The 'Baulandumlegung' Prognosis Algorithm: A Non-Stationary Spatiotemporal Approach for Forecasting Land Readjustment Projects Using Machine Learning and 120 Years of Unstructured Data from Frankfurt am Main (1902–2022)

This study integrates historical institutionalism with geostatistical methods to examine the evolution and impact of land readjustment (Baulandumlegung) projects in Frankfurt am Main over the past 120 years. Drawing on critical junctures and path dependence as theoretical pillars, we trace how pivotal legal, political, and economic events—such as new planning laws, macroeconomic crises, and wartime disruptions—shaped planning policies and practices. We hypothesize that once certain institutional pathways were chosen, subsequent changes became increasingly constrained, leading to distinct project designs, financing schemes, and dispute-resolution practices. Methodologically, we present a novel Baulandumlegung Prognosis Algorithm that combines machine learning, spatiotemporal statistics, and a historical institutionalist ground to forecast land readjustment outcomes. The dataset, which spans from 1902 to 2022, integrates unstructured data from archival records, property transaction listings, and Sütterlin-scripted documents across distinct historical periods, including the German Empire, Weimar Republic, and post-World War II eras. Applied to distinct periods of voluntary (139) and enforcement-based (183) project implementations, the algorithm incorporates seven key components into a single predictive pipeline—geographically weighted regression, Mahalanobis distance matching, difference-in-differences estimation, anisotropic variogram analysis, odds ratio clustering, a historical weight function, and a cumulative probability predictor. These components collectively capture spatial dependence, isolate treatment effects, and incorporate path-dependent institutional changes. At the core of this framework lies a machine-learning classification module. By applying cross-entropy loss to guide training, we iteratively optimize hyperparameters (e.g., learning rate, maximum tree depth, regularization terms) to minimize predictive bias and enhance model generalizability in real estate markets. We pair 'treated' land readjustment sites with comparable 'control' areas to quantify the causal impact of Baulandumlegung on property prices and urban form over time. Additionally, anisotropic (semi)variogram analyses uncover hidden spatial heterogeneities by illustrating how direction-specific correlation structures influence project outcomes. A historical weight function further operationalizes how critical junctures shaped path-dependent trajectories across German political regimes. Our results demonstrate significant spatial heterogeneity in the effects of land readjustment across Frankfurt am Main, with varying impacts on property values and urban development patterns. Empirical validation using spatial cross-validation confirms that our machine learning approach successfully predicts project locations at specific times while accounting for complex spatiotemporal relationships outperforming traditional parametric and non-spatial hedonic models. Preliminary findings suggest that critical junctures—such as the post-war reconstruction and amendments to the German Building Code—acted as catalysts for reaffirming or altering Frankfurt’s planning trajectory. This methodological innovation provides urban planners and policymakers with a robust tool for analyzing not only prospective land readjustment areas but also regions affected by other planning instruments, effectively combining a historical-institutionalist framework with modern machine-learning techniques.