A common benchmark for Aedes albopictus phenological models reveals surveillance data, not model complexity, as the limiting factor

Forecasting the seasonal dynamics of Aedes albopictus, a highly invasive arbovirus vector now established across much of Europe, is a growing public health priority. Numerous mechanistic and correlative models have been developed over the last two decades, but largely in isolation, each calibrated and validated against its own data, which has prevented any assessment of where predictive capacity is genuinely limited. Rather than ranking models, which were built for different purposes and cannot be reduced to a single skill score, we establish common ground: a shared evaluation framework in which six mechanistic models, one correlative machine learning model, and two naive baseline predictors are driven by identical ERA5-Land forcing and assessed against the harmonised VectAbundance ovitrap dataset (Albania, Italy, Switzerland) using a single protocol, with robustness quantified by bootstrap resampling. We additionally combine the models into a weighted ensemble via approximate Bayesian computation. Both naive baselines outperformed every process-based and machine learning model, and the ensemble matched but did not exceed the best individual models. Predictive skill varied markedly in space, with all model classes, including the uncalibrated baselines, performing best in intensively monitored regions and worst at sparsely sampled sites. This pattern suggests that geographic variation in skill reflects the density and consistency of surveillance data more than differences in model complexity. We present this benchmark not as a league table but as shared infrastructure, common evaluation data, covariates, and protocol, on which comparable, policy-relevant forecasting tools can be built, and we argue that sustained, harmonised surveillance is the principal precondition for operational Ae. albopictus prediction.