Robust inference methods are essential for parameter estimation and model selection in stochastic modeling approaches, which provide interpretable frameworks for describing time series of complex phenomena. However, applying such methods is often challenging, as they typically demand either high-frequency observations or access to the model's analytical solution, resources that are rarely available in practice. Here, we address these limitations by designing a novel Monte Carlo method based on full-path statistics and bridge processes, which optimize sampling efforts and performance even under coarse sampling. We systematically investigate how experimental design -- particularly sampling frequency and dataset size -- shapes inference accuracy, revealing optimal sampling regimes and the fundamental limits of model distinguishability. We validate our approach on four datasets -- optical tweezers, human microbiome, topic mentions in social media, and forest population dynamics -- where resolution-dependent limits to inference emerge, offering fresh insight into ongoing debates about the dominant sources of noise in these systems. Overall, this work shows how combining minimal stochastic models with path-inference tools and model selection can guide the experimental design of optimized strategies in data collection and clarify the boundaries of model-based understanding in complex systems.
翻译:暂无翻译