Construct a weighted combination of control units to create a synthetic counterfactual for one treated unit, validate pre‑period fit, and quantify post‑period impact.
Uses constrained optimization to find non-negative weights that sum to 1, minimizing pre-treatment RMSPE (root mean squared prediction error).
Runs synthetic control for each control unit as if it were treated, calculating p-values based on effect size rankings.
Compares pre/post RMSPE ratios between treated and control units to assess statistical significance of treatment effects.
Your data needs a unit_column (region/store ID), time_column (period), outcome (metric to analyze), plus identification of treated_unit and treatment_period.
Data format: Panel data with one treated unit and multiple control units. Algorithm uses Nelder-Mead optimization to find optimal weights that minimize pre-treatment prediction error.
Minimum requirements: At least 10 pre-treatment periods for reliable fit, 5+ control units (ideally 20+). Control units must remain untreated throughout entire period.
What you get: Synthetic control weights, average treatment effect, pre/post RMSPE, placebo test p-values, weights table showing donor contributions.
From donor selection to inference with placebos
Finds optimal non-negative weights that sum to 1, minimizing squared prediction errors in pre-treatment periods using constrained optimization.
Applies synthetic control method to each control unit (up to 20) to generate distribution of placebo effects for significance testing.
Calculates gap between actual and synthetic control in post-period, reports average effect, RMSPE ratios, and p-value from placebo distribution.
When there’s one treated region/product and many similar controls, Synthetic Control provides an interpretable counterfactual with strong visual diagnostics.
Ideal for policy evaluations, city‑level interventions, one‑off feature launches, and events where randomized experiments are infeasible. By matching pre‑period behavior closely and stress‑testing with placebos, results are intuitive to present and defensible.
Key Idea: A weighted average of control units can approximate the treated unit’s counterfactual trajectory absent treatment.
Evaluate interventions with a transparent counterfactual
Read the article: Synthetic Control