Total Order Broadcast (TOB) is a fundamental building block at the core of a number of strongly consistent, fault-tolerant replication schemes. While it is widely known that the performance of existing TOB algorithms varies greatly depending on the workload and deployment scenarios, the problem of how to forecast their performance in realistic settings is, at current date, still largely unexplored. In this paper we address this problem by exploring the possibility of leveraging on machine learning techniques for building, in a fully decentralized fashion, performance models of TOB protocols. Based on an extensive experimental study considering heterogeneous workloads and multiple TOB protocols, we assess the accuracy and efficiency of alternative machine learning methods including neural networks, support vector machines, and decision tree-based regression models. We propose two heuristics for the feature selection phase, that allow to reduce its execution time up to two orders of magnitude incurring in a very limited loss of prediction accuracy.