Scalable Neural Synthetic Control with Individual Counterfactuals under Hidden Confounding

Maha Ouali, Badih Ghattas, Emmanuel Flachaire, Philippe Charpentier, BOZZI Laurent

Keywords: Causal inference, Time series, Representations Learning, Scalability

Abstract: As causal inference scales to high-frequency, individual-level observational data, traditional estimators face a dual challenge: computational intractability and bias from hidden confounding. In settings with staggered treatment adoption, prevalent in energy grid management and clinical monitoring, classical Synthetic Control (SC) methods suffer from prohibitive computational overhead when applied to individual units, while neural "black-box" regressors often diverge under non-stationary dynamics. We propose B-Twin (Balanced-Twin), a scalable neural framework that bridges representation learning with the structural rigor of synthetic control. B-Twin learns latent representations of individual trajectories that act as proxies for hidden confounders and uses a neural weight regressor to construct synthetic controls as weighted combinations of observed units. This replaces costly per-unit optimization with efficient neural inference while preserving interpretable counterfactual construction. Our experiments on synthetic and real-world datasets demonstrate that B-Twin effectively recovers treatment effects where outcome-regression baselines fail, while scaling efficiently to large populations.

Submission Number: 6