A benchmarking study on process model forecasting: univariate vs. multivariate approaches

Process Model Forecasting (PMF) extends predictive process monitoring by shifting the predictive focus from individual cases to the evolution of the overall process model. This is typically done by forecasting directly-follows (DF) relations to construct future process graphs that reflect anticipated system behavior. However, most existing approaches rely on univariate forecasting, modeling each DF relation in isolation. This neglects structural dependencies and interactions inherent in real-world processes. In this paper, we present a comprehensive benchmark study comparing univariate and multivariate forecasting techniques, including statistical, machine learning, and deep learning models, for predicting DF relations in PMF. We evaluate these methods on a diverse set of real-life event logs and assess their performance using both traditional statistical metrics and process-specific evaluation criteria. Our results show that forecasting effectiveness varies substantially across datasets and methods, highlighting the importance of context-aware model selection. All implementations and datasets are publicly available in our GitHub repository: https://github.com/YongboYu/PMF-Benchmark .