Synthetic Production Dataset for Causal Machine Learning in Operations Management

Abstract

Causal machine learning represents a promising paradigm for estimating intervention effects in operations management, yet its adoption is constrained by the scarcity of structured datasets suitable for method development and validation. This paper presents a synthetic production dataset comprising 10,000 production order records from a simulated metalworking factory with ten welding lines and ten product types. The dataset incorporates a binary treatment variable representing the implementation of drum-buffer-rope synchronization from the Theory of Constraints on two production lines. Data generation follows established statistical distributions with explicit causal structure, including line-product interaction effects, Little’s Law relationships, and treatment-dependent outcome variations. The dataset includes 15 variables capturing order characteristics, operational parameters, performance outcomes, and a line-level carry-over state that induces temporal serial correlation between successive orders on the same line. Primary reuse value lies in benchmarking causal machine learning methods, particularly meta-learners, tree-based algorithms, and neural network–based approaches for heterogeneous treatment effect estimation in operations management contexts. The dataset supports methodological research on ex-ante effect estimation without requiring explicit simulation models of production systems.