modal predictive framework integrates structural transaction data, address-derived geospatial location signals via Bayesian out-of-fold target encoding and macroeconomic indicators from the Central Statistics Office. Five models are trained and evaluated, including Ridge Regression, Random Forest and three HistGradientBoosting variants on an 80/20 train-test split. Geospatial location features increase R² by 0.2552, from 0.3494 to 0.6046, reducing mean absolute error by €28,271 per property. Macroeconomic lag features covering the consumer price index, unemployment and earnings change R² by less than 0.0001. The best model, HistGradientBoosting v3, achieves R² = 0.6046 and MAE = €80,755 on 144,470 held-out observations. The study is constrained by the absence of floor area, bedroom count and BER information for most transactions in the Irish Property Price Register. To the best of the authors’ knowledge, this is the first population-scale machine learning study applied to the complete Irish Property Price Register. It introduces Bayesian out-of-fold target encoding of raw address strings as a geocoding-free geospatial method generalisable to any national transaction register and delivers an ablation study isolating geospatial and macroeconomic modality contributions in the Irish residential market.