Chronological Prediction of Undated Archaeological Assemblages: A Comparative Study of Bayesian Statistics and Machine Learning Approaches

The chronological assignment of archaeological assemblages without radiocarbon dating represents a fundamental challenge in archaeological practice. This study presents the first systematic comparison between Bayesian statistical methods and machine learning algorithms for chronological prediction of undated archaeological contexts. Using a comprehensive dataset of radiocarbon determinations from Late Neolithic to Early Bronze Age sites in Eastern Iberia, we evaluate the predictive performance of Dirichlet multinomial models against Support Vector Machines, neural networks, and ensemble methods. Archaeological variables include arrowhead typologies, metal artifacts, and Bell Beaker ceramics across six chronological phases. Results demonstrate that machine learning approaches, particularly ensemble methods, achieve superior predictive accuracy compared to traditional Bayesian methods when validated through expert archaeological assessment. Cultural markers, especially metal presence and Bell Beaker ceramics, provide significantly stronger chronological signals than lithic typological features. Temporal cross-validation reveals systematic performance degradation compared to standard validation, emphasizing the need for specialized validation protocols in archaeological contexts. Random Forest demonstrates the highest temporal robustness, while uncertainty quantification through Bayesian approaches proves valuable for identifying problematic predictions. These findings establish machine learning as a reliable tool for systematic chronological attribution of undated archaeological assemblages, with significant implications for museum collections and rescue archaeology.