The Impact of Model Complexity on Data Shift Adaptation: A Comparative Analysis in Hospitality Domain

Data shifts—changes in distribution between training and deployment environments—pose a critical challenge to model generalization in the hospitality industry. This study investigates how model complexity influences resilience to such shifts, examining six architectures ranging from logistic regression (249 parameters) to convolutional neural networks (2 million+ parameters). Using 119,000+ hotel booking records with deliberately introduced geographic and temporal distribution shifts, we reveal a striking “double descent” pattern in generalization performance. As complexity increases, models first become more robust to distribution shifts, then dramatically more vulnerable at intermediate complexity (40,000 parameters), before regaining and exceeding their earlier generalization capabilities at higher complexities. While CNNs achieved superior accuracy (90% vs 87% under shifted distributions), tree-based ensemble methods demonstrated disproportionate stability relative to their parameter count. Our findings provide empirical evidence that the relationship between model complexity and generalization under data shifts is non-monotonic, challenging the bias-variance tradeoff paradigm. Code available at https://github.com/DannyMeb/Data-shift-Effect-on-Model-Generalization.git.